Accurate photovoltaic power forecasting is crucial for the grid integration and dispatch of renewable energy sources and for ensuring the stable operation of power systems. Existing models inadequately capture physical laws and struggle to adapt to the strong periodicity, non-stationarity, and multivariate coupling characteristics of photovoltaic sequences. To this end, a data-driven PV forecasting framework (DPFF) has been established. First, by explicitly incorporating the day-night phase through physical encoding, the representation of time remains continuous and consistent while enhancing periodic variations. Second, utilize series decomposition to extract high-frequency seasonal components while suppressing non-stationary trend interference. Finally, a convolutional neural network autoencoder-based score attention is employed to capture the nonlinear interactions between photovoltaic power and meteorological variables, thereby enhancing the ability to fuse multi-source features. Experiments on Australian public datasets demonstrate that DPFF exhibits outstanding predictive performance across various prediction horizons. Compared to the linear baseline model, the mean absolute error decreased by ∼16.7%. Compared to transformers and their variants, the root mean squared error decreased by an average of 15%–30%.
Li et al. (Thu,) studied this question.