Mach. Learn.: Earth 1 (2025) 011001 https://doi.org/10.1088/3049-4753/adde60 OPEN ACCESS RECEIVED 3 February 2025 REVISED 21 May 2025 ACCEPTED FOR PUBLICATION 29 May 2025 PUBLISHED 30 July 2025 Original content from this work may be used under the terms of the Creative Commons Attribution 4.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. TOPICAL REVIEW The need of explainability in low-carbon urban system design using AI: A systematic review Tong Chen and Ramit Debnath∗  University of Cambridge, Cambridge CB2 1PX, United Kingdom ∗ Author to whom any correspondence should be addressed. E-mail: rd545@cam.ac.uk Keywords: explainable artificial intelligence, climate change, urban system, urban planning Abstract Urban activities account for over 70% of carbon emissions, making urban systems crucial targets for strategies aimed at reducing emissions. The use of big data and artificial intelligence (AI) holds promise in reducing urban emissions through efficient data analysis in various sectors. However, current reviews on AI’s role in urban decarbonisation are broad and do not specifically tackle urban systems or the detailed application of big data and AI techniques. This study systematically reviewed 91 articles using preferred reporting items for systematic reviews and meta-analysis protocol to assess AI-driven approaches in low-carbon urban decision making and their effectiveness for real-world needs. The findings highlight the significant potential of AI-driven methods in carbon emission forecasting, energy consumption driver identification, and low-carbon planning. Building energy prediction emerges as the most common focus, representing nearly half of the reviewed studies. Machine learning, especially random forest for energy prediction and K-Means for consumption pattern recognition, is more prevalent than deep learning, with applications mainly supporting government and enterprise decisions. However, notable gaps remain. Research goals often lack multi-objective optimisation and cross-sector integration. Data limitations persist in behavioural, technological, spatial, and temporal dimensions. Crucially, explainable AI (XAI) remains underused, undermining model transparency, trust, and uptake, particularly in policymaking contexts where understanding by officials and stakeholders is essential. The absence of interpretability hampers real-world application and public engagement. This study contributes a more integrated perspective, explicitly incorporates XAI, and evaluates AI models through the lens of practical relevance, aiming to enhance their applicability in urban climate action. 1. Introduction Global decarbonisation is one of the defining challenges of our time. Unrestrained carbon emissions have driven global warming, creating a decisive crisis with potentially catastrophic consequences such as rising sea levels, extreme weather events, and large- scale displacement [1]. From 1990 to 2023, global carbon emissions have increased by approximately 75%, increasing from around 20 billion tonnes to more than 35 billion tonnes [2]. Rapid decarbonisation is crucial for a sustainable future. Although cities occupy only 3% of the Earth’s surface, they consume two-thirds of global energy and account for more than 70% of carbon emissions [3]. The main sources of urban greenhouse gas emissions include transport, energy use in buildings, electricity consumption, and waste and industrial production processes [4]. These patterns highlight the importance of reducing the emissions produced by urban systems. The term urban systems refers to the essential components and networks within urban areas [5, 6]. According to the existing literature and reports on carbon emissions, urban systems are characterised by the combinations and interactions of buildings, transport networks and other infrastructure, open spaces, energy and water supplies and waste management systems (figure 1) [3, 6]. These combinations and interactions contribute to © 2025 The Author(s). Published by IOP Publishing Ltd https://doi.org/10.1088/3049-4753/adde60 https://crossmark.crossref.org/dialog/?doi=10.1088/3049-4753/adde60&domain=pdf&date_stamp=2025-7-30 https://creativecommons.org/licenses/by/4.0/ https://creativecommons.org/licenses/by/4.0/ https://orcid.org/0000-0002-8242-2412 https://orcid.org/0000-0003-0727-5683 mailto:rd545@cam.ac.uk Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Figure 1. Components of urban systems. carbon emissions and absorption within urbanised areas. Specifically, the construction and operation of buildings [7–9], transportation activities [8, 9], energy consumption [8], the decomposition of municipal solid waste [8], and the collection, treatment, and transport of water [10] all generate substantial greenhouse gas emissions. In contrast, open spaces play a critical role in carbon sequestration. The green infrastructure included such as vegetation and soil-based systems within these open spaces has been shown to significantly absorb and store atmospheric carbon [11]. Embedded between buildings and transport corridors, these vegetated areas help offset emissions and contribute to balancing urban carbon dynamics [8, 9]. The integration of big data and artificial intelligence (AI) has become a promising approach for decarbonisation of urban systems. For example, in terms of monitoring, supported by the internet of things (IoT), AI enables highly efficient tracking of electricity usage, transport flows, and human activities, thereby illustrating real-time carbon emissions [12–15]. For prediction, AI uses historical data to forecast energy demand and supply in buildings, transport and other urban sectors, providing actionable insights to prevent inefficiencies [16–18]. Moreover, in optimisation, AI improves resource allocation and operational efficiency, providing low-carbon solutions for building layouts, transport modes, and green space designs [19–21]. In collaboration with IBM and the Environmental Systems Research Institute, Scottish Gas Networks developed the future gas system: Digital Twin project to optimise energy transmission within urban infrastructure. AI contributes to achieving decarbonisation goals by enabling smarter and more sustainable urban systems. Among the various AI technologies, XAI is receiving increasing attention. It is capable of unveiling the decision-making logic behind complex machine learning (ML) models, particularly ‘black-box’ models, thereby enhancing the transparency and trustworthiness of such systems [22]. XAI holds significant potential for urban system decarbonisation: it can assist researchers and policymakers in identifying key factors influencing carbon emissions [23], and in understanding why models generate specific predictions [24]. This is particularly essential for real-world applications, where XAI can provide credible and transparent technical support in policy-sensitive contexts such as multi-stakeholder coordination, accountability, and fairness assessment [22, 25, 26]. Existing reviews have explored the progress made in applying AI to address climate change challenges within cities, but their discussions often remain broad and generalised, without a focus on urban systems. For example, the role of smart city technologies in achieving the sustainable development goals (SDGs) has been reviewed, highlighting their potential to tackle climate-related issues [27]. Similarly, Bibri et al [28] and Shaamala et al [29] have examined AI-driven and IoT-based approaches for climate change, with a particular focus on eco-cities and green infrastructure. In particular, Yan and Qi [30] provided an overview of AI applications in low carbon urban planning, specifically in the context of China. However, their analysis focusses primarily on planning and design methodologies, without comprehensively discussing urban systems or offering a clear and statistical discussion of the big data and AI-driven technologies leveraged, 2 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath especially for XAI. This underscores the need for a more detailed examination of what specific big data and AI methods are applied and how they function, to establish a clearer framework for future applications in urban system decarbonisation. Furthermore, as big data and AI technologies continue to advance, the practical challenges associated with their implementation have become increasingly apparent. Evaluating not only the technological progress of AI but also its feasibility and gaps in real-world applications is critical for enabling effective deployment and bridging the theoretical potential and practical impact. Therefore, this paper presents a comprehensive review of the big data and AI-driven technologies supporting urban system decarbonisation, mainly addressing the research questions of: i. What and how are AI-driven approaches used for the decision making of low-carbon urban systems? ii. To what extent are AI-driven models reliable and adequate in meeting the decision- making needs of real-world applications? This study aims to answer these questions by systematic review of the literature following the preferred reporting items for systematic reviews and meta-analysis protocol (PRISMA). The PRISMA protocol was adopted to ensure the systematicity and transparency of the literature selection process. The application of PRISMA helped clarify the inclusion and exclusion criteria, thereby ensuring that the selected studies are representative, scientifically sound, and comparable. This approach strengthens the reliability and reproducibility of the review findings [31]. The Python programming language (spaCy) is used for structured information extraction to identify keywords in sections including introduction, data, methods, conclusion, and future work. Quantitative analysis is conducted to identify and understand the specific pathways through which big data and AI-driven methods are applied to achieve the decarbonisation of urban systems. Finally, this study examines the challenges and future potential of integrating AI technology and urban decarbonisation. The contribution of this study lies in providing a comprehensive perspective for academia and policymakers through a structured literature review, thereby enabling a nuanced understanding of the practical roles of big data and AI in developing low-carbon urban systems and advancing toward a sustainable future. This paper is organised as follows: section 2 outlines the research methodology and details the systematic literature review approach employed in this study. Section 3 presents general observations, objectives, data use, and methodological approaches of the reviewed articles. Section 4 provides a discussion of the pathways for AI applications in low-carbon urban systems and the distance between current studies and real-world application. Finally, section 5 concludes the paper by summarising the key findings and clarifying the implications of this research for advancing the role of AI in urban system decarbonisation. 2. Methods Clarifying the definition of AI-driven methods is a necessary step in delineating the scope of this study. John McCarthy originally defined AI as the science and engineering of making intelligent machines [32]. As the field has evolved, AI has become a cornerstone of computer science and is now widely regarded as the development of algorithms that replicate human cognitive processes through computational means [33, 34]. Within the existing literature, AI-driven methods are frequently reviewed as a distinct class of approaches encompassing machine learning (ML), deep learning (DL), optimisation algorithms (OAs), and XAI [33, 35–37]. Some recent reviews have further extended the classification to include generative AI, such as ChatGPT and diffusion-based models [38, 39]. However, given the nascent state of their application to urban system decarbonisation and the limited number of relevant publications to date, such models were not considered within the analytical scope of this review. Specifically, AI encompasses methods that train models using various types of data to answer specific questions by learning from experience, with ML serving as a core methodology within AI. As a significant subset of ML, DL employs artificial neural networks (ANNs) to develop intelligent models capable of solving complex problems. In this study, the term ML specifically refers to models that exclude DL techniques. OAs are used to fine-tune model parameters, improving performance, while XAI focusses on providing clear interpretations of the methods, procedures, and outputs involved in AI processes [34, 40]. A systematic literature review includes retrieving, mapping, aggregating, configuring, and critically evaluating the studies published to address and discuss the research topic. The PRISMA method was applied in the retrieval stage. This method includes four steps: identification, selection, eligibility, and inclusion of studies (figure 2). The Web of Science (WoS) database was selected for its comprehensive coverage of peer-reviewed academic literature [41–43]. The WoS is one of the most influential sources for scholarly metadata. WoS provides structured indexing across multiple citation indices, including the Science Citation Index Expanded 3 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Figure 2. The PRISMA flowchart for literature search and selection. Reproduced from [31]. CC BY 4.0. (SCIE), Social Sciences Citation Index (SSCI), and Emerging Sources Citation Index (ESCI), which collectively help minimize the inclusion of low-quality or predatory publications. This curated structure ensures a consistent level of source quality and metadata reliability [44]. In addition, WoS offers advanced filtering functions and structured export capabilities, which facilitate efficient screening and support the transparency and reproducibility required in systematic reviews. Three categories of keywords were used to ensure a comprehensive and systematic retrieval of relevant studies. These categories correspond to the main components of the research focus. In particular, this study focusses on scales at or above the urban level, so related keywords were also included. Keyword strings were generated using the Boolean operators AND and OR to optimise the search. AND was used to combine the three categories and the research scales to ensure the retrieval of studies addressing all aspects. OR was used within each category to allow for variations in terminology. Table 1 details the specific keywords used for each category. The search strings were applied to titles, abstracts, and keywords within the WoS. This review limits the timeframe to publications from 2017 to 2024 to ensure that only recent and relevant studies were considered. Only journals indexed in the ESCI, SSCI, and SCIE were included to ensure the credibility of the source. The type of source was limited to journal articles, excluding reviews, books, and conference reports. Furthermore, only papers published in English were selected. Studies considered entirely unrelated were excluded on the basis of research scales. An initial total of 1503 papers entered the screening phase. After removing duplicates, the titles and abstracts were reviewed and the studies were excluded if their 4 https://creativecommons.org/licenses/by/4.0/ Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Table 1. Keywords used for literature search strings. Category Keywords Urban systems ‘built-up areas’ OR ‘urban systems’ OR ‘urban planning’ OR ‘smart cities’ OR ‘urban infrastructure’ OR ‘energy systems’ OR ‘transport networks’ OR ‘street’ OR ‘water supplies’ OR ‘buildings’ OR ‘green spaces’ OR ‘waste management’ Low-carbon ‘low-carbon’ OR ‘carbon footprint’ OR ‘GHG’ OR ‘carbon emissions reduction’OR ‘decarbonisation’ OR ‘energy efficiency’ OR ‘energy consumption’ OR ‘climate mitigation’ OR ‘green buildings’ Big data and AI—driven approaches ‘big data’ OR ‘data-driven’ OR ‘AI’ OR ‘machine learning’ OR ‘deep learning’ OR ‘prediction’ OR ‘modelling’ OR ‘monitoring’ Research scales ‘city’ OR ‘cities’ OR ‘metropolitan area’ OR ‘re-gional’ OR ‘multi-city’ OR ‘national’ OR ‘continen-tal’ OR ‘global’ Table 2. Examples of exclusion cases based on the predefined screening criteria. Exclusion criteria Examples Inconsistent research scales Studies focusing on individual buildings, small samples of a few buildings, or limited to community-, district-, campus-, or industrial park-level analyses. Studies where the urban scale was ambiguous or the administrative level could not be clearly determined were also excluded. Inconsistent research scopes Studies focused on climate modelling without explicit connection to urban contexts; applications of AI for climate-related urban services without addressing carbon reduction; studies on air quality or urban heat without targeting decarbonisation; building control research where ‘low-carbon’ was mentioned only descriptively rather than as a modelling objective. Unclear dataset description Studies that did not specify data sources, provided vague or non-specific city samples, omitted the time span of the dataset, or failed to explain how the data were collected. Unclear method description Studies lacking detailed descriptions of model structure, including input/output variables, data splitting strategy (e.g. training/testing sets), hyperparameter settings, or performance evaluation metrics. AI-driven methods not used as primary approach Studies where AI methods were only used for comparison but the main analysis relied on traditional statistical models; where AI was limited to data preprocessing or visualisation; or where AI was included among multiple models but played only a minor role in the analysis. scales or scopes were inconsistent with the research focus. This process yielded 157 articles that progressed to the eligibility phase. In this phase, further reading led to the exclusion of 66 articles due to insufficient clarity in the description of datasets, ambiguous method explanations, or the absence of AI-driven methods as the primary approach in their analyses. The examples of the exclusion criteria could be found in table 2. Subsequently, the reference sections of the remaining papers were examined and 12 relevant studies were added to the final dataset, resulting in a total of 91 papers. This study employed a semi-automated workflow developed in Python to extract structured information from PDF-format academic literature, in order to support statistical analysis. The workflow was designed to retrieve key information from different sections of each paper, including the geographic focus (i.e. countries and cities) identified in the methods section, datasets used as indicated in the data or methods sections, AI- driven methods described in the methods section, and the main findings and research gaps outlined in the introduction and conclusion sections. 5 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Figure 3. Number of published papers by year from 2017 to 2024. In terms of implementation, a keyword matching strategy was adopted for the extraction of geographic focus and AI-driven methods. Geographic entities were identified using the named entity recognition function of the spaCy natural language processing library, specifically through the GPE label, and then matched against an official list of place names to ensure consistency and accuracy. AI methods were derived from preliminary reading and synthesis of a sample of relevant literature [35–37, 40], from which a list of commonly used algorithm terms was manually compiled. The full list is provided in appendix table 2. These keywords were then matched to identify relevant mentions in the methods sections. However, the extraction of datasets used, main findings, and research gaps followed a fuzzy locating strategy, as the expressions used in the literature vary widely and are difficult to capture with fixed keywords. Typical patterns and indicative expressions used for extraction are summarised in appendix table 3. To ensure the accuracy and quality of the extracted information, all results were manually reviewed and corrected where necessary. Python tools were primarily used to improve efficiency, while the interpretation and validation of content were carried out by the researchers. It should also be noted that the research themes and objectives of each paper were not identified through automated methods but were instead manually summarised by the researchers based on close reading. To address RQ2 on the adequacy and reliability of AI-driven models, this study examines how existing literature has evaluated model performance and explainability under different climate action objectives, using representative examples for comparison and synthesis. In addition, it also investigates how AI models have been applied to generate policy recommendations for various stakeholder groups, focusing on the types of recommendations proposed and the models on which they were based. These analyses, viewed from multiple dimensions, including model performance, generalizability, explainability, and policy relevance, help to assess the extent to which AI-driven approaches can deliver reliable insights and support effective, stakeholder-oriented decision-making in real-world contexts. 3. Results 3.1. General observations Figure 3 illustrates the growing interest in applying AI-driven methods to low-carbon urban systems since 2017. Between 2017 and 2019, the number of publications remained relatively low, the number of publications per year being fewer than 10. A notable increase was observed in 2020 when the number of publications increased to 15. The number of publications remained steady, ranging from 12 to 15 annually. This trend culminated in a significant increase in 2024, during which 26 publications were found, marking the highest level of research in this field to date. Figure 4 shows that the research areas are primarily concentrated in China (n= 28) and the United States (n= 20), and 80.69% of the countries and regions worldwide have not been studied. This scenario may still 6 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Figure 4. Distribution of research areas by country in the literature. Figure 5. Number of research areas by city in the literature. reflect imbalances in research distribution, particularly affecting the Global South. In this study, a detailed analysis of the specific cities investigated was performed (figure 5), highlighting the cities with a research frequency of four or more. The results reveal that all of these cities are located in the United States or China, with New York being the most studied, appearing 14 times. A primary contributor to this disparity is the variation in data availability between different countries and cities. The literature indicates that the United States [45], the United Kingdom [46], those in the European Union [47], and the Republic of Korea [48], have disclosed detailed statistics on building-level energy consumption. This transparency facilitates extensive analyses of building energy consumption, whereas the majority of other countries lack comprehensive, research-ready data. However, the literature search was conducted solely in English and limited to the WoS database. This may have resulted in the underrepresentation of studies from the Global South, where local research is often published in other languages or in grey literature. Thus, the regional concentration of studies may partly reflect these search constraints. 7 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Figure 6. Keyword co-occurrence diagram where the size of nodes represents keyword frequency, and the thickness of edges represents the strength of co-occurrence relationships. 3.2. Objective landscape: AI-facilitated urban climate initiatives Figure 6 presents the co-occurrence network of keywords. ML is noticeably the most prominent keyword. It is primarily associated with the analysis of energy consumption in buildings, as highlighted by the yellow cluster. The green cluster indicates the application of DL in studies related to energy efficiency (EE), which is likely to be connected to smart city management. Other clusters demonstrate research focus on areas such as prediction, simulation, the development of benchmarking systems, and the optimisation of urban carbon emissions. A treemap was created to clarify the primary objectives and pathways of the current research in the urban climate action domain. It visualizes the hierarchical structure and distribution of ‘goal–method–data’ relationships. For example, in figure 7 we show that for the’target’ relating to the prediction of building energy consumption, the methods include ML and the’data’ that are used to train these ML models consists of characteristic building data, energy data, climate and environmental data, and so on. The size of each rectangle represents the relative number of studies and its overall distribution across different goals, methods, and data types (see figure 7). The table 3 further expands this illustration of ’goal-method-data’ in detail. 3.2.1. Building energy consumption prediction The most prevalent objective is to predict the energy consumption of the building (n= 50), which represents 54. 95% of all articles. Studies predict the electricity consumption [49], energy use [50, 52], and energy rating [55] of buildings. This area is dominated by ML that uses heavily building characteristic data, energy data, and climate and environmental data, which are considered the most important factors influencing building energy consumption [82, 83]. Although some literary works consider the entire life cycle of building energy consumption, most studies focus on emissions during the operational stage. However, embodied emissions (arising from materials and equipment), construction stage emissions (for example, machinery use), and demolition stage emissions require a comprehensive examination [84]. This is particularly important given that some research indicates that emissions from the operational stage are only approximately 30% greater than those of the construction stage [85]. 3.2.2. City carbon prediction and management The prediction and management of carbon in cities (n= 16) represents 17.58% of the reviewed articles, including analyses in multiple urban sectors, such as the power system [61], urban expansion [64], and cross-sectoral collaboration [65] of the entire city. ML and DL methods are widely employed. They use a 8 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Figure 7. Treemap of climate action objectives, AI-driven methods and data types. combination of energy data, demographic and socioeconomic data, and geospatial data. However, this research area suffers from a relative scarcity of studies on urban infrastructure and waste management, despite their significant contributions to urban energy consumption [62, 86]. The limited availability of urban-scale data hampers comprehensive analyses in these domains. 3.2.3. The development of new energy sources The development of new energy sources (n= 10) has become an increasing research objective in recent years, with the majority of the literature focussing primarily on identifying suitable locations for the implementation of solar energy [72, 87]. Such studies rely on the integration of climate and environmental data, building characteristic data, and geospatial data. DL is often applied to process satellite imagery, which facilitates large-scale detection from a top-down perspective [74]. 3.2.4. Building energy consumption reduction The objective of reducing building energy consumption (n= 8) involves simulating possible energy changes, identifying inefficient building characteristics for energy retrofit projects, and assessing the potential investment in EE [66, 71]. Compared to other objectives, explainable AI (XAI) is often employed in this context because understanding the influence of various factors on building energy consumption is essential for informed decision-making in energy reduction strategies [49]. 3.2.5. Transportation carbon prediction and management Studies on transportation carbon prediction and management (n= 7) mainly address the prediction or monitoring of transport carbon emissions and the creation of effective decarbonisation systems [20, 88]. These studies often leverage IoT to enable complete control on a city scale [79]. 3.3. Data typologies: big data for urban systems Figure 8 presents the types of data sets used in existing research and their respective data sources. 3.3.1. Data sources The availability of high-quality data is essential for advance research in this field. From the perspective of data source, public databases (n= 75) and government and public departments (n= 41) are the main 9 M ach. Learn.: E arth 1 (2025) 011001 T C h en an d R D ebn ath Table 3. Examples of AI-driven methods supporting climate action objectives. Climate action objective AI-driven methods Data types Model perfor-mance Explainability Application Examples Building energy consumption prediction ML: DT, RF, XGB, AdaBoost, SVM Building characteristic data, energy data, cli- mate and environmental data R2 ⩾ 0.85, MAE/MAPE≈ 5%–15% Some studies ranked variable importance Predict residential electricity use, fore- cast changes in energy use intensity (EUI), estimate building energy con- sumption [49–51] DL: DNN, CNN, LSTM R2 ⩾ 0.90 Not explained Predict residential energy use, forecast cross-regional thermal comfort, esti- mate building energy consumption [52–54] XAI: SHAP R2 ⩾ 0.90, MAPE≈ 40% Variable importance ranking, visual explanation per build-ing, variable interaction anal- yeses Optimise energy rating, develop cool-ing strategies, identify heating retrofit priorities [45, 55, 56] OA: Adam, GA, PSO R2 ⩾ 0.85 Some studies ranked variable importance Predict building energy consumption, forecast indoor temperature, predict city-wide hourly energy load [54, 57, 58] City carbon prediction and management ML: KNN, RF, DT, SVM, K-Means Demographic and socioeconomic data, geospatial data, building characteristic data R2 ⩾ 0.60 Some studies identified main contributing variables Predict household-level carbon emis-sions, analyse emissions–greenspace links, examine energy use and neigh- bourhood form [19, 59, 60] DL: TCN, LSTM, ANN, MLP Energy data, demo- graphic and socioeco- nomic data, climate and environmental data MAE≈ 3.4%, R2 ≈ 0.99 Some studies ranked variable importance Forecast power system and EV charg-ing load, map carbon emissions spa- tially, predict city-level electricity de- mand [61–63] (Continued.) 