An Intelligent Data-Driven Method for Enhanced Solar Power Interval Forecasting Based on Mixture-of-Deep-Learning-Experts

Considering the pronounced variability and intermittent characteristics inherent in solar photovoltaic (PV) power generation, the adoption of interval forecasting (IF) emerges as a pivotal tool, furnishing grid operators with essential insights for optimizing energy management in contrast to traditional point forecasting. In this context, the article presents an intelligent data-driven approach based on deep learning (DL) to advance short-term solar power interval forecasting (SPIF). To this end, a novel mixture-of-DL-experts (MoDLE) for IF, termed MoDLE-IF, is proposed. Within the MoDLE, the temporal convolutional networks (TCNs) serve as the primary DL experts, contributing to the enhancement of the overall forecasting accuracy. A power transformation technique is seamlessly incorporated into the MoDLE-IF framework to address forecast uncertainties. The proposed approach facilitates efficient quantification of SPIF uncertainty by utilizing prediction intervals (PIs). To validate the effectiveness of this approach, we executed comprehensive experimental studies using realistic solar power data. This assessment examines the quality and reliability of our method by benchmarking it against several cutting-edge IF techniques. The results demonstrate the exceptional performance of the proposed methodology in generating highly accurate PIs compared to other benchmark techniques.