Multi-Objective Estimation of Optimal Prediction Intervals for Wind Power Forecasting
Yinsong Chen, Samson S. Yu, Chee Peng Lim, Peng Shi
Abstract
Accurate and reliable wind power forecasting is crucial for efficient operation in a power system. Due to the substantial uncertainties associated with wind generation, probabilistic interval forecasting offers a distinct approach for assessing and quantifying the potential impacts and risks that may arise from the integration of wind energy into a power system. This paper proposes a novel multi-objective lower upper bound estimation method to directly construct optimal wind power intervals without the assumption of any specific distribution function. Prediction intervals at a nominal confidence level are formulated through simultaneously optimizing the Winkler loss and coverage probability. The proposed framework is gradient descent-enabled and therefore allows flexible integration of various deep learning algorithms. An evaluation using four wind power datasets is conducted, and the results are analyzed and compared with those from several benchmark models. The findings indicate the proposed method outperforms its counterparts in terms of both reliability and overall performance.