A Dual-Channel Filtering Network for Day-Ahead Photovoltaic Power Forecasting

The volatility and uncertainty of photovoltaic (PV) power generation pose significant challenges to the safe and stable operation of power grids. In the context of day-ahead scheduling, accurate PV power forecasting is particularly critical. This paper proposes a deep forecasting network based on dual-channel frequency-domain filtering, which integrates Reversible Instance Normalization (RevIN), seasonal-trend decomposition, and a Transformer encoder to enhance prediction accuracy and robustness. Specifically, the input time-series data are first normalized using RevIN to mitigate distributional discrepancies across different time periods. The original sequence is then decomposed into low-frequency trends and high-frequency residuals via seasonal-trend decomposition. These two components are individually transformed and filtered in the frequency domain to precisely extract low-pass and high-pass information. Finally, a Transformer encoder is employed to fuse multi-channel features and generate day-ahead forecasting results. Experimental results demonstrate that the proposed method outperforms conventional deep learning forecasting models across multiple evaluation metrics.