Dynamic-Error-Compensation-Assisted Deep Learning Framework for Solar Power Forecasting
Heng-Yi Su, Chen Tang
Abstract
A reliable approach to forecasting solar energy generation using deep learning (DL) models is presented. The approach relies on a prediction-correction (PC) framework. It is composed of a primary model that performs preliminary prediction, followed by a secondary model that is charged with the task of dynamic error compensation (DEC), based on hierarchical residual (HR) learning and Choquet fuzzy integral (CFI) technique. An improved gated recurrent unit (IGRU) is designed and integrated into the PC framework. Moreover, a practical algorithm is developed to facilitate the calculation of the CFI aggregation. Empirical studies on real-world data sets are presented, illustrating the gains in accuracy and reliability of the proposed approach.