Observations Driven Machine Learning Prediction of the Urban Surface Energy Balance

Accurate urban surface energy balance (SEB) predictions are essential for reliable weather forecasting, influencing near-surface variables and serving as boundary conditions for city-scale atmospheric flow. However, results from the Urban-PLUMBER intercomparison project (Lipson et al., 2024) indicate that despite advances in urban SEB modelling, conventional models often underperform compared to simple empirical benchmark models derived from flux tower observations.

This study investigates whether machine learning (ML) models, trained on flux tower observations and land surface descriptors, can outperform traditional urban SEB models. Our results show that ML approaches achieve lower errors across standard metrics than nearly all Urban-PLUMBER participant models. However, ML models face challenges, particularly in maintaining expected physical relationships such as energy budget closure. We discuss these limitations and propose solutions, including hybrid approaches that combine ML with conventional models to address the sparse global coverage of urban SEB observations.

Since the SEB underpins urban surface-atmosphere interactions, its representation will be essential in urban digital twins. The neural network approach presented is computationally efficient, utilizes standard urban land surface model input parameters, and can be readily integrated into digital twin frameworks. By improving SEB predictions, this method has the potential to enhance forecasts of urban hazards such as heatwaves and flooding.