Fast prediction of spatial temperature distributions in urban areas with WRF and temporal fusion transformers

Abstract

Urban Heat Island (UHI) poses a significant challenge to the sustainable development of global cities. It is of great importance to efficiently characterize the spatiotemporal distribution of urban temperatures for UHI mitigation strategies, such as urban ecosystem planning and control. Numerical Weather Prediction (NWP) methods are used to obtain the urban temperature distribution. However, NWP requires significant hardware resources and long computation time. The development of artificial intelligence approaches have been applied in expediting the weather forecasting, yet their forecasting precision remains significantly inferior to that of NWP. Hence, this study aims to propose a hybrid fast prediction model, considering the accuracy of WRF (Weather Research and Forecasting) and efficiency of Temporal Fusion Transformer (TFT) neural networks. By integrating high-precision temperature time series boundaries generated by WRF into TFT, this method (WRF-TFT) is able to realize the rapid predictions of urban temperature distributions (around 15 times faster compare to WRF) while maintaining the physical characteristics of atmospheric dynamics. With this method, we also conducted for future temperature forecast for cities. It is estimated that the temperature can exceed 35 °C more than 12 hours per day in July 2050. This hybrid model facilitates swift acquisition of urban temperature trends, providing a crucial basis for urban risk management and planning.