Evaluating Critical Weather Parameters Using Machine Learning Models

Abstract Wind speed and temperature forecasting accurately for an urban area are two critical elements to mitigate cost and energy in engineering calculations. Recently, researchers have employed a combination of mesoscale Weather Research and Forecasting (WRF) and microscale models of Computational Fluid Dynamics (CFD) to simulate wind flow in urban areas. Employing machine learning algorithms is also introduced as another promising alternative/complementary tool. In this work, a WRF model is developed to calculate wind speed at 10m above the ground in two different residential regions with different upwind roughness in Singapore. The results are validated by the literature. Using the validated model, Recurrent Neural Network (RNN) models are developed to compare with numerical weather models and evaluate their performance in prediction. Prediction results’ accuracy is evaluated using Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE), and obtained errors below 0.19 and 5%, respectively. The models can provide high-resolution data as boundary conditions for CFD modeling at the local/building scale, which leads to understanding the more realistic behavior of the flow field in a built environment. In addition, they support an extended range of urban microclimate analysis, renewable energy integration practices, and smart city planning efforts.