10 M ach. Learn.: E arth 1 (2025) 011001 T C h en an d R D ebn ath Table 3. (Continued.) Climate action objective AI-driven methods Data types Model perfor-mance Explainability Application Examples XAI: SHAP Geospatial data, demo-graphic and socioeco- nomic data, energy data R2 > 0.95, MAPE≈ 6%–9% Variable importance (direc-tion and strength) Analyse carbon impacts of urban ex-pansion, assess cross-sectoral information’s contribution to predictions [64, 65] Building energy consumption reduction ML: KNN, DT LGBM, SVR, RF, XGB Energy data, building characteristic data, de- mographic and socioeco-nomic data R2 ⩾ 0.90 Not explained Assess building energy performance, analyse building type–carbon links, identify retrofit-priority buildings [66–68] XAI: SHAP, CAM, PDP, PFI R2 ⩾ 0.70 Variable importance (direc-tion and strength), variable importance in image models Analyse urban morphology–energy links, identify hard-to-decarbonise homes, assess energy traits impacting house prices [46, 69, 70] DL: DenseNet, ANN, RBFNN Building character-istic data, energy data, geospatial data ACC⩾ 0.80 Some studies ranked variable importance Identify hard-to-decarbonise homes, estimate retrofit investment costs [70, 71] The development of new energy sources ML: RF, SVR, LGBM, SVM Climate and environ- mental data, geospatial data, building character- istic data F1⩾ 0.85 Not explained Compare SRP across LCZs, assess so- lar district heating feasibility, simulate building energy use [72, 73] DL: U-Net+ ResNet50, ANN R2 ⩾ 0.95, error< 5% Not explained Identify green and solar rooftops, pre-dict household electricity use, evaluate solar incentive schemes [18, 74, 75] (Continued.) 11 M ach. Learn.: E arth 1 (2025) 011001 T C h en an d R D ebn ath Table 3. (Continued.) Climate action objective AI-driven methods Data types Model perfor-mance Explainability Application Examples XAI: PDP Climate and environ- mental data, energy data, demographic and socioeconomic data Not reported Variable importance ranking Assess coordination of energy supply—demand transition, identify transition drivers [76] OA: NSGA-II Climate and environ- mental data, geospatial data, building character- istic data R2 = 0.98 Not explained Optimise block massing, improve solar absorption efficiency [77] Transportation carbon prediction and management ML: SVR, K-Means Energy data, demo-graphic and socioeco-nomic data, geospatial data MAPE< 10% Not explained Predict traffic flow, model transport-related carbon emissions [20, 78] DL: RBF-NN, ANN-RNN, MLP R2 > 0.95, MAPE≈ 1% Not explained Optimise multimodal transport con- figuration, predict transport-related emissions [17, 79, 80] XAI: ALE plots RMSE= 1.21 kg/per- son/day Nonlinear and threshold analysis Optimise multimodal transport con- figuration, predict transport-related emissions [81] 12 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Figure 8. Distribution of data source types across different data types. A detailed classification is shown in appendix table 1. sources of data. Public databases have the broadest coverage, covering all major data categories, while government and public departments primarily supply energy data, building characteristics data, and demographic and socioeconomic data. Businesses and industry (n= 17), experimental and simulation databases (n= 16), and research institutions and universities (n= 11) also play a critical role in providing energy data. This highlights that most studies are still based on publicly accessible data, which supports the generalizability and scalability of research efforts. At the same time, it underscores the importance of cross-sector collaboration in enriching data sources. 3.3.2. Data categories From the perspective of data categories, energy data (n= 44) stand out as the most essential and widely sourced category. As shown in figure 8, energy data is obtained from a variety of providers, including government agencies, public databases, and private companies. Although experimental and simulation data are less commonly used, they contribute to innovative research approaches. For example, Tardioli et al [89] demonstrated the value of clustering buildings with similar characteristics and conducting physical simulations on representative samples to generate high-resolution energy consumption data to evaluate the energy use of buildings in the whole city. This method shows how simulation data can fill gaps in real-world datasets while reducing computational demands. 3.3.3. Data gaps However, significant data gaps remain. Table 4 summarises additional data gaps identified by the authors across key domains, including carbon emissions, behaviour, technology, spatial coverage, and temporal coverage. With respect to carbon emissions, although there has been increasing availability of various types of carbon- related data, such as carbon footprint datasets [90], emitter-based emission tracking [91, 92], global carbon accounting initiatives based on unified protocols [93–95], and multi- source aggregated 13 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Table 4. Summary of identified data gaps and their potential research uses. Domain Data gaps Potential research use Carbon emissions Globally standardised and annually continuous carbon accounting data for cities Cross-city comparative analysis at a global scale Behaviour Publicly available electricity usage data at daily/hourly resolution for residential, industrial, and commercial sectors Accurate modelling of energy consumption patterns Data on building energy-saving behaviour, e.g. implementation records of energy-saving measures Optimisation of energy efficiency evaluation models and green building policies Technology City-level data on renewable energy consumption and infrastructure lay- out (e.g. solar PV, wind, heat pumps) Assessment of technical feasibility and urban energy transition potential Harmonised multi-year data on electric vehicle ownership and charging infrastructure, including spatial distribution Modelling of transport energy transition and support for spatial planning Carbon- or energy-related data on urban infrastructure and waste management system Studies on these two types of urban systems Spatial coverage Open data on energy, carbon emissions, and policy in developing countries Expanding regional representativeness in research Open data on energy, carbon emissions, and policy in small and medium-sized cities Expanding regional representativeness in research Temporal coverage Policy intervention tracking data before, during, and after implementation, especially with quantitative impact data Causal inference and policy effectiveness evaluation Missing staged energy consumption data for buildings during construction, operation, and maintenance periods Lifecycle energy consumption and carbon footprint assessment for buildings estimation data [96], city-level carbon accounting data still lack strong consistency. For instance, Nangini et al [97] compiled carbon emissions data for 343 cities worldwide, but the datasets are not always comparable due to variations in reporting years and accounting methodologies. Initiatives such as the Carbon Disclosure Project (CDP) [93], C40 Cities Climate Leadership Group (C40) [94], and the Global Protocol for Community-Scale Greenhouse Gas Emission Inventories [95] have made notable progress in addressing this issue, and the number of reporting cities has been steadily increasing. Nevertheless, the degree of temporal coverage remains insufficient for large- scale comparisons, as cities tend to report in different years and follow different update cycles, making coordinated multi-year analysis challenging. A second major data gap concerns electricity usage behaviours. Approximately 18. 68% of the studies report challenges related to the lack of detailed data on human electricity usage. These data are critical for accurately modelling spatial and temporal energy consumption patterns and improving the precision of peak demand forecasting [98, 99]. Another emerging area of concern is the data availability on new technologies. As more decarbonisation technologies are being introduced, ranging from renewable energy systems to electric vehicles, tracking their adoption, usage patterns, and infrastructure deployment becomes increasingly important. However, many datasets on renewables and electric vehicle ownership remain incomplete or inconsistent. While a few major cities may provide harmonised multi-year datasets, most urban areas still lack unified or publicly accessible records [100, 101]. Addressing these gaps will be crucial to further advancing the potential of AI in urban system decarbonisation. 14 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Figure 9. Clustered word clouds for AI-driven methods applied in the literature. A detailed classification is shown in appendix table 2. 3.4. Methodological spectrum: AI-driven approaches in the literature Figure 9 provides an overview of the AI-driven methods employed in the reviewed literature and table 5 shows the application of major methods. The distribution of these methods indicates that ML dominates the field, with XAI emerging as a critical yet underutilised approach to address the limitations of black-box models. 3.4.1. ML ML is a technique that enables computers to identify patterns, make precise predictions, and improve performance through iterative learning without requiring explicit programming [34]. Among the reviewed literature, ML emerged as the most widely adopted approach (n= 55), with random forest (RF), K-means clustering, and light gradient boosting machine (LightGBM) identified as the most widely used algorithms. Specifically, RF is one of the most widely used foundational models, used in 36 studies. As a ML method capable of addressing complex non- linear data relationships, RF is distinguished by its relatively low computational and hardware requirements compared to more sophisticated models and its robust predictive performance. Therefore, RF is widely applicable in fields such as the prediction of carbon emissions, energy usage [49, 59, 84] and the solar radiation potential [72, 87]. Furthermore, RF serves as a critical base model for XAI variable importance analysis [50, 76]. Compared to RF, LightGBM excels in handling high-dimensional sparse data, capturing intricate nonlinear relationships, and addressing time-sensitive forecasting tasks due to its gradient-boosting-based optimisation mechanism, efficient leaf-wise growth strategy, and histogram-based computation. These features enable LightGBM to outperform RF in building energy consumption simulation tasks with superior predictive accuracy [102–104]. K-means are used primarily for data clustering and pattern recognition. By categorising regional characteristics [105], operational patterns [65], building attributes [19, 106], and behavioural patterns [67], it simplifies subsequent prediction and optimisation by ensuring consistency of input data. In addition, it facilitates the stratification of outcomes, constructing clear hierarchical frameworks for policy formulation tailored to different categories [107]. ML’s ability to efficiently handle nonlinear relationships and recognise patterns has led to its widespread application. Compared with DL, ML models, such as tree- based algorithms, often offer greater interpretability, which is valuable in contexts where understanding model decisions is important. 3.4.2. DL In the reviewed literature, 38 studies employed DL techniques, with ANNs, convolutional neural networks (CNNs), and recurrent neural networks (RNNs) identified as the three foundational models, cited with decreasing frequency. Long- and short-term memory (LSTM), an advanced variant of RNN, was also frequently used. 15 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Table 5. Application of major AI-driven methods in urban system decarbonisation. Method category Algorithm Urban system Application Examples MF RF Buildings, energy supplies, transport networks and infrastructure, open spaces, waste management systems Prediction of building energy use (annual/hourly), energy use intensity (EUI) modelling, construction-phase emissions estimation, anomaly detection, analysis of urban form and carbon storage dynamics at city scale, solar radiation potential estimation [49, 59, 72, 84, 87] LightGBM Buildings, energy supplies, open spaces Building energy consumption and EUI prediction (hourly/yearly), GHG emissions modelling, solar radiation forecasting [102–104] K-means Clustering Buildings, energy supplies, transport networks and infrastructure, open spaces Clustering of regional characteristics, building attributes, operational and behavioural patterns; classification of urban form and energy performance; enhancement of prediction models, policy stratification based on consistent input patterns [19, 65, 67, 105–107] DL ANNs Buildings, energy supplies, transport networks and infrastructure, open spaces Energy demand prediction, thermal load estimation, carbon emissions forecasting, solar energy generation analysis, energy cost modelling, policy impact assessment, multi-objective optimization for urban layout and energy planning [62, 75, 108, 109] CNNs Buildings Image-based roof extraction, solar potential assessment, building classification, energy efficiency prediction via streetview, carbon emissions estimation from building layouts, identification of hard-to-decarbonize buildings [52, 70, 74, 110–112] RNNs Energy supplies, transport networks and infrastructure Time-series forecasting of energy load and traffic flow, smart grid optimization, temporal pattern learning for IoT-enabled systems [63, 98, 106, 113–115] LSTM Buildings, energy supplies, transport networks and infrastructure Short- and long-term forecasting of energy demand, electricity consumption, and building thermal dynamics; spatiotemporal pattern learning for smart grids and smart cities [54, 63, 114, 116] OAs GA Buildings, energy supplies Data selection optimisation, sample size reduction, multi-objective tradeoffs (e.g. thermal comfort vs emissions) [54, 58, 77, 117, 118] PSO Buildings, transport networks and infrastructure Model parameter optimization for traffic flow and building energy load forecasting [20, 57, 119] XAI SHAP Buildings, energy supplies, transport networks and infrastructure Explains how each input feature affects model predictions (positive/negative direction and magnitude), identifies dominant drivers of energy use or emissions, detects unreliable or biased models, supports variable selection and policy prioritization [24, 55, 56, 64] (Continued.) 16 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Table 5. (Continued.) LIME Buildings Explains EUI predictions at the individual building level, identify key influencing factors (e.g. property type) [120] CAM Buildings Visually interprets which building features (e.g. roofs, windows) drive model predictions in image-based energy classification tasks [70] ANNs, the most fundamental neural network model (n= 13), mimic the structure of biological neural networks, employing feedforward and backpropagation to train weights. In the reviewed articles, ANNs were applied to predict urban energy consumption and thermal demand [75, 108, 109], as well as carbon emissions [62]. CNNs were widely used for image data processing (n= 8). For satellite imagery, CNNs were used to extract roof contours, assess the status and potential of solar installations [74, 110], or extract building characteristics for classification and evaluation of energy consumption [52]. In addition, CNNs analysed residential floor plans to map layout characteristics to carbon emissions [111]. Streetview imagery was another significant data source, with CNN extracting facade features to predict and optimise EE [70, 112]. RNNs excel at capturing dependencies in time-series data. In the reviewed studies, RNNs processed time-series energy load data for dynamic forecasting [98, 113] and traffic flow time-series data for prediction [80]. To address gradient vanishing in traditional RNNs when handling long sequences, LSTMs were developed to capture long-term dependencies. LSTMs were widely applied in forecasting hourly energy consumption, particularly in prediction of electricity demand [63, 106, 114, 115]. DL methods are particularly advantageous for analysing high-dimensional, unstructured, and multimodal data, such as images and time series datasets, thereby establishing their critical role in urban-scale research. 