Improving the accuracy of microclimate coupled urban building energy modeling using convolutional neural networks

Abstract

Meteorological data plays a crucial role in building energy simulations. Currently, most urban building energy modeling approaches rely primarily on Typical Meteorological Year (TMY) data. However, TMY data represents city-wide climate conditions, which often differ significantly from actual microclimates, leading to simulation inaccuracies. To address this issue, this study proposes a convolutional neural network (CNN)-based approach to reliably predict year-round microclimate data for the study area, thereby enhancing simulation accuracy. The method utilizes meteorological data from nearby weather stations (NWS), adjusted using the Urban Weather Generator (UWG) model, as the prediction target. A convolutional approach is employed to resolve dimensional inconsistencies between urban morphology features and hourly meteorological variables such as temperature, wind direction, and wind speed. By applying pointwise multiplication, the model effectively integrates these two data types and leverages CNNs to achieve high-precision microclimate predictions, improving simulation accuracy by 8 %. The robustness of the proposed method is validated using real energy consumption data from dozens of buildings in Shanghai. Notably, this approach enables accurate microclimate prediction in areas lacking NWS coverage, significantly enhancing the accuracy of energy consumption simulations.