Combining computational fluid dynamics and neural networks to characterize microclimate extremes: Learning the...

The urban form and extreme microclimate events can have an important impact on the energy performance of buildings, urban comfort and human health. State-of-the-art building energy simulations require information on the urban microclimate, but typically rely on ad-hoc numerical simulations, expensive in-situ measurements, or data from nearby weather stations. As such, they do not account for the full range of possible urban microclimate variability and findings cannot be generalized across urban morphologies. To bridge this knowledge gap, this study proposes two data-driven models to downscale climate variables from the meso to the micro scale in arbitrary urban morphologies, with a focus on extreme climate conditions. The models are based on a feedforward and a deep neural network (NN) architecture, and are trained using results from computational fluid dynamics (CFD) simulations of flow over a series of idealized but representative urban environments, spanning a realistic range of urban morphologies. Both models feature a relatively good agreement with corresponding CFD training data, with a coefficient of determination R² = 0.91 (R² = 0.89) and R² = 0.94 (R² = 0.92) for spatially-distributed wind magnitude and air temperature for the deep NN (feedforward NN). The models generalize well for unseen urban morphologies and mesoscale input data that are within the training bounds in the parameter space, with a R² = 0.74 (R² = 0.69) and R² = 0.81 (R² = 0.74) for wind magnitude and air temperature for the deep NN (feedforward NN). The accuracy and efficiency of the proposed CFD-NN models makes them well suited for the design of climate-resilient buildings at the early design stage.