Integrating Physics-Informed Vectors for Improved Wind Speed Forecasting with Neural Networks

This paper introduces an approach to enhance wind speed prediction by integrating Physics-Informed Vectors with neural network architectures, specifically Long Short-Term Memory and Temporal Convolution Networks. It also proposes a hybrid decaying loss function aimed at improving the efficiency of model training and its predictive performance. The methodology utilizes physical knowledge and atmospheric data from various global locations to develop predictive models. Evaluations conducted using datasets from Canada, Chile, Kazakhstan, and Mongolia illustrate the advantages of including Physics-Informed Vectors. The incorporation of these vectors leads to improvements in Mean Squared Error, Mean Absolute Error, and R2 Score across different volumes of data. The analysis reveals improvements of up to 8.43% in Mean Absolute Error, 16.39% in Mean Squared Error, and 0.82% in R2 Score for models based on Long Short-Term Memory. For models based on Temporal Convolution Networks, improvements of up to 17.27% in Mean Absolute Error, 29.24% in Mean Squared Error, and 1.55% in R2 Score were observed. The introduction of a custom loss function, which merges mean squared error with Physics-Informed Vector estimates, aids in hastening the convergence of models by modulating the influence of these vectors during the training phase. The study underscores the effectiveness of incorporating physics-informed techniques into machine learning for predicting renewable energy sources, thereby opening paths for further research and application in this domain.