Abstract. Accurate urban climate modeling is crucial for addressing the growing impacts of urban heat islands (UHI) and climate change. Physics-based tools such as the Urban Weather Generator (UWG) are widely used but often limited by high parameterization needs and a lack of specialized data. In this study, we develop a hybrid framework combining UWG simulations with deep learning, introducing two models: NUWG-Sim (Neural Urban Weather Generator on Simulations), trained solely on simulated data, and NUWG-city, which is fine-tuned with ground weather station data. To systematically evaluate model performance across heterogeneous urban contexts, we structure our experiments around Local Climate Zones (LCZs) in Toulouse, France. Our methodology involves generating over 3400 UWG initialization files, simulating urban air temperatures time series for diverse surface parameters, and training a neural model on these series. We then fine-tune the model with observed data from selected weather stations, analyzing how the number and diversity of stations environments impact performance on unseen stations from different LCZs. Results show that even limited fine-tuning significantly improves performance, particularly when training includes stations from LCZs similar to the test set. The approach highlights the potential of physics-informed neural models for city-specific urban climate monitoring.
Hamdi et al. (Thu,) studied this question.