A new wind power interval prediction approach based on reservoir computing and a quality-driven loss function

Abstract

Understanding the uncertainty of wind power forecasting is crucial for its practical application. This paper proposes a new forecasting approach to estimate the wind power prediction intervals (PIs) to quantify the prediction uncertainty. This approach integrates the reservoir computing methodology into a three-layer neural network architecture, and outputs the final PIs by minimizing a quality-driven loss function. The reservoir computing methodology can help study the nonlinear relationship implicit in data as well as accelerate computational time, while the quality-driven loss function is assumption-free and can help improve the forecasting capability of the proposed model. The proposed model is applied to real wind power data to test its effectiveness. Case studies show that for the data used in this paper, the proposed model can reduce the mean prediction interval width (MPIW) by up to 16.69%, reduce root mean square error (RMSE) by up to 7.36%, and save up to 5 times computation time compared to the benchmark models, these indicate that the proposed model has strong predictive capability.