基于Kolmogorov-Arnold网络与timeGAN混合架构并结合考虑运行机制的数据增强的可解释光伏功率建模

Abstract The variable and stochastic nature of photovoltaic (PV) power generation poses significant challenges to its large-scale integration into power grids, limiting the full exploitation of solar energy as a clean power source. To address this issue, this study proposes a hybrid modeling framework that synergistically combines data preprocessing, feature expansion, and advanced deep learning architectures. First, the integrate Variational Autoencoder (VAE) is used for feature selection and dimensionality reduction, followed by seasonal and diurnal pattern division. The Ordering Points to Identify the Clustering Structure (OPTICS) algorithm is employed to identify intrinsic operational regimes, while TimeGAN is adopted to augment data diversity through synthetic generation of temporal features. The core innovation resides in the adoption of a dual-model architecture. For the first time, this architecture integrates the Transformer with the Kolmogorov-Arnold Network (KAN) to establish a hybrid framework. Comparative experiments involving ten benchmark models, including five standalone deep learning models and their KAN-enhanced variants, demonstrate the superiority of the proposed approach. The Transformer-KAN hybrid achieves 29.51 % and 12.0 % reductions in RMSE and MAE respectively compared to the standalone Transformer, while maintaining computational efficiency. This study provides an interpretable artificial intelligence solution for PV modeling that can effectively support renewable energy utilization optimization.