A systematic review of explainable artificial intelligence in urban building energy modeling: methods, applications, and future directions
Leila Darvishvand, Babak Kamkari, Ming Jun Huang, Neil Hewitt
Abstract
Urban Building Energy Modeling (UBEM) has emerged as a critical tool for addressing the growing energy demands of cities and promoting sustainability by optimizing energy use, reducing emissions, and improving efficiency at the urban scale. Machine learning (ML) techniques have transformed UBEM applications by enabling energy consumption prediction, demand optimization, and retrofitting strategies. However, the complexity of many ML models limits their interpretability and can undermine stakeholder trust in decision-making. Explainable Artificial Intelligence (XAI) addresses these issues by increasing transparency and making ML models more understandable and actionable. Given the multidisciplinary nature of urban energy planning, XAI enables policymakers, urban planners, engineers, and building managers to extract relevant insights tailored to their needs. By translating complex ML predictions into understandable explanations, XAI fosters trust and supports inclusive decision-making. This systematic review examines the role of XAI in UBEM and presents a taxonomy of XAI methodologies based on explainability approach, explanation scope, and model applicability. It highlights how XAI methods contribute to key UBEM applications, including energy forecasting, retrofitting, clustering, and occupancy modeling. Despite recent advances, several challenges remain, such as the computational cost of post-hoc techniques at city scale and balancing interpretability with predictive performance. The review also identifies underexplored areas, including interpretable clustering and occupancy behavior modeling, alongside the limited adoption of model-specific explainability methods. To address these gaps, future research should explore the scalability and usability of XAI methods in UBEM, investigate model-specific techniques for deep and graph-based models, and develop approaches for evaluating the effectiveness of XAI in real-world UBEM contexts.