3.4.3. Optimization algorithms OAs play a fundamental role in many ML methodolo- gies, where surrogate models are constructed and optimised to achieve optimal perfor- mance by maximising an objective function or minimising a loss function [121]. Among the reviewed studies, 10 applied OAs, with genetic algorithms (GA) and particle swarm optimisation (PSO) being the most frequently employed. GA is a search and OA inspired by natural selection and genetics, which incorporates selection, crossover (recombination), and mutation. GA was used in urban energy consumption prediction to optimise data selection, reduce sample size [117], and improve model performance for energy consumption classification and forecasting [58, 118]. It also addressed multi-objective problems, such as balancing volume minimisation with maximising solar absorption [77], and resolving integrated optimisation issues involving energy consumption, carbon emissions, and thermal comfort [54]. PSO, a swarm intelligence-based metaheuristic OA, simulates the cooperative behaviour of particles within a group to identify global optima. PSO improved the accuracy of the prediction of traffic flow [20] and improved the accuracy of the prediction of building energy loads [57, 119] by optimising model hyperparameters. 3.4.4. XAI XAI encompasses techniques and methods that enhance the transparency and interpretability of AI systems’ decision-making processes. In the context of energy consumption prediction and management, white-box, grey-box, and black- box models are commonly employed. White-box models effectively elucidate causal relationships and system complexities, offering fundamental insights into heat- and mass-transfer mechanisms. Grey-box models, often grounded in building simulations, partially reveal the intricate relationships between various building parameters and energy consumption. Black-box models, represented by ML and DL approaches, are frequently used in the reviewed literature. These models address key limitations of white- and grey- box approaches, such as the requirement for detailed building data and high computational costs, making them efficient tools for urban research [55]. However, black-box models are often criticised for their lack of interpretability. XAI techniques improve the transparency and credibility of these models by elucidating their prediction processes. In the reviewed articles, 31. 87% (n= 29) of the studies incorporated XAI methods, with SHapley additive explanations (SHAP), local interpretable model-agnostic explanations (LIME) and class activation maps (CAM) applied. XAI, rooted in cooperative game theory, quantifies the contribution of each feature to 17 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath the output of a model by evaluating all possible combination of features. For example, Alvarez-Sanz et al [55] used SHAP to pinpoint critical factors influencing heating demand, helping to design and retrofit the targeted building. In addition to illustrating the contribution of the factor, Amiri et al [24] demonstrated the utility of SHAP to reveal unreliable models where seemingly irrelevant features disproportionately influence output. CAMs improve spatial interpretability by identifying regions in input images that most influence the predictions of a model, making them particularly effective for CNNs. Sun and Bardhan [70] applied CAMs to streetview images, highlighting elements such as roofs and windows that significantly impact EE predictions for hard-to- decarbonise buildings, providing valuable insights for urban-scale policy decisions. LIME creates simplified local models around specific predictions by perturbing input data and analysing output changes. Jin et al [120] employed LIME to investigate the intensity of energy use in individual buildings, offering customised planner recommendations to improve EE. These XAI methods enhance the transparency, reliability, and applicability of complex AI models in various domains. 3.5. Synthesis: applications of AI in urban decarbonisation AI-driven approaches in urban decarbonisation have made significant contributions in multiple areas. Based on current literature, these contributions include: (a) forecasting carbon emissions, (b) identifying factors influencing energy consumption, and (c) developing data-driven strategies for low-carbon urban planning. 3.5.1. Forecasting carbon emissions Current studies have contributed to carbon emission predictions across diverse dimensions, including prediction targets and spatial and temporal scales, with high model accuracy. Current predictive models enable highly accurate prediction of electricity, heating and cooling demand, as well as overall energy demand, carbon emissions and carbon footprints at the building, community, district, city and county levels across multiple timescales, from minute-by-minute to annual predictions. This objective represents the contribution most studied in the literature and forms the foundation for subsequent efforts to identify influencing factors and develop actionable solutions for urban decarbonisation. 3.5.2. Identifying factors influencing energy consumption A detailed understanding of the factors influencing energy consumption is fundamental for diagnosing carbon emission sources and informing targeted mitigation efforts. Existing studies have explored these factors on different scales. At the building scale, research has extensively examined the contributions of building physical characteristics, environmental and climate factors, energy usage patterns, and user behaviour to energy consumption. Among the physical characteristics of the building, factors such as building age, gross floor area, and construction period consistently emerge as the most significant determinants [24, 53, 103, 108, 122]. In terms of environmental and climate factors, the climate zone and the temperature of the outdoor air are repeatedly identified as the most critical variables [56, 82, 83, 104]. Regarding energy usage patterns, the performance of the HVAC system and the heating technologies are highlighted as dominant contributors to energy demand [58, 83, 108]. On the urban scale, spatial morphology such as building density, sky view factor, road density, and land use mix are the most influential determinants of energy consumption [64, 123, 124]. Environmental and climatic conditions remain crucial, with weather variability and compactness of green spaces frequently highlighted due to their impact on urban carbon emissions [60, 115]. Socioeconomic dimensions, particularly population density and income levels, are consistently recognised as central factors influencing energy use [59, 64, 115, 125]. In addition, ecological preservation scenarios and cross-sector information integration have been shown to be effective measures for reducing emissions [65, 126]. Although AI-driven methods have been widely used to explore the effects of physical, climatic, and socioeconomic factors on urban energy consumption, the role of policy-related factors remains underexamined in such frameworks. Callaghan et al [127] underscore the importance and growing research attention to climate policy instruments across sectors. Yet, few studies have employed AI-driven approaches to systematically assess how these policies shape urban energy demand and decarbonisation pathways [71, 73]. In addition, few studies have explored the interdependencies between different influencing factors. Various components of urban systems often exhibit complex interrelationships, and their collective contributions to carbon emissions remain underexplored. 3.5.3. Developing data-driven strategies for low-carbon urban planning Developing data-driven strategies enables precise planning and actionable guidance for urban decarbonisation. Current research offers tools and frameworks for low-carbon urban management. For example, a spatial assessment tool for solar photovoltaic installations was developed by Chen et al [72], who combined a physical solar radiation calculation model with an ML model to predict the solar radiation 18 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath potential on rooftops and external walls across a city. Skiba et al [71] provided a tool for evaluating investments in energy retrofit projects using ANNs to model the economic dependency between urban development policies and investments in EE. Excell et al [66] developed an energy retrofit simulation tool by integrating parametric simulations with ML, offering detailed guidance on the scope of energy retrofit projects. Despite these advancements, existing approaches often focus on isolated objectives, lacking integrated frameworks that consider multi-objective trade-offs in real-world scenarios. Strengthening these frameworks and incorporating broader socio-economic and environmental factors would enhance the practical relevance and effectiveness of data-driven strategies. 3.6. Synthesis: AI-drivenmethods for real-world applications With regard to RQ 2, in order to examine whether the AI-driven methods employed in the reviewed literature are capable of reliably and adequately supporting real-world decision-making, this study conducts an evaluation across model performance, model generalizability, explainability, and policy recommendations. This approach aims to assess the extent to which current research bridges the gap between theoretical development and practical application, while also identifying potential directions for future methodological refinement. 3.6.1. Model performance and generalizability For model performance, the reported accuracies generally exceed 70%, with approximately 25% of the research achieving accuracies of 90% or more and eight studies demonstrating accuracies higher than 98%. As shown in table 3, various climate action objectives and different types of AI-driven models overall exhibit relatively high levels of performance. Most studies also adopt multiple rounds of cross-validation, sensitivity analysis, and uncertainty quantification to evaluate model robustness, thereby ensuring consistency and reliability of the results under different samples or parameter settings. AI methods possess clear advantages in modelling complex nonlinear relationships, often yielding more accurate predictions than traditional techniques. This forms the fundamental guarantee for real-world applications, as the policy-making process is heavily reliant on the accuracy and stability of model outputs. The generalizability of models is a key criterion in assessing the real-world applicability of data-driven approaches [128]. Currently, the majority of carbon mitigation studies are based on specific urban contexts, rendering their models highly context- dependent and difficult to generalise. In contrast, models with strong generalizability can adapt to varying urban structures, energy systems, and policy environments, and thus offer greater practical relevance. The carbon transition is influenced by regional development levels, environmental regulations, and industrial structures [76]. In addition, building types and climate zones are considered crucial because different types of buildings exhibit varying patterns of energy usage in climate zones [16]. Thus, the generalizability and reliability of these models can only be validated when tested in countries with different conditions. Moreover, for urban decision-makers, the construction of locally tailored high-accuracy models is often prohibitively expensive and difficult to replicate. Therefore, models with cross-regional applicability are particularly valuable. In this review, approximately 48.35% of the reviewed literature (n= 44) highlighted the absence of cross-regional validation. Nevertheless, there are several noteworthy cases. Some studies have used approaches such as transfer learning, training models in data-rich regions before applying them to data-scarce areas [54, 129] and resampling methods for specific building types [120]. These strategies offer feasible research pathways for enhancing model generalizability and enabling studies in data-scarce developing countries. 3.6.2. Explainability and policy recommendations For real-world applications, a key requirement is the ability to provide clear and actionable policy recommendations for different types of decision-makers. Table 6 presents an overview of the stakeholders and types of policy recommendations discussed in the existing literature. The findings reveal that most studies primarily target governments (69 articles) and enterprises (81 articles), offering technically-oriented recommendations such as energy forecasting, improvements in building EE, smart grid optimisation, and renewable energy integration. These recommendations are typically derived from AI models that simulate complex relationships among multiple variables, thus requiring substantial technical expertise. As a result, they are often not readily comprehensible or actionable for policy recipients without technical backgrounds. More critically, such recommendations tend to serve the optimisation goals of the research itself, such as enhancing predictive accuracy or improving EE, rather than being directly aligned with the actual tasks and priority objectives of real-world decision-makers. In practice, government departments are often more concerned with identifying key variables that influence policy performance and understanding their boundaries of applicability to support scenario simulations and resource allocation [24, 56, 103]. Enterprises, in comparison, tend to focus on the operational applicability of model outputs, such as the 19 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath feasibility and prioritisation of energy retrofitting strategies [66]. Therefore, encouraging the development of models that are designed and interpreted from the perspective of practical needs would significantly enhance the real-world applicability of AI-driven research. In this context, the importance of XAI methods has become increasingly pronounced. Techniques such as SHAP enable attribution of positive or negative contributions of individual input features to the model output, thereby supporting stakeholder-specific communication of model insights. However, according to this review, only 31.87% of the literature (n= 29) employed XAI techniques to interpret model mechanisms. As shown in table 6, this issue is particularly acute among studies using DL methods, in which explainability analysis is rarely applied. Previous research has highlighted the inherent trade-off between model complexity and interpretability—the more complex a model, the less transparent its reasoning process tends to be [22]. Furthermore, from a stakeholder perspective, the current literature demonstrates a clear lack of attention to affected individuals: only a very small number of studies have involved community representatives (5 articles) or the public (4 articles). Yet these groups are precisely those in most need of clear and intuitive explanations. They often lack the technical background to understand model structures, but have a greater need to comprehend, in accessible terms, how individual behaviours translate into carbon mitigation outcomes [59]. Future research should therefore prioritise the development of interpretative and recommendation mechanisms oriented towards the general public and communities, for instance, by incorporating behavioural and socio-economic modelling frameworks, supported by intuitive visualisation and interactive presentation techniques to enhance the comprehensibility and actionability of policy recommendations. In sum, while AI-driven research has demonstrated strong capabilities in terms of model performance, there remain notable limitations in generalizability, interpretability, stakeholder adaptability, and recommendation mechanisms tailored to real-world policy- making. 