An Improved Photovoltaic Power Forecasting Method Using Variational Mode Decomposition and Kolmogorov-Arnold Networks

With the large-scale development of photovoltaic (PV) power generation, its intermittency and uncertainty pose significant challenges to the stability of power systems. To address the complex multi-frequency characteristics and nonlinear modeling challenges in PV power forecasting, this paper proposes an innovative forecasting framework based on Variational Mode Decomposition (VMD), Long Short-Term Memory networks (LSTM), and Kolmogorov-Arnold Networks (KAN). First, VMD is employed to decompose the PV power time series into multiple frequency modes, and a multi-input multi-to-multi mapping structure is constructed to integrate modal signals and original power features. Second, LSTM is utilized to extract temporal dependencies, and KAN further enhances the nonlinear modeling capability through adaptive spline activation functions. Finally, a correlation-based weighted loss function is proposed to optimize the prediction accuracy of both multi-modal signals and original power. Experimental results demonstrate that the proposed VMD-LSTM-KAN model significantly outperforms several baseline models, highlighting its effectiveness in accurate photovoltaic power forecasting.