Abstract
Modeling the behavior of wind in the atmosphere is a long-term ambition for humanity and an indispensable initiative to promote wind energy. The real challenge lies in the high complexity of modeling non-linear dynamic systems to simulate intermittent behaviors. Simulations of current forecasting models are approximate, time-consuming, and computationally expensive. This work presents a novel hybrid adaptive spatiotemporal domain-informed neural network for short-term wind-component forecasting targeting five regions with varying wind dynamics, three of which are in the state of Bahia, Brazil: Esplanada, Mucuge, Mucuri; one in the United Kingdom: Kelmarsh, located in Northamptonshire; and one in the United States: Roscoe, located in Texas. The performance of a fully data-driven multi-layer perceptron was compared to a domain-informed multi-layer perceptron that shares the same architecture with an addition of the numerical methodology to incorporate meteorological inputs and introduces local spatiotemporal regularization through the loss function. The results were evaluated assessing predictive accuracy (NRMSE, RMSE, MAE, đ
2 , Pearsonâs R), computational efficiency, and model interpretability. The quantitative results indicate a modest gain in predictive performance, with the mean Pearsonâs R increasing from 0.905 to 0.919 for the u-component and from 0.937 to 0.944 for the v-component. The statistical validity was confirmed using the ShapiroâWilk, Kruskal-Walis, Wilcoxon signed-rank and DieboldâMariano tests.