4. Discussion 4.1. The pathway for using AI for urban system decarbonisation Big data and AI-driven approaches significantly improve efficient and multidimensional low carbon decision making in urban settings. These methods provide highly accurate tools for forecasting carbon emissions. In this way, key factors influencing emissions can be identified and specific improvement guidelines are recommended. These methods also monitor building energy consumption, integrate renewable energy, develop sustainable transportation, and manage urban carbon emissions comprehensively. Such efforts contribute to multiple SDGs, including SDG 11 (Sustainable Cities and Communities), SDG 13 (Climate Action) and SDG 3 (Good Health and Well-Being), thereby promoting a sustainable future. The implementation pathway is clear through data-driven decision-making. For example, data on energy, buildings, environment and socioeconomic factors can fundamentally support carbon emission forecasting and management, thereby forming the basis of low-carbon decisions. As described earlier, the ML and DL models can uncover complex data patterns to support accurate emissions predictions and trend analyses. XAI helps decision makers understand critical influencing factors, whereas OA increases efficiency across these pathways. This combination of data insights and intelligent analysis methods can provide an actionable path to achieving low carbon urban environments. Compared with traditional empirical methods, data-driven pathways are increasingly seen, which offers great adaptability to complexity, dynamic resilience, and computational efficiency. This approach results in precise and effective urban carbon management. 4.2. Scope, objective and data challenges Big data and AI approaches have limitations that prevent them from fully aligning with real-world complexities. A summary of the gaps identified from the reviewed literature highlights the following. (1) Research objectives: For one thing, research focussing on multi-objective optimi- sation, including considerations of environmental, economic, and social sustainability, is lacking. This limitation restricts the application of research findings in complex real-world scenarios. Thus, avoiding potential conflicts of interest becomes difficult. In contrast, few studies provide a view of interactions between different decision-making sectors. In prac-tice, the influence of different decision sectors often overlaps—for example, adjustments in energy policies can affect electric vehicle usage, potentially influencing urban transporta- tion planning. Considering these synergies is crucial to ensure that research outcomes serve practical applications effectively. Currently, only a limited number of studies ac- count for the multi-objective synergy between emission reduction and financial benefits [46, 68, 80]. 20 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Table 6. Summary of policy recommendations from the reviewed literature categorized by stakeholders. Stakeholder category Specific role Policy recommendation Models used Number of papers Examples Policy makers Governments Develop urban energy forecasting and carbon emission models, optimise building design and urban form, pro-mote integration of smart grids and renewable energy, provide energy ef-ficiency regulations and labelling sys-tems, and support the use of AI and GIS in policy-making. ML: RF, XGBoost, 69 [79, 98, 102] SVR, SVM, ANN, KMeans, Light- GBM; DL: LSTM, CNN, GRU, DNN, RNN, TCN; XAI: SHAP International organiza-tions Promote standardisation of energy systems and cross-border knowledge transfer, develop global guidelines on smart transport and green roofs, and support data sharing and cross-city validation of energy efficiency methods. ML: RF, XG- 12 [73, 74, 78] Boost, SVR, SVM, KMeans; DL: CNN, U-Net, DNN, ANN; XAI: SHAP Action tak-ers Enterprises Implement AI-based optimisation and prediction of building energy effi-ciency, adopt energy-saving and re-newable technologies, improve opera-tional scheduling and equipment main-tenance, and introduce contactless au-diting and IoT management systems. ML: RF, XGBoost, 81 [108, 111, 122] ANN, SVR, Light- GBM, SVM, KNN, CART; DL: LSTM, CNN, GRU, DNN, TCN, RNN; XAI: SHAP Government implement-ing agencies Deploy IoT and real-time data sys-tems for energy monitoring, promote energy-efficient retrofitting of build-ings, manage urban expansion and transport systems for emission reduc-tion, and implement forecasting mod-els to support energy allocation. ML: RF, XGBoost, 11 [60, 73, 113] LightGBM, KNN, DT; DL: RNN; XAI: SHAP (Continued.) 21 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Table 6. (Continued.) Stakeholder category Specific role Policy recommendation Models used Number of papers Examples Affected in-dividuals Community representa-tives Promote community-level energy-saving behaviours, guide participation in building energy-efficient design, or-ganise energy awareness activities, and support local green energy projects. ML: RF, XGBoost; 5 [56, 67, 76] XAI: SHAP The public Participate in energy-saving schemes, increase awareness of energy efficiency, adopt high-efficiency appliances, ad-just household energy usage, and pay attention to carbon footprints and en-vironmental impacts. ML: RF, XGBoost; 4 [67, 76] XAI: SHAP (2) Data: Significant information gaps exist in carbon emissions, behaviour, technology, spatial coverage and temporal coverage. Addressing these gaps can enable a complete understanding of carbon emission sources and potential mitigation strategies. The coverage and interpretative power of data on real-world phenomena remain substantial barriers to AI for decarbonisation. (3) AI-driven methods: First, limited attention to generalizability may undermine the reliability of AI model predictions when applied beyond their original contexts. Most existing studies are based on data from the Global North, while those focusing on the Global South remain scarce. As a result, the applicability of many models to the complex and diverse contexts of the Global South remains uncertain. Research on cross-regional transferability is still limited [54, 74, 78, 129]. Second, the integration of XAI remains limited in current literature. Although many studies aim to support government decision- making, the lack of XAI often leads to insufficient transparency regarding technical processes and contextual suitability. This opacity may hinder effective communication between technical experts and policymakers, ultimately limiting the uptake and practical use of model outputs in policy formulation. 4.3. Gaps in XAI for urban climate action Given the increasing impacts of climate change, we must consider the real-world applicability of these accurate models. Future work sections in relevant studies reveal gaps that prevent full reliance on big data and AI-driven technologies for low-carbon urban decision-making: (1) Trustworthiness: High accuracy alone does not guarantee trust. Decision-makers and the public often distrust black-box models because they cannot understand the underlying decision-making logic. (2) Transparency: Complex ML and DL models lack transparent decision pathways. Thus, identifying the factors driving the changes in carbon emissions and implementing targeted improvements become challenging. (3) Fairness: Low-carbon decisions must account for diverse regions and socioeconomic groups. However, data availability often varies, potentially leading to biased model outcomes that undermine the inclusion of decisions. (4) Interpretability for multiple decision makers: Climate policies require cooperation amongst multiple decision- makers. In addition, communicating model mechanisms clearly to all stakeholders involved in collaborative decision-making in real-world scenarios is crucial. (5) Precise management: Real-world governance requires precise interventions to address complex and evolving situations, making enhancing the temporal and spatial precision of decision- making essential. XAI can address these issues to some extent by clarifying decision paths and interaction mechanisms amongst different sectors and highlighting model biases. However, as shown in this article, XAI applications remain limited, which restricts the progress of low-carbon urban decision-making. These challenges can be addressed by fully integrating XAI into future urban decarbonisation efforts. This approach can facilitate an efficient and trustworthy future. The key areas for focus include the following. 22 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath 4.3.1. Trustworthiness XAI can reveal the contributing factors behind the model results, allowing experts to assess the reliability of the model by assessing the reasonableness of these contributions based on theoretical knowledge. If experts identify unreasonable variable weights, then they can adjust these variable weights based on their expertise and relevant standards; as a result, a reliable model can be constructed [24]. 4.3.2. Transparency XAI can uncover the key factors driving carbon emissions within urban systems. In this way, planners can help design targeted policies. For example, in the building sector, XAI can pinpoint the primary factors contributing to carbon emissions, such as building density and floor area ratio, and maximum limits for these metrics can be suggested to avoid excessive energy consumption; as a result, policy makers can receive scientifically grounded numerical references for planning [69]. 4.3.3. Fairness Poor-quality data produce poor-quality results. Many underdeveloped cities or regions lack quality data to input into models. This scenario can lead to poor representativeness and biased output. XAI can help mitigate this issue in three ways. Firstly, the use of local explanations (such as LIME) can detect whether the behaviour of the model is biased toward different subjects, which allows a fair assessment to some extent [130]. Secondly, equitable data sources can be used in conjunction with their XAI tools, such as satellite images; these sources have become a valuable substitute for traditional data sources in countries and regions struggling with a lack of data, as well as expansive rural areas [131]. The emergence of image-based XAI techniques (e.g. CAM) has facilitated the interpretation of satellite-derived predictions, thereby providing an equitable modelling approach for many under-represented areas. Third, XAI can help assess the geographic transferability of models or policies. XAI can precisely locate the key influencing factors on which a model relies; this approach aids in determining whether the regions with similar characteristics can effectively adopt these models [132] or whether regions with similar influencing factors can share improved policies [133]. These three approaches can help address the lack of research and guidance for many countries and regions worldwide, thereby promoting fair decarbonisation research worldwide. 4.3.4. Interpretability for multiple decision-makers XAI has the potential to provide explanations to each stakeholder in the decision-making chain, thereby fostering collaboration. It enables developers and researchers to gain an in-depth understanding of the mechanisms of the AI model, allowing for algorithmic optimisation and performance enhancement. Subsequently, it enables policymakers to understand the basis of decisions and evaluate the reliability of the outcomes. For intervention implementers, such as urban planners and architects, XAI provides insight into the underlying logic of the model, allowing them to adjust strategies based on specific requirements. Finally, XAI supports people affected by decisions, such as the public, in understanding, trusting, and seeking recourse for these decisions [26]. 4.3.5. Precise management XAI can reveal the interdependencies between variables, thereby offering refined policy recommendations. For example, XAI can uncover the variable interdependencies on household electricity consumption. In economically developed and highly urbanised areas, an increase in per capita income and urbanisation leads to high electricity usage; however, in economically underdeveloped and less urbanised areas, increased income may not necessarily be directed towards increasing electricity demand [134]. Such insights can guide the development of differentiated electricity and subsidy policies for areas with varying economic levels, rather than applying uniform policies. From a temporal perspective, XAI can reveal how the influence of factors of the built environment and land use on transportation carbon emissions changes over time. This insight leads planners to implement dynamic management strategies [23]. Thus, XAI facilitates the transition towards precise, spatiotemporal and dynamic management. 4.4. Contributions of this review Existing reviews on the application of AI in climate change contexts have rarely focused specifically on urban systems. This paper offers an innovative and comprehensive review of urban system decarbonisation supported by AI-driven methods, aiming to clarify the pathways, applicability, and future directions of AI in supporting urban climate action. Compared with previous studies, the key contributions of this paper are as follows: (1) This review adopts a more integrative perspective on AI-driven methods by incorporating both XAI and OAs, with a particular emphasis on XAI as a critical enabler for real- world applications. Previous reviews on AI-driven methods for climate action have mostly focused on ML and DL techniques [33, 36]. (2) 23 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Table 7. Summary of contributions of this review. Dimension Contribution of this review Limitations of previous reviews Methodological perspective Incorporates XAI and OAs Predominantly focused on ML and DL, neglecting XAI and real- world compatibility Three- dimensional pathways Summarises ‘goals–methods–data’ frame- work Lacks systematic connection and detailed application of specific AI methods Evaluation of real-world applicability Includes generalizability, explainability, and policy interface analysis; emphasises the importance of explainability for real-world application Rarely discusses this critical perspective Identification of research gaps Highlights research gaps in objectives, data, and AI-method applications Missing discussions on multi- objective optimisation, cross-sectoral interaction, data gaps, and insufficient XAI adoption This review synthesises various climate action goals–AI methods–data pathways. This framework enhances researchers’ ability to select appropriate methodologies and design more effective studies. Earlier reviews [30] have seldom provided a systematic explanation of how different types of data and AI methods serve distinct climate objectives, lacking in-depth clarification of the interconnections among the three. (3) This review systematically evaluates the alignment between existing AI models and real-world needs, particularly in terms of accuracy, generalizability, explainability, and policy interfaces. This practical lens has often been neglected in prior reviews [28, 29], yet it is crucial for the on-ground application of AI in climate action. (4) This review identifies major research gaps across three dimensions: research objectives, data availability, and the application of AI methods. These findings provide valuable guidance for future studies. Table 7 presents a clear summary of these contributions. 