Neural Network Approach for Wind Forecasting

Physical approaches to wind speed forecasting have been widely studied in the past. More recent developments in machine learning have seen success in modeling nonlinear phenomena. Short-term forecasting of wind speed is a highly nonlinear problem, and a statistical approach could reduce prediction error. Here we present a neural network approach that is trained with historical measurements of temperature and pressure from multiple locations to predict hourly wind speeds 48 hours in advance. Multiple approaches are considered for wind forecasting and treating input-data locations, such as sensitivity to station proximity and wind directions. The results suggest that this model is capable of capturing some of the nonlinear aspects of wind. Physical and practical insights on future improvements will be discussed as well.