Multi-timescale photovoltaic station power prediction based on Reformer model

Photovoltaic (PV) power generation, as the primary technology for utilizing solar energy, faces challenges due to intermittency and volatility, which pose significant issues for grid scheduling and power system stability. To improve the accuracy of PV power prediction, this paper proposes a PV power prediction method based on one-dimensional wavelet convolution (WTC), Reformer, and Kolmogorov-Arnold Networks (KAN), incorporating multi-timescale integration of deep learning models. The method extracts cross-features at finer spatio-temporal scales via one-dimensional WTC. A lightweight design of Reformer is employed to reduce the model's complexity and enhance its ability to capture long-term dependencies. The KAN model, which learns univariate spline functions with superior nonlinearity mapping ability, is utilized to predict PV power. In this study, PV datasets from two different PV sites in Australia and a photovoltaic station in northern China are selected for 1-day, 3-day, and 7-day power prediction. The experimental results demonstrate that the proposed WTC-Reformer-KAN method exhibits outstanding prediction performance. This study provides a novel methodological approach for the operational and maintenance needs of PV sites in multi-timescale PV power prediction.