4.5. Limitations of the literature review While this review provides a comprehensive synthesis of AI applications for urban system decarbonisation, several limitations should be acknowledged. First, this study relied exclusively on the WoS for literature retrieval. Although WoS maintains rigorous indexing standards and offers high metadata reliability, it may not capture all relevant publications. Future research could address this limitation by integrating additional bibliographic databases such as Scopus or Google Scholar, which are known to offer broader disciplinary and linguistic coverage. Second, while the keywords were carefully selected to reflect major themes, the search strings may not have captured all relevant variations in terminology. For example, some studies may have used forms like ‘low carbon’ without a hyphen or other wording not included in the query. This could have led to the exclusion of relevant papers. Future reviews may improve coverage by applying fuzzy search techniques or expanding synonyms during the search process. Finally, this review only included English-language publications. As a result, studies published in other languages or in grey literature may have been missed. Including multilingual sources and broader databases in future reviews could help provide a more complete global perspective. 5. Conclusion Rapid decarbonisation is important on a global scale, and cities provide important avenues of reducing emissions. AI technologies provide efficient tools for reducing carbon emissions at the urban scale by integrating and analysing massive datasets from various urban sectors. To explore how AI-driven approaches facilitate the development of low carbon urban systems and identify the gaps between these technologies and real-world decision- making needs, this research conducted a systematic review of the literature of 91 selected papers. Through statistical analysis of these studies, a roadmap for urban decarbonisation planning was outlined that takes advantage of big data and intelligent models. The main findings are as follows: For RQ 1: (1) Overall, AI-driven approaches show significant potential in forecasting carbon emissions, identifying drivers of energy consumption, and supporting the development of low-carbon planning strategies. (2) In terms of objectives, building energy consumption prediction is the most frequently studied topic, with nearly half of the reviewed literature focusing on this climate-action- related goal. However, research objectives need to be more multi-sectoral in order to better address the complex and interconnected nature of real-world urban systems. (3) Regarding data used, public databases 24 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath and data from government or public departments are the most commonly utilised sources. Nevertheless, there remain substantial data gaps in carbon emissions, behavioural patterns, technological attributes, spatial coverage, and temporal coverage, which limit the comprehensiveness of existing studies. (4) Among AI-driven methods, ML is more widely applied, particularly RF for energy prediction and K-means clustering for identifying energy consumption patterns. However, the limited adoption of XAI restricts the interpretability of these models. For Research Question 2: (1) In terms of model performance, current studies generally achieve high accuracy, but considerations of generalizability remain insufficient. (2) With regard to policy recommendations, most studies primarily provide technical suggestions to governments and enterprises, such as how to accurately predict energy consumption or improve low-carbon building design. However, there is a lack of explanation regarding contextual applicability, optimal indicator prioritisation, and model reliability, which contributes to the gap between current research and real-world applications. In addition, studies focusing on affected individuals, such as the general public, are still limited, which hinders a more comprehensive consideration of diverse stakeholders. This study highlights that the lack of model explainability undermines the practicality and adoption of these approaches. Greater implementation of XAI is essential to enhance trustworthiness, transparency, and fairness, while also improving interpretability for diverse stakeholders and enabling precise urban management. It is crucial to recognise that AI technologies themselves contribute to carbon emissions. To improve the environmental sustainability of AI-driven research, future studies should prioritise practical, real-world applications that address key urban challenges. Ultimately, by integrating innovation with sustainability, AI has the potential to redefine the pathways towards a greener, more resilient future for cities worldwide. Data availability statement No new data were created or analysed in this study. Acknowledgment RD acknowledges the support of the Cambridge Humanities Research Grant (ALBN) and the CRASSH climaTRACES lab. 25 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Appendix Table A1. Data source classification standards. Category Data sources Government and public de- partments International Energy Agency (IEA), Wuhan Municipal Bureau of Natural Resources and Planning, Seattle Office of Sustainability and Environment, Planning Department of Hong Kong, Ministry of Land, Public Security Bureau Research institutions and universities China’s Statistical Institute, Chinese Academy of Sciences Resource and Environmental Science Data Center, University of Texas at Austin, LandScan Global Population Data (Oak Ridge National Laboratory), University of Florida campus utility data, AIMSUN platform simulations Business and industry Local electricity providers, Pacific Gas and Electric (California), Qatar General Electricity and Water Corporation (KAHRAMAA), Red El´ectrica de Espan˜a, Consolidated Edison (ConEd), National Grid, Ausgrid Public databases World Bank databases, EXIOBASE and ecoinvent databases, OpenStreetMap (OSM), census block data from New York City, open-source data inventory for anthropogenic CO2 (ODIAC), Geneva’s open data repository (SITG), NASA NEX-GDDP dataset Experimental and simulation data Ladybug plugin for Grasshopper, EnergyPlus simulation data, Swedish Energy Performance Certification (EPC) dataset, Ecotect simulation software, Agent-based simulation frameworks, Dynamo simulation data Table A2.Method classification standards. Category Methods Machine learning AdaBoost, AdaBoost Regressor, Gaussian naive Bayes, K-means clustering, extra trees regressor, random forest classifier Deep learning Artificial neural network (ANN), back propagation neural network (BPNN), convolutional neural network (CNN), long short-term memory (LSTM), deep neural network (DNN) Optimization algorithms Particle swarm optimization (PSO), grey wolf optimizer (GWO), genetic algorithm (GA), non-dominated sorting genetic algorithm II (NSGA-II), Adam optimizer Explainable AI Feature importance, sHapley additive exPlanations (SHAP), class activation maps (CAM), local interpretable model-agnostic explanation (LIME), partial dependence plots (PDPs) 26 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath Table A3. Indicative expressions used for fuzzy extraction of ‘datasets used’, ‘main findings’, and ‘research gaps’. Target field Indicative expressions/patterns Datasets used The dataset used in this study…; We collected data from…; The study relies on data from…; Data were obtained from…; The data source is…; We used publicly available data…; The dataset is available at…; This work uses data provided by…; Open-source data from…was used; Experimental data were collected from…; Data were retrieved from…; We made use of datasets compiled by…; Our empirical analysis is based on data from…; The analysis is conducted using data sourced from…; The research utilises data published by…; This paper employs data extracted from…; Data used in the analysis include… Main findings The main findings are…; Our results show that…; This study demonstrates that…; The key contribution of this work is…; It was found that…; The results indicate that…; This study provides evidence that…; We conclude that…; Empirical results suggest that…; The findings support the hypothesis that…; We have shown that…; The outcomes reveal…; The research confirms…; This study establishes a link between…; Our findings highlight…; The paper concludes by… Research gaps/fu-ture work This study has several limitations…; Future work will focus on…; Further research is needed to…; One remaining challenge is…; This paper does not address…; An unanswered question remains…; We plan to explore…; The following aspects warrant further investigation…; Opportunities for future studies include…; This research opens new directions for…; There is a need for more research on…; Further investigation should consider…; It remains unclear whether…; This work could be extended by…; Limitations of the present study include…; It would be beneficial to examine… ORCID iDs Tong Chen https://orcid.org/0000-0002-8242-2412 Ramit Debnath https://orcid.org/0000-0003-0727-5683 References [1] Masson-Delmotte V et al 2018 Global Warming of 1.5 ◦C. An IPCC Special Report on the Impacts of Global Warming of 1.5 ◦C above Pre-industrial Levels and Related Global Greenhouse Gas Emission Pathways, in the Context of Strengthening the Global Response to the Threat of Climate Change, Sustainable Development, and Efforts to Eradicate Poverty (Intergovernmental Panel on Climate Change (IPCC)) (available at: www.ipcc.ch/sr15/) [2] Ritchie H, Roser M and Rosado P 2023 CO2 and greenhouse gas emissions (available at: https://ourworldindata.org/co2- emissions) (Accessed 13 December 2024) [3] International Energy Agency 2021 Empowering Cities for a Net Zero Future Licence: CC BY 4.0 (International Energy Agency) (available at: www.iea.org/reports/empowering-cities-for-a-net-zero-future) [4] Kennedy C A, Ibrahim N and Hoornweg D 2014 Low-carbon infrastructure strategies for cities Nat. Clim. Change 4 343–6 [5] McHale M R et al 2015 The new global urban realm: complex, connected, diffuse, and diverse social-ecological systems Sustainability 7 5211–40 [6] Rus K, Kilar V and Koren D 2018 Resilience assessment of complex urban systems to natural disasters: a new literature review Int. J. Disaster Risk Reduct. 31 311–30 [7] Müller D B, Liu G, Løvik A N, Modaresi R, Pauliuk S, Steinhoff F S and Brattebø H 2013 Carbon emissions of infrastructure development Environ. Sci. Technol. 47 11739–46 27 https://orcid.org/0000-0002-8242-2412 https://orcid.org/0000-0002-8242-2412 https://orcid.org/0000-0003-0727-5683 https://orcid.org/0000-0003-0727-5683 https://www.ipcc.ch/sr15/ https://ourworldindata.org/co2-emissions https://ourworldindata.org/co2-emissions https://creativecommons.org/licenses/by/4.0/ https://www.iea.org/reports/empowering-cities-for-a-net-zero-future https://doi.org/10.1038/nclimate2160 https://doi.org/10.1038/nclimate2160 https://doi.org/10.3390/su7055211 https://doi.org/10.3390/su7055211 https://doi.org/10.1016/j.ijdrr.2018.05.015 https://doi.org/10.1016/j.ijdrr.2018.05.015 https://doi.org/10.1021/es402618m https://doi.org/10.1021/es402618m Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath [8] Kavehei E, Jenkins G A, Adame M F and Lemckert C 2018 Carbon sequestration potential for mitigating the carbon footprint of green stormwater infrastructure Renew. Sustain. Energy Rev. 94 1179–91 [9] Younger M, Morrow-Almeida H R, Vindigni S M and Dannenberg A L 2008 The built environment, climate change, and health: opportunities for Co-benefits Am. J. Preventive Med. 35 517–26 [10] Lu S, Qiu J, Bai L, Hu R and Xin K 2023 Carbon emissions of water supply systems in China: characteristic, right, right balance and reduction potential assessment J. Clean. Prod. 428 139446 [11] Baes Jr C F, Goeller H E, Olson J S and Rotty R M 1977 Carbon dioxide and climate: the uncontrolled experiment Am. Sci. 65 310–20 (available at: https://www.jstor.org/stable/27847841) [12] Zini M and Carcasci C 2023 Machine learning-based monitoring method for the electricity consumption of a healthcare facility in Italy Energy 262 125576 [13] Fay C D, Corcoran B and Diamond D 2024 Green IoT event detection for carbon- emission monitoring in sensor networks Sensors 24 162 [14] Nocera S, Ruiz-alarco´n-quintero C and Cavallaro F 2018 Assessing carbon emissions from road transport through traffic flow estimators Transp Res. C 95 125–48 [15] Liu Z et al 2020 Carbon Monitor, a near-real-time daily dataset of global CO2 emission from fossil fuel and cement production Sci. Data 7 392 [16] Jin X, Xiao F, Zhang C and Chen Z 2022 Semi-supervised learning based framework for urban level building electricity consumption prediction Appl. Energy 328 120210 [17] Ji T et al 2024 Urban transport emission prediction analysis through machine learning and deep learning techniques Trans. Res. D 135 104389 [18] Ghadami N, Gheibi M, Kian Z, Faramarz M G, Naghedi R, Eftekhari M, Fathollahi-Fard A M, Dulebenets M A and Tian G 2021 Implementation of solar energy in smart cities using an integration of artificial neural network, photovoltaic system and classical Delphi methods Sustain. Cities Soc. 74 103149 [19] Oh M and Kim Y 2019 Identifying urban geometric types as energy performance patterns Energy Sustain. Dev. 48 115–29 [20] Peng T, Yang X, Xu Z and Liang Y 2020 Constructing an environmental friendly low-carbon-emission intelligent transportation system based on big data and machine learning methods Sustainability 12 8118 [21] Ren Y, Zhou M, Zhu A, Shi S, Zhu H, Chen Y, Li S and Fan T 2024 Evolution, reconfiguration and low-carbon performance of green space pattern under diverse urban development scenarios: a machine learning- based simulation approach Ecol. Indicators 169 112945 [22] Minh D, Wang H X, Li Y F and Nguyen T N 2022 Explainable artificial intelligence: a comprehensive review Artif. Intell. Rev. 55 3503–68 [23] Wu J et al 2023 Spatiotemporal analysis of built environment restrained traffic carbon emissions and policy implications Trans. Res. D 121 103839 [24] Amiri S, Mueller M and Hoque S 2023 Investigating the application of a commercial and residential energy consumption prediction model for urban planning scenarios with machine learning and Shapley additive explanation methods Energy Build. 287 112965 [25] Linardatos P, Papastefanopoulos V and Kotsiantis S 2021 Explainable AI: a review of machine learning interpretability methods Entropy 23 18 [26] Amarasinghe K, Rodolfa K T, Lamba H and Ghani R 2023 Explainable machine learning for public policy: use cases, gaps, and research directions Data Policy 5 e5 [27] Sharifi A, Allam Z, Bibri S E and Khavarian-Garmsir A R 2024 Smart cities and sustainable development goals (SDGs): a systematic literature review of co-benefits and trade-offs Cities 146 104659 [28] Bibri S E, Krogstie J, Kaboli A and Alahi A 2024 Smarter eco-cities and their leading-edge artificial intelligence of things solutions for environmental sustainability: a comprehensive systematic review Environ. Sci. Ecotechnol. 