Accurate ultra-short-term photovoltaic (PV) power forecasting is essential for secure grid dispatch and renewable-rich system operation, yet it remains difficult because of rapid weather fluctuations and error accumulation in multi-step prediction. This paper proposes a decoupled physics-guided residual-correction framework, built on an attention-based sequence-to-sequence (Seq2Seq) architecture, for deterministic 4 h ahead rolling PV forecasting at 15 min resolution. In the first stage, a physical model maps numerical weather prediction (NWP) inputs to a deterministic baseline trajectory while preserving physical bounds. In the second stage, an Attention-Seq2Seq network learns the structured residual trajectory from historical sequences. The global attention mechanism allows the decoder to focus on the most informative historical states, helping reduce information loss and error accumulation over extended horizons. Experiments on a 22-month real-world PV dataset show that the proposed framework outperforms conventional linear and nonlinear benchmarks, reducing root mean square error (RMSE) and mean absolute error (MAE) by 23.79% and 39.17%, respectively, relative to the physical baseline. The framework also maintains robust instantaneous tracking under rapidly changing cloud conditions and preserves a 30–40% error reduction rate at Steps 12–16, supporting reliable intraday scheduling.
Yang et al. (Mon,) studied this question.