Self-supervised pretraining for PV power forecasting: Toward enhanced interpretability

Accurate photovoltaic (PV) power forecasting is essential for stable and efficient grid operation. To address the challenges posed by coupled meteorological and temporal variability, this study proposes an interpretable two-stage forecasting framework integrating self-supervised TabNet pretraining with a multilayer perceptron (MLP) regressor. In the first stage, TabNet jointly learns meteorological and temporal representations through a masked self-supervised task, while its sparse attention mechanism enables automatic feature selection and quantitative attribution. In the second stage, the pretrained representations are transferred to an MLP for efficient regression forecasting. Experiments on a 50 MW grid-connected PV plant demonstrate significant performance improvements. On the annual dataset, the proposed model reduces the root mean square error (RMSE) and mean absolute error by 42.7% and 39.2%, respectively, compared with the long short-term memory model. Under cloudy and rainy/snowy conditions, the model achieves additional RMSE reductions of 20%–30% relative to conventional baselines. Multi-horizon forecasting further demonstrates slower error accumulation and consistently superior accuracy. Overall, the proposed TabNet-MLP framework provides an effective and interpretable solution for multi-timescale PV power forecasting.