A novel photovoltaic power probabilistic forecasting model based on monotonic quantile convolutional neural network and multi-objective optimization

Abstract

Photovoltaic (PV) power probabilistic forecasting that provides decision makers with probabilistic information and ranges of PV power generation is critical to the power system. Existing studies have demonstrated that QR-based nonlinear models can generate probability distributions directly from historical data. However, the accuracy of these methods may be degraded when confronting with PV power at high latitude meteorological factors and they inherently have flaws in the model structure and loss function. This paper proposes a novel approach called monotonic quantile convolutional neural network-multi-layer nondominated fast sort genetic algorithm II (MQCNN-MLNSGAII) for solving these challenges. MQCNN first uses the convolutional structure to extract the valid deep features from the high latitude factor, and then designs a monotonic quantile structure to output monotonically increasing probability distributions at once. Considering the high impact of the probability distribution width on the quality of the forecasting, we design two loss functions, average quantile loss (AQS) and quantile distribution average width (QDAW), based on multi-objective optimization (MOO) to balance the reliability and width. Finally, a novel multi-objective evolutionary algorithm (MOEA), MLNSGAII, is proposed for training MQCNN. It develops a multi-layer mechanism based on global and historical information to assist the algorithm in generating diverse offspring and improve the performance in convergence and diversity. Compared to the benchmark models, the proposed model achieves significant strengths in the real Australian dataset.