基于GAN-QRCNLSTM与高分辨率数据重构的日前光伏功率概率密度预测

In solar energy systems, accurate photovoltaic (PV) power forecasting is intricately linked to meteorological factors. Day-ahead prediction based on similar weather scenarios is pivotal for enhancing forecasting performance. However, existing approaches are constrained by simplistic similar sample selection methods, inefficient utilization of PV datasets, and inadequate attention to generation uncertainty. This study proposes a day-ahead probabilistic forecasting model that improves input quality via a high-resolution multidimensional adaptive similar moment selection. In addition a Generative Adversarial Network (GAN) learns the joint distribution of historical measurements and Numerical Weather Prediction (NWP), reducing redundancy and enriching training data. The model employs a multibranch architecture, integrating Convolutional Neural Networks and Long Short-Term Memory networks for spatio-temporal feature extraction. It also leverages Quantile Regression with Kernel Density Estimation to generate multiquantile predictions and comprehensive probability distributions. Experiments on authentic PV datasets demonstrate that the proposed GAN-QRCNLSTM model outperforms baseline models, achieving superior accuracy and robustness, particularly in complex weather conditions such as extreme weather scenario.