Physics-informed Mamba network for ultra-short-term photovoltaic power forecasting: integrating WGAN-GP augmentation and CEEMDAN-SST decomposition

Minute-level ultra-short-term photovoltaic power forecasting is vital for real-time grid stability and renewable energy integration. However, forecasting performance is often hindered by limited samples under specific weather conditions, strong non-stationarity, and lack of physical consistency in data-driven models. This study proposes an integrated framework combining data augmentation, time-frequency decomposition and physics-informed modeling. A density-based clustering algorithm (HDBSCAN) is used to identify minority weather patterns, and an improved Wasserstein GAN with gradient penalty (WGAN-GP) augments time-series data within these clusters. A hybrid CEEMDAN-SST method is then applied to extract multi-scale features, in which dominant intrinsic mode functions are selected based on energy and correlation to enhance sensitivity to non-stationary dynamics. A physics-informed PINN-Mamba model is further developed by integrating physical constraints and a state-space architecture, ensuring both forecasting accuracy and interpretability. Experimental results demonstrate that the proposed method outperforms several mainstream benchmarks in both accuracy and generalization performance. Compared to a strong baseline model, the PINN-Transformer, the proposed model reduces MAE, RMSE and NRMSE by 38.16 %, 9.13 %, and 17.75 %, respectively, and improves R2 by 0.62 %, demonstrating its practical effectiveness. Furthermore, PINN-Mamba achieves an average inference latency of 1.19 ms, confirming its real-time applicability in ultra-short-term PV forecasting tasks.