Multi-scale patch and frequency-domain gated learning for high-resolution day-ahead photovoltaic forecasting

Accurate day-ahead forecasting of photovoltaic (PV) power under high temporal resolution is crucial for the reliable operation of smart grids. However, long-sequence PV data often exhibit strong coupling between global trends and local fluctuations, posing significant challenges to traditional forecasting methods. To address this, we propose a novel forecasting framework that integrates trend decomposition, multi-scale patch segmentation, and frequency-domain gated learning. The framework first decomposes the PV time series into trend and residual components using a dual-level structure. The TrendNet module captures long-term patterns via daily cycle-based segmentation and normalization, employing average pooling, a channel attention mechanism, and frequency-domain modeling with a parallel gated network (PGN) enhanced by fast Fourier transform (FFT). Meanwhile, the ResidualNet module focuses on short-term fluctuations by applying multi-scale patch division to the residual component, enabling localized temporal feature extraction. These two branches are trained in separate feature spaces and later fused to generate final predictions, allowing the model to effectively learn and integrate both long-range dependencies and short-term variability. Extensive experiments on multiple real-world PV datasets with time resolutions from 1 hour to 5 minutes demonstrate the model’s strong generalization ability and superior performance across different temporal granularities, highlighting its practical applicability for high-resolution PV power forecasting.