Deterministic Forecasts and Prediction Intervals for Wind Speed Using Enhanced Multi-Quantile Loss Based Dilated Causal Convolutions

Abstract

With rising wind power penetration into power systems obtaining wind speed forecasts with associated uncertainty becomes crucial for better planning and dispatch. This study proposes an enhanced multi quantile regression-based loss function specially tailored to train models to generate both deterministic forecast and the corresponding prediction intervals. Though the regression architecture of the model plays an important role in extracting precise forecasts, however, its efficiency is often ignored which may be a downside for short term forecasting scenarios where model training time may also be a significant factor. The present study therefore designed a multi-scale dilated convolution-based architecture for enhanced efficiency. The architecture generates predictions at different scales which are combined using particle swarm optimization to obtain optimal forecasts. The model is trained using the proposed loss function on datasets from both NREL simulations and operational Chinese state grid measurements across three different locations. The proposed model exhibits excellent forecasting performance in comparative experiments with both simulated and real-world operational datasets.