ReCliDaR: An open-source ML-based tool to identify representative climate days for urban environmental modeling

Source abstract

Building Energy Modeling and Computational Fluid Dynamics-based Urban Climate Modeling tools are essential environmental modeling tools for designers to make informed, climate-responsive design decisions. However, these tools are computationally intensive, particularly when applied at large spatial scales or within optimization workflows. Traditional approaches, such as simulating for solstice, equinox, or peak cooling days, often fail to capture the diversity of climatic conditions. This can lead to overestimation or undermining of urban environmental performance. The Representative Climate Days Recognizer (ReCliDaR) tool bridges this gap by utilizing unsupervised machine learning to identify representative climate days, condensing a typical meteorological year into a manageable subset of 24-hour profiles. The framework of the tool automates the data preprocessing, feature scaling, and dimensionality reduction before executing a comparative clustering analysis using k-means, Gaussian Mixture, and Hierarchical Agglomerative Clustering. Evaluation of 108 global cities across 11 ASHRAE climate zones demonstrates that the k-means algorithm provides the most robust performance, achieving almost a 100% seasonal continuity rate in most regions while minimizing data loss. The results indicate that while temperate and high-latitude climates typically follow binary seasonal structures, tropical environments often exhibit complex ternary patterns driven by monsoon cycles. By providing a validated, user-friendly interface for extracting medoid-based representative days, the ReCliDaR tool empowers researchers and practitioners to perform accurate, long-term environmental assessments. The tool and its source code are openly available on GitHub to support reproducible research in climate-responsive urban design and energy-efficient building performance.