Lightweight multiscale time-series mixing with periodic encoding and physical constraints for ultra-short-term photovoltaic power forecasting

Photovoltaic (PV) power generation is crucial to modern power systems, yet its output exhibits significant randomness, volatility, and periodicity due to factors like solar irradiance and cloud cover, posing severe challenges to grid dispatch. Existing ultra-short-term forecasting models often struggle to capture minute-level fluctuations, explicitly model diurnal/annual periodic patterns, and ensure physical consistency. To address these issues, this study proposes DTP-TSMixer, an ultra-short-term multi-step forecasting model based on the lightweight Time-Series Mixer (TSMixer). The framework integrates a multi-scale depthwise separable convolutional front-end to capture local fluctuations across different receptive fields, employs sine–cosine encoding to enhance intraday and annual periodic feature representation, and incorporates physical constraints (non-negativity, ramp rates, local upper bounds) during training to ensure forecast plausibility. Experiments on four PV systems from two independent public datasets with 5-min and 15-min sampling intervals show that DTP-TSMixer achieves the best overall performance across all evaluated systems. On the DKASC Site 1 test set, it achieves RMSE, MAE, and R 2 values of 0.4826, 0.2895, and 0.8914, respectively, reducing RMSE and MAE by 16.9 % and 26.1 % relative to the baseline TSMixer. Additional analyses across weather states and seasons, together with SHAP and LIME explanations, further support the robustness and interpretability of the proposed method.