Enhanced wind speed forecasting for sustainable power systems: A deep learning framework unifying deterministic predictions and uncertainty quantification

This paper presents an innovative deep learning framework that enhances renewable energy integration by simultaneously generating both deterministic forecasts and prediction intervals (PIs) for wind speed. Traditional deterministic forecasts, while useful, lack a measure of uncertainty. In contrast, Prediction Intervals (PIs) communicate uncertainty but lack the specific point forecasts needed for precise operational planning. Our model addresses this by simultaneously generating point estimates and prediction intervals directly from a single model trained using a novel loss function called “Enhanced Mean-Variance Estimation (EMVE)". The loss function optimizes the accuracy of predicted distribution parameters and calibrates PIs while incorporating an adaptive quantile violation mechanism that compares standardized residuals with scaled uncertainty estimates. To effectively leverage this loss function, we designed a branched ensemble model architecture utilizing LSTM. Synergy between this specialized architecture and EMVE loss function with quantile violation correction enables the model to effectively learn the underlying wind speed distribution, leading to superior performance in generating both point estimates and prediction intervals. Empirical evaluation for three different locations encompassing NREL WIND Toolkit datasets and operational Chinese state grid measurements demonstrate the model achieves 38 % and 21 % improvement in coverage width criterion compared to state-of-the-art approaches, establishing superior performance across both synthetic and real-world operational environments making it a leading solution for reliable wind speed forecasting.