19 100330 [29] Shaamala A, Yigitcanlar T, Nili A and Nyandega D 2024 Algorithmic green infrastructure optimisation: review of artificial intelligence driven approaches for tackling climate change Sustain. Cities Soc. 101 105182 [30] Yan F and Qi X 2024 Application and prospect of artificial intelligence technology in low-carbon cities—from the perspective of urban planning content and process Land 13 1834 [31] Page M J et al 2021 The PRISMA 2020 statement: an updated guideline for reporting systematic reviews BMJ 372 n71 [32] Thomason R H (ed) 1989 Philosophical Logic and Artificial Intelligence (Springer Book Archive 2012) 1st edn (Springer) (https:// doi.org/10.1007/978-94-009-2448-2) [33] Salem A H, Azzam S M, Emam O E and Abohany A A 2024 Advancing cybersecurity: a comprehensive review of AI-driven detection techniques J. Big Data 11 105 [34] Saranya A and Subhashini R 2023 A systematic review of explainable artificial intelligence models and applications: recent developments and future trends Decis. Anal. J. 7 100230 [35] Singh S et al 2024 Survey of AI-driven techniques for ovarian cancer detection: state-of-the-art methods and open challenges Netw. Mod. Anal. Health Inf. Bioinf. 13 56 [36] Ouyang F and Zhang L 2024 AI-driven learning analytics applications and tools in computer-supported collaborative learning: a systematic review Educ. Res. Rev. 44 100616 [37] Albahri A, Khaleel Y L, Habeeb M A, Ismael R D, Hameed Q A, Deveci M, Homod R Z, Albahri O S, Alamoodi A H and Alzubaidi L 2024 A systematic review of trustworthy artificial intelligence applications in natural disasters Comput. Electr. Eng. 118 109409 [38] Sarker I H, Furhad M H and Nowrozy R 2021 AI-driven cybersecurity: an overview, security intelligence modeling and research directions SN Comput. Sci. 2 173 [39] Wang S, Hu T, Xiao H, Li Y, Zhang C, Ning H, Zhu R, Li Z and Ye X 2024 GPT, large language models (LLMs) and generative artificial intelligence (GAI) models in geospatial science: a systematic review Int. J. Digit. Earth 17 2353122 [40] Abbas S, Asif M, Rehman A, Alharbi M, Khan M A and Elmitwally N 2024 Emerging research trends in artificial intelligence for cancer diagnostic systems: a comprehensive review Heliyon 10 e36743 [41] García-garcía A R, Jacobo-Hernández C A, Ochoa-Jiménez S and Valdez-del Río S 2023 Sustainable development and tourism: a review of the literature in WoS from 2001 to 2020 Sustainability 15 16805 [42] Sarkar A, Wang H, Rahman A, MemonWH and Qian L 2022 A bibliometric analysis of sustainable agriculture: based on the Web of Science (WOS) platform Environ. Sci. Pollut. Res. 29 38928–49 28 https://doi.org/10.1016/j.rser2018.07.002 https://doi.org/10.1016/j.rser2018.07.002 https://doi.org/10.1016/j.amepre.2008.08.017 https://doi.org/10.1016/j.amepre.2008.08.017 https://doi.org/10.1016/j.jclepro.2023.139446 https://doi.org/10.1016/j.jclepro.2023.139446 https://www.jstor.org/stable/27847841 https://doi.org/10.1016/j.energy.2022.125576 https://doi.org/10.1016/j.energy.2022.125576 https://doi.org/10.3390/s24010162 https://doi.org/10.3390/s24010162 https://doi.org/10.1016/j.trc.2018.07.020 https://doi.org/10.1016/j.trc.2018.07.020 https://doi.org/10.1038/s41597-020-00708-7 https://doi.org/10.1038/s41597-020-00708-7 https://doi.org/10.1016/j.apenergy.2022.120210 https://doi.org/10.1016/j.apenergy.2022.120210 https://doi.org/10.1016/j.trd.2024.104389 https://doi.org/10.1016/j.trd.2024.104389 https://doi.org/10.1016/j.scs.2021.103149 https://doi.org/10.1016/j.scs.2021.103149 https://doi.org/10.1016/j.esd.2018.12.002 https://doi.org/10.1016/j.esd.2018.12.002 https://doi.org/10.3390/su12198118 https://doi.org/10.3390/su12198118 https://doi.org/10.1016/j.ecolind.2024.112945 https://doi.org/10.1016/j.ecolind.2024.112945 https://doi.org/10.1007/s10462-021-10088-y https://doi.org/10.1007/s10462-021-10088-y https://doi.org/10.1016/j.trd.2023.103839 https://doi.org/10.1016/j.trd.2023.103839 https://doi.org/10.1016/j.enbuild.2023.112965 https://doi.org/10.1016/j.enbuild.2023.112965 https://doi.org/10.3390/e23010018 https://doi.org/10.3390/e23010018 https://doi.org/10.1017/dap.2023.2 https://doi.org/10.1017/dap.2023.2 https://doi.org/10.1016/j.cities.2023.104659 https://doi.org/10.1016/j.cities.2023.104659 https://doi.org/10.1016/j.ese.2023.100330 https://doi.org/10.1016/j.ese.2023.100330 https://doi.org/10.1016/j.scs.2024.105182 https://doi.org/10.1016/j.scs.2024.105182 https://doi.org/10.3390/land13111834 https://doi.org/10.3390/land13111834 https://doi.org/10.1136/bmj.n71 https://doi.org/10.1136/bmj.n71 https://doi.org/10.1007/978-94-009-2448-2 https://doi.org/10.1007/978-94-009-2448-2 https://doi.org/10.1186/s40537-024-00957-y https://doi.org/10.1186/s40537-024-00957-y https://doi.org/10.1016/j.dajour.2023.100230 https://doi.org/10.1016/j.dajour.2023.100230 https://doi.org/10.1007/s13721-024-00491-0 https://doi.org/10.1007/s13721-024-00491-0 https://doi.org/10.1016/j.edurev.2024.100616 https://doi.org/10.1016/j.edurev.2024.100616 https://doi.org/10.1016/j.compeleceng.2024.109409 https://doi.org/10.1016/j.compeleceng.2024.109409 https://doi.org/10.1007/s42979-021-00557-0 https://doi.org/10.1007/s42979-021-00557-0 https://doi.org/10.1080/17538947.2024.2353122 https://doi.org/10.1080/17538947.2024.2353122 https://doi.org/10.1016/j.heliyon.2024.e36743 https://doi.org/10.1016/j.heliyon.2024.e36743 https://doi.org/10.3390/su152416805 https://doi.org/10.3390/su152416805 https://doi.org/10.1007/s11356-022-19632-x https://doi.org/10.1007/s11356-022-19632-x Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath [43] Chen C, Li C, Reniers G and Yang F 2021 Safety and security of oil and gas pipeline transportation: a systematic analysis of research trends and future needs using WoS J. Clean. Prod. 279 123583 [44] Pranckute˙ R 2021 Web of science (WoS) and scopus: the titans of bibliographic information in today’s academic world Publications 9 12 [45] Arjunan P, Poolla K and Miller C 2020 EnergyStar++: towards more accurate and explanatory building energy benchmarking Appl. Energy 276 115413 [46] Konha¨user K and Werner T 2024 Uncovering the financial impact of energy-efficient building characteristics with eXplainable artificial intelligence Appl. Energy 374 123960 [47] Guy H, Vittoz S, Caputo G and Thiery T 2023 Benchmarking the energy performance of European commercial buildings with a Bayesian modeling framework Energy Build. 299 113595 [48] Seo Yoo Y, Shin H, Woo Kim D and Soo Park C 2024 Performance gap analysis for Korean building energy efficiency certification Energy Build. 314 114294 [49] Kontokosta C and Tull C 2017 A data-driven predictive model of city-scale energy use in buildings Appl. Energy 197 303–17 [50] Mohammadiziazi R and Bilec MM 2020 Application of machine learning for predicting building energy use at different temporal and spatial resolution under climate change in USA Buildings 10 139 [51] Moulla D, Attipoe D, Mnkandla E and Abran A 2024 Predictive model of energy consumption using machine learning: a case study of residential buildings in South Africa Sustainability 16 4365 [52] Streltsov A, Malof J M, Huang B and Bradbury K 2020 Estimating residential building energy consumption using overhead imagery Appl. Energy 280 116018 [53] Jiang F, Ma J, Li Z and Ding Y 2022 Prediction of energy use intensity of urban buildings using the semi- supervised deep learning model Energy 249 123631 [54] Shafiq M, Bhavani N P G, Venkata Naga Ramesh J, Veeresha R K, Talasila V and Sulaiman Alfurhood B 2024 Thermal modeling and Machine learning for optimizing heat transfer in smart city infrastructure balancing energy efficiency and Climate Impact Therm. Sci. Eng. Prog. 54 102868 [55] Alvarez-Sanz M, Satriya F A, Terés-Zubiaga J, Campos-Celador Á and Bermejo U 2024 Ranking building design and operation parameters for residential heating demand forecasting with machine learning J. Build. Eng. 86 108817 [56] Chakraborty D et al 2021 Scenario-based prediction of climate change impacts on building cooling energy consumption with explainable artificial intelligence Appl. Energy 291 116807 [57] Roth J, Martin A, Miller C and Jain R K 2020 SynCity: using open data to create a synthetic city of hourly building energy estimates by integrating data-driven and physics-based methods Appl. Energy 280 115981 [58] Zhang D, Ding Y and Jiang X 2023 A segmented evaluation model for building energy performance considering seasonal dynamic fluctuations Energy Convers. Manage. 298 117780 [59] Froemelt A, Buffat R and Hellweg S 2020 Machine learning based modeling of households: a regionalized bottom-up approach to investigate consumption-induced environmental impacts J. Ind. Ecol. 24 639–52 [60] Ji R, Wang K, Zhou M, Zhang Y, Bai Y, Wu X, Yan H, Zhao Z and Ye H 2023 Green space compactness and configuration to reduce carbon emissions from energy use in buildings Remote Sens. 15 1502 [61] Kusumoto Y, Delage R and Nakata T 2024 Machine learning application for estimating electricity demand by municipality Energy 296 131138 [62] Park J and Yang B 2020 GIS-enabled digital twin system for sustainable evaluation of carbon emissions: a case study of Jeonju City, South Korea Sustainability 12 9186 [63] Lara-benítez P, Carranza-García M, Luna-Romera J M and Riquelme J C 2020 Temporal convolutional networks applied to energy- related time series forecasting Appl. Sci. 10 2322 [64] Feng X, Wang S, Li Y, Yang J, Lei K and Yuan W 2024 Spatial heterogeneity and driving mechanisms of carbon emissions in urban expansion areas: a research framework coupled with patterns and functions Land Use Policy 143 107209 [65] Zhou Y and Cui X 2023 Can cross-sector information improve multi-energy demand forecasting accuracy? English Energy Rep. 9 886–93 [66] Excell L, Andrews A and Jain R 2024 E-Audit: a ”no-touch” energy audit that integrates machine learning and simulation Energy Build. 317 114360 [67] Liu X, Hu S and Yan D 2023 A statistical quantitative analysis of the correlations between socio-demographic characteristics and household occupancy patterns in residential buildings in China Energy Build. 284 112842 [68] Sarmas E, Spiliotis E, Marinakis V, Koutselis T and Doukas H 2022 A meta-learning classification model for supporting decisions on energy efficiency investments Energy Build. 258 111836 [69] Li Z, Ma J, Jiang F, Zhang S and Tan Y 2024 Assessing the impacts of urban morphological factors on urban building energy modeling based on spatial proximity analysis and explainable machine learning J. Build. Eng. 85 108675 [70] Sun M and Bardhan R 2024 Identifying Hard-to-Decarbonize houses from multi-source data in Cambridge, UK Sustain. Cities Soc. 100 105015 [71] Skiba M, Mro´wczynska M and Bazan-Krzywoszanska A 2017 Modeling the economic dependence between town development policy and increasing energy effectiveness with neural networks. Case study: the town of Zielona Gora Appl. Energy 188 356–66 [72] Chen X, Tu W, Yu J, Cao R, Yi S and Li Q 2024 LCZ-based city-wide solar radiation potential analysis by coupling physical modeling, machine learning, and 3D buildings Comput. Environ. Urban Syst. 113 102176 [73] Eslami S, Noorollahi Y, Marzband M and Anvari-Moghaddam A 2022 District heating planning with focus on solar energy and heat pump using GIS and the supervised learning method: case study of Gaziantep, Turkey Energy Convers. Manage. 269 116131 [74] Wu A and Biljecki F 2021 Roofpedia: automatic mapping of green and solar roofs for an open roofscape registry and evaluation of urban sustainability Landscape Urban Plan. 214 104167 [75] Meliani M et al 2022 Smart grid challenges in morocco and an energy demand forecasting with time series Int. J. Eng. Res. Afr. 61 195–215 [76] Dang N, Wang Q, Zhou K and Zhou T 2024 Coordinated transition of the supply and demand sides of China’s energy system Renew. Sustain. Energy Rev. 203 114744 [77] Veisi O, Shakibamanesh A and Rahbar M 2022 Using intelligent multi-objective optimization and artificial neural networking to achieve maximum solar radiation with minimum volume in the archetype urban block Sustain. Cities Soc. 86 104101 [78] Tu R, Wang A and Hatzopoulou M 2019 Improving the accuracy of emission inventories with a machine-learning approach and investigating transferability across cities J. Air Waste Manage. Assoc. 69 1377–90 29 https://doi.org/10.1016/j.jclepro.2020.123583 https://doi.org/10.1016/j.jclepro.2020.123583 https://doi.org/10.3390/publications9010012 https://doi.org/10.3390/publications9010012 https://doi.org/10.1016/j.apenergy.2020.115413 https://doi.org/10.1016/j.apenergy.2020.115413 https://doi.org/10.1016/j.apenergy.2024.123960 https://doi.org/10.1016/j.apenergy.2024.123960 https://doi.org/10.1016/j.enbuild.2023.113595 https://doi.org/10.1016/j.enbuild.2023.113595 https://doi.org/10.1016/j.enbuild.2024.114294 https://doi.org/10.1016/j.enbuild.2024.114294 https://doi.org/10.1016/j.apenergy.2017.04.005 https://doi.org/10.1016/j.apenergy.2017.04.005 https://doi.org/10.3390/buildings10080139 https://doi.org/10.3390/buildings10080139 https://doi.org/10.3390/su16114365 https://doi.org/10.3390/su16114365 https://doi.org/10.1016/j.apenergy.2020.116018 https://doi.org/10.1016/j.apenergy.2020.116018 https://doi.org/10.1016/j.energy.2022.123631 https://doi.org/10.1016/j.energy.2022.123631 https://doi.org/10.1016/j.tsep.2024.102868 https://doi.org/10.1016/j.tsep.2024.102868 https://doi.org/10.1016/j.jobe.2024.108817 https://doi.org/10.1016/j.jobe.2024.108817 https://doi.org/10.1016/j.apenergy.2021.116807 https://doi.org/10.1016/j.apenergy.2021.116807 https://doi.org/10.1016/j.apenergy.2020.115981 https://doi.org/10.1016/j.apenergy.2020.115981 https://doi.org/10.1016/j.enconman.2023.117780 https://doi.org/10.1016/j.enconman.2023.117780 https://doi.org/10.1111/jiec.12969 https://doi.org/10.1111/jiec.12969 https://doi.org/10.3390/rs15061502 https://doi.org/10.3390/rs15061502 https://doi.org/10.1016/j.energy.2024.131138 https://doi.org/10.1016/j.energy.2024.131138 https://doi.org/10.3390/su12219186 https://doi.org/10.3390/su12219186 https://doi.org/10.3390/app10072322 https://doi.org/10.3390/app10072322 https://doi.org/10.1016/j.landusepol.2024.107209 https://doi.org/10.1016/j.landusepol.2024.107209 https://doi.org/10.1016/j.egyr.2022.11.194 https://doi.org/10.1016/j.egyr.2022.11.194 https://doi.org/10.1016/j.enbuild.2024.114360 https://doi.org/10.1016/j.enbuild.2024.114360 https://doi.org/10.1016/j.enbuild.2023.112842 https://doi.org/10.1016/j.enbuild.2023.112842 https://doi.org/10.1016/j.enbuild.2022.111836 https://doi.org/10.1016/j.enbuild.2022.111836 https://doi.org/10.1016/j.jobe.2024.108675 https://doi.org/10.1016/j.jobe.2024.108675 https://doi.org/10.1016/j.scs.2023.105015 https://doi.org/10.1016/j.scs.2023.105015 https://doi.org/10.1016/j.apenergy.2016.12.006 https://doi.org/10.1016/j.apenergy.2016.12.006 https://doi.org/10.1016/j.compenvurbsys.2024.102176 https://doi.org/10.1016/j.compenvurbsys.2024.102176 https://doi.org/10.1016/j.enconman.2022.116131 https://doi.org/10.1016/j.enconman.2022.116131 https://doi.org/10.1016/j.landurbplan.2021.104167 https://doi.org/10.1016/j.landurbplan.2021.104167 https://doi.org/10.4028/p-2guf6 https://doi.org/10.4028/p-2guf6 https://doi.org/10.1016/j.rser.2024.114744 https://doi.org/10.1016/j.rser.2024.114744 https://doi.org/10.1016/j.scs.2022.104101 https://doi.org/10.1016/j.scs.2022.104101 https://doi.org/10.1080/10962247.2019.1668872 https://doi.org/10.1080/10962247.2019.1668872 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath [79] Chavhan S, Gupta D, Gochhayat S P, Cb N, Khanna A, Shankar K and Rodrigues J J P C 2022 Edge computing AI-IoT integrated energy-efficient intelligent transportation system for smart cities ACM Trans. Internet Technol. 22 1–18 [80] Chen G and Zhang J 2024 Intelligent transportation systems: machine learning approaches for urban mobility in smart cities Sustain. Cities Soc. 107 105369 [81] Shao Q, Zhang W, Cao X and Yang J 2023 Built environment interventions for emission mitigation: a machine learning analysis of travel-related CO2 in a developing city J. Transp. Geogr. 110 103632 [82] Huang R, Zhang X and Liu K 2024 Assessment of operational carbon emissions for residential buildings comparing different machine learning approaches: a study of 34 cities in China Build. Environ. 