TE-SDNet: Temporally Enhanced Sequential Dependency Network for photovoltaic power forecasting

Accurate forecasting of photovoltaic (PV) power output is essential for ensuring the reliable and economic operation of grid-connected PV systems. However, due to the inherently strong fluctuations, pronounced periodicity, and the influence of various complex environmental factors on PV generation, existing time-based or variable-based forecasting models often fail to fully capture and exploit these characteristics. As a result, their accuracy and stability in PV power forecasting remain limited. To address these challenges, this paper proposes TE-SDNet, a three-component forecasting framework that efficiently exploits both temporal and variable-level information for short-term PV power prediction. The Temporal–Feature Decoupling Block (TFD) proposes a two-stage convolutional network to independently capture temporal dependencies and enrich feature representations. The Multi-Scale Temporal Mixing Block (MTM) introduces frequency-domain analysis to effectively identify salient periodic patterns and performs their decomposition and aggregation dynamically. In addition, the Inter-Variable Dependency Learning Block (IDL) is proposed to capture cross-variable correlations through a dedicated attention network operating along the variable dimension, enabling the framework to generate more accurate predictions. To evaluate the effectiveness of the proposed framework, extensive experiments are conducted using data collected from fourteen PV plants located in two different regions of China. Averaged over two PV datasets and three forecasting horizons, TE-SDNet reduces RMSE by 2.2% and MAE by 3.1% relative to the strongest baseline, PatchTST, while requiring 94.6% fewer parameters and 99.6% fewer FLOPs. The source code of TE-SDNet is publicly available at https://github.com/Wokanood/TE-SDNet .