250 111176 [83] Golafshani E, Chiniforush A A, Zandifaez P and Ngo T 2024 An artificial intelligence framework for predicting operational energy consumption in office buildings Energy Build. 317 114409 [84] Fang Y, Lu X and Li H 2021 A random forest-based model for the prediction of construction-stage carbon emissions at the early design stage J. Cleaner Prod. 328 129657 [85] Kumanayake R and Luo H 2018 Life cycle carbon emission assessment of a multi-purpose university building: a case study of Sri Lanka Front. Eng. Manag. 5 393 [86] Carrera B, Peyrard S and Kim K 2021 Meta-regression framework for energy consumption prediction in a smart city: a case study of Songdo in South Korea Sustain. Cities Soc. 72 103025 [87] Yue Y et al 2024 Promoting solar energy utilization: prediction, analysis and evaluation of solar radiation on building surfaces at city scale Energy Build. 319 114561 [88] Astrain J, Falcone F, Lopez-Martin A J, Sanchis P, Villadangos J and Matias I R 2022 Monitoring of electric buses within an urban smart city environment IEEE Sens. J. 22 11364–72 [89] Tardioli G, Narayan A, Kerrigan R, Oates M, O’Donnell J and Finn D P 2020 A methodology for calibration of building energy models at district scale using clustering and surrogate techniques Energy Build. 226 110309 [90] Moran D et al 2018 Carbon footprints of 13,000 cities Environ. Res. Lett. 13 064041 [91] Climate TRACE 2025 Climate TRACE emissions inventory v4.1.0 (available at: https://climatetrace.org) (Accessed 19 April 2025) [92] Huo D et al 2022 Carbon Monitor Cities near-real-time daily estimates of CO2 emissions from 1500 cities worldwide Sci. Data 9 533 [93] CDP 2022 CDP cities dataset 2022 (available at: www.cdp.net/en/cities) (Accessed 19 April 2025) [94] C40 Cities 2022 C40 cities climate action planning framework and emissions data (available at: www.c40.org) (Accessed 19 April 2025) [95] Local Governments for Sustainability, Cities Climate Leadership Group and World Resources Institute 2014 Global protocol for community-scale greenhouse gas emission inventories (GPC) Tech. Rep (World Resources Institute) (available at: https:// ghgprotocol.org/greenhouse-gas-protocol-accounting-reporting-standard-cities) (Accessed 19 April 2025) [96] Crippa M et al 2024 GHG Emissions of All World Countries Tech. Rep. JRC138862 (Publications Office of the European Union) (https://doi.org/10.2760/4002897) [97] Nangini C et al 2019 A global dataset of CO2 emissions and ancillary data related to emissions for 343 cities Sci. Data 6 180280 [98] Amalou I, Mouhni N and Abdali A 2022 Multivariate time series prediction by RNN architectures for energy consumption forecasting Energy Rep. 8 1084–91 [99] Cerquitelli T, Malnati G and Apiletti D 2019 Exploiting scalable machine-learning distributed frameworks to forecast power consumption of buildings Energies 12 2933 [100] Yang G, Zhang G, Cao D, Gao X, Wang X, Yang S, Jiang P, Zha D and Shan Y 2024 A comprehensive city-level final energy consumption dataset including renewable energy for China, 2005–2021 Sci. Data 11 738 [101] Pevec D, Babic J and Podobnik V 2019 Electric vehicles: a data science perspective review Electronics 8 1190 [102] Cui X, Lee M, Koo C and Hong T 2024 Energy consumption prediction and household feature analysis for different residential building types using machine learning and SHAP: toward energy-efficient buildings Energy Build. 309 113997 [103] Zhang Y, Teoh B K, Wu M, Chen J and Zhang L 2023 Data-driven estimation of building energy consumption and GHG emissions using explainable artificial intelligence Energy 262 125468 [104] Liu X, Tang H, Ding Y and Yan D 2022 Investigating the performance of machine learning models combined with different feature selection methods to estimate the energy consumption of buildings Energy Build. 273 112408 [105] Garc´ıa-l´opez J, Dom´ınguez-amarillo S and Sendra J J 2024 Clustering open data for predictive modeling of residential energy consumption across variable scales: a Case Study in Andalusia, Spain Buildings 14 2335 [106] Fathi S, Srinivasan R S, Kibert C J, Steiner R L and Demirezen E 2020 AI-based campus energy use prediction for assessing the effects of climate change Sustainability 12 3223 [107] Papadopoulos S and Kontokosta C E 2019 Grading buildings on energy performance using city benchmarking data Appl. Energy 233–234 244–53 [108] Wenninger S, Kaymakci C and Wiethe C 2022 Explainable long-term building energy consumption prediction using QLattice Appl. Energy 308 118300 [109] Zeki´c-suˇsac M, Has A and Kneˇzevi´c M 2021 Predicting energy cost of public buildings by artificial neural networks, CART, and random forest Neurocomputing 439 223–33 [110] Ren H, Xu C, Ma Z and Sun Y 2022 A novel 3D-geographic information system and deep learning integrated approach for high-accuracy building rooftop solar energy potential characterization of high-density cities Appl. Energy 306 117985 [111] Yan S, Zhang Y, Sun H and Wang A 2023 A real-time operational carbon emission prediction method for the early design stage of residential units based on a convolutional neural network: a case study in Beijing, China J. Build. Eng. 75 106994 [112] Sun M et al 2022 Understanding building energy efficiency with administrative and emerging urban big data by deep learning in Glasgow (https://doi.org/10.31219/osf.io/g8p4f) [113] Aljohani A 2024 Deep learning-based optimization of energy utilization in IoT-enabled smart cities: a pathway to sustainable development Energy Rep. 12 2946–57 [114] Gao L, Liu T, Cao T, Hwang Y and Radermacher R 2021 Comparing deep learning models for multi energy vectors prediction on multiple types of building Appl. Energy 301 117486 [115] Peteleaza D, Matei A, Sorostinean R, Gellert A, Fiore U, Zamfirescu B-C and Palmieri F 2024 Electricity consumption forecasting for sustainable smart cities using machine learning methods Internet Things 27 101322 [116] Carrera B and Kim K 2023 A regression framework for energy consumption in smart cities with encoder-decoder recurrent neural networks Energies 16 7508 30 https://doi.org/10.1145/3507906 https://doi.org/10.1145/3507906 https://doi.org/10.1016/j.scs.2024.105369 https://doi.org/10.1016/j.scs.2024.105369 https://doi.org/10.1016/j.jtrangeo.2023.103632 https://doi.org/10.1016/j.jtrangeo.2023.103632 https://doi.org/10.1016/j.buildenv.2024.111176 https://doi.org/10.1016/j.buildenv.2024.111176 https://doi.org/10.1016/j.enbuild.2024.114409 https://doi.org/10.1016/j.enbuild.2024.114409 https://doi.org/10.1016/j.jclepro.2021.129657 https://doi.org/10.1016/j.jclepro.2021.129657 https://doi.org/10.15302/J-FEM-2018055 https://doi.org/10.15302/J-FEM-2018055 https://doi.org/10.1016/j.scs.2021.103025 https://doi.org/10.1016/j.scs.2021.103025 https://doi.org/10.1016/j.enbuild.2024.114561 https://doi.org/10.1016/j.enbuild.2024.114561 https://doi.org/10.1109/JSEN.2021.3077846 https://doi.org/10.1109/JSEN.2021.3077846 https://doi.org/10.1016/j.enbuild.2020.110309 https://doi.org/10.1016/j.enbuild.2020.110309 https://doi.org/10.1088/1748-9326/aac72a https://doi.org/10.1088/1748-9326/aac72a https://climatetrace.org https://doi.org/10.1038/s41597-022-01657-z https://doi.org/10.1038/s41597-022-01657-z https://www.cdp.net/en/cities https://www.c40.org https://ghgprotocol.org/greenhouse-gas-protocol-accounting-reporting-standard-cities https://ghgprotocol.org/greenhouse-gas-protocol-accounting-reporting-standard-cities https://doi.org/10.2760/4002897 https://doi.org/10.1038/sdata.2018.280 https://doi.org/10.1038/sdata.2018.280 https://doi.org/10.1016/j.egyr.2022.07.139 https://doi.org/10.1016/j.egyr.2022.07.139 https://doi.org/10.3390/en12152933 https://doi.org/10.3390/en12152933 https://doi.org/10.1038/s41597-024-03529-0 https://doi.org/10.1038/s41597-024-03529-0 https://doi.org/10.3390/electronics8101190 https://doi.org/10.3390/electronics8101190 https://doi.org/10.1016/j.enbuild.2024.113997 https://doi.org/10.1016/j.enbuild.2024.113997 https://doi.org/10.1016/j.energy.2022.125468 https://doi.org/10.1016/j.energy.2022.125468 https://doi.org/10.1016/j.enbuild.2022.112408 https://doi.org/10.1016/j.enbuild.2022.112408 https://doi.org/10.3390/buildings14082335 https://doi.org/10.3390/buildings14082335 https://doi.org/10.3390/su12083223 https://doi.org/10.3390/su12083223 https://doi.org/10.1016/j.apenergy.2018.10.053 https://doi.org/10.1016/j.apenergy.2018.10.053 https://doi.org/10.1016/j.apenergy.2021.118300 https://doi.org/10.1016/j.apenergy.2021.118300 https://doi.org/10.1016/j.neucom.2020.01.124 https://doi.org/10.1016/j.neucom.2020.01.124 https://doi.org/10.1016/j.apenergy.2021.117985 https://doi.org/10.1016/j.apenergy.2021.117985 https://doi.org/10.1016/j.jobe.2023.106994 https://doi.org/10.1016/j.jobe.2023.106994 https://doi.org/10.31219/osf.io/g8p4f https://doi.org/10.1016/j.egyr.2024.08.075 https://doi.org/10.1016/j.egyr.2024.08.075 https://doi.org/10.1016/j.apenergy.2021.117486 https://doi.org/10.1016/j.apenergy.2021.117486 https://doi.org/10.1016/j.iot.2024.101322 https://doi.org/10.1016/j.iot.2024.101322 https://doi.org/10.3390/en16227508 https://doi.org/10.3390/en16227508 Mach. Learn.: Earth 1 (2025) 011001 T Chen and R Debnath [117] Jahani E and Cetin K 2022 Energy savings and retrofit assessment for city-scale residential building stock during extreme heatwave events using genetic algorithm- numerical moment matching Clean Technol. Environ. Policy 24 2081–98 [118] Chui K T, Lytras M D and Visvizi A 2018 Energy sustainability in smart cities: artificial intelligence, smart monitoring, and optimization of energy consumption Energies 11 2869 [119] Le L T et al 2019 Estimating the heating load of buildings for smart city planning using a novel artificial intelligence technique PSO-XGBoost Appl. Sci. 9 2714 [120] Jin X, Xiao F, Zhang C and Li A 2022 GEIN: an interpretable benchmarking framework towards all building types based on machine learning Energy Build. 260 111909 [121] Zhang W, Gu X, Tang L, Yin Y, Liu D and Zhang Y 2022 Application of machine learning, deep learning and optimization algorithms in geoengineering and geoscience: comprehensive review and future challenge Gondwana Res. 109 1–17 [122] Ye Z, Cheng K, Hsu S-C, Wei H-H and Cheung C M 2021 Identifying critical building-oriented features in city-block-level building energy consumption: a data-driven machine learning approach Appl. Energy 301 117453 [123] Zhang X, Yan F, Liu H and Qiao Z 2021 Towards low carbon cities: a machine learning method for predicting urban blocks carbon emissions (UBCE) based on built environment factors (BEF) in Changxing City, China Sustain. Cities Soc. 69 102875 [124] Tian P, Cai M, Sun Z, Liu S, Wu H, Liu L and Peng Z 2024 Effects of 3D urban morphology on CO2 emissions using machine learning: towards spatially tailored low-carbon strategies in Central Wuhan, China Urban Clim. 57 102122 [125] Raji´c M N, Milovanovíc M B, Antíc D S, Maksimovíc R M, Milosavljevíc P M and Pavlovíc D L 2020 Analyzing energy poverty using intelligent approach Energy Environ. 31 1448–72 [126] Zhang Y, Liao X and Sun D 2024 A coupled InVEST-PLUS model for the spatiotemporal evolution of ecosystem carbon storage and multi-scenario prediction analysis LAND 13 509 [127] Callaghan M et al 2025 Machine learning map of climate policy literature reveals disparities between scientific attention, policy density, and emissions npj Clim. Action 4 7 [128] Pena Acosta M, Dikkers M, Vahdatikhaki F, Santos J and Dorée A G 2023 A comprehensive generalizability assessment of data-driven Urban Heat Island (UHI) models Sustain. Cities Soc. 96 104701 [129] Man Q, Yu H, Feng K, Olofsson T and Lu W 2024 Transfer of building retrofitting evaluations for data-scarce conditions: an empirical study for Sweden to China Energy Build. 310 114041 [130] Alikhademi K et al 2021 Can explainable AI explain unfairness? a framework for evaluating explainable AI (arXiv:2106.07483) [131] Burke M, Driscoll A, Lobell D B and Ermon S 2021 Using satellite imagery to understand and promote sustainable development Science 371 eabe8628 [132] Chen C et al 2024 Explainable artificial intelligence framework for urban global digital elevation model correction based on the SHapley additive explanation- random forest algorithm considering spatial heterogeneity and factor optimization Int. J. Appl. Earth Observ. Geoin. 129 103843 [133] Bellantuono L, Palmisano F, Amoroso N, Monaco A, Peragine V and Bellotti R 2023 Detecting the socio-economic drivers of confidence in government with eXplainable artificial intelligence Sci. Rep. 13 839 [134] Wang Y et al 2024 A SHAP machine learning-based study of factors influencing urban residents’ electricity consumption—evidence from Chinese provincial data Environ. Dev. Sustain. 26 30445–76 31 https://doi.org/10.1007/s10098-022-02299-w https://doi.org/10.1007/s10098-022-02299-w https://doi.org/10.3390/en11112869 https://doi.org/10.3390/en11112869 https://doi.org/10.3390/app9132714 https://doi.org/10.3390/app9132714 https://doi.org/10.1016/j.enbuild.2022.111909 https://doi.org/10.1016/j.enbuild.2022.111909 https://doi.org/10.1016/j.gr.2022.03.015 https://doi.org/10.1016/j.gr.2022.03.015 https://doi.org/10.1016/j.apenergy.2021.117453 https://doi.org/10.1016/j.apenergy.2021.117453 https://doi.org/10.1016/j.scs.2021.102875 https://doi.org/10.1016/j.scs.2021.102875 https://doi.org/10.1016/j.uclim.2024.102122 https://doi.org/10.1016/j.uclim.2024.102122 https://doi.org/10.1177/0958305X20907087 https://doi.org/10.1177/0958305X20907087 https://doi.org/10.3390/land13040509 https://doi.org/10.3390/land13040509 https://doi.org/10.1038/s44168-024-00196-0 https://doi.org/10.1038/s44168-024-00196-0 https://doi.org/10.1016/j.scs.2023.104701 https://doi.org/10.1016/j.scs.2023.104701 https://doi.org/10.1016/j.enbuild.2024.114041 https://doi.org/10.1016/j.enbuild.2024.114041 https://arxiv.org/abs/2106.07483 https://doi.org/10.1126/science.abe8628 https://doi.org/10.1126/science.abe8628 https://doi.org/10.1016/j.jag.2024.103843 https://doi.org/10.1016/j.jag.2024.103843 https://doi.org/10.1038/s41598-023-28020-5 https://doi.org/10.1038/s41598-023-28020-5 https://doi.org/10.1007/s10668-024-05263-4 https://doi.org/10.1007/s10668-024-05263-4 The need of explainability in low-carbon urban system design using AI: A systematic review 1. Introduction 2. Methods 3. Results 3.1. General observations 3.2. Objective landscape: AI-facilitated urban climate initiatives 3.2.1. Building energy consumption prediction 3.2.2. City carbon prediction and management 3.2.3. The development of new energy sources 3.2.4. Building energy consumption reduction 3.2.5. Transportation carbon prediction and management 3.3. Data typologies: big data for urban systems 3.3.1. Data sources 3.3.2. Data categories 3.3.3. Data gaps 3.4. Methodological spectrum: AI-driven approaches in the literature 3.4.1. ML 3.4.2. DL 3.4.3. Optimization algorithms 3.4.4. XAI 3.5. Synthesis: applications of AI in urban decarbonisation 3.5.1. Forecasting carbon emissions 3.5.2. Identifying factors influencing energy consumption 3.5.3. Developing data-driven strategies for low-carbon urban planning 3.6. Synthesis: AI-driven methods for real-world applications 3.6.1. Model performance and generalizability 3.6.2. Explainability and policy recommendations 4. Discussion 4.1. The pathway for using AI for urban system decarbonisation 4.2. Scope, objective and data challenges 4.3. Gaps in XAI for urban climate action 4.3.1. Trustworthiness 4.3.2. Transparency 4.3.3. Fairness 4.3.4. Interpretability for multiple decision-makers 4.3.5. Precise management 4.4. Contributions of this review 4.5. Limitations of the literature review 5. Conclusion Appendix References