Stability-Driven Multistage Transformer–BiLSTM for Short-Term PV Power Forecasting

This study proposes an integrated short-term photovoltaic (PV) power forecasting framework, termed MSD-STB-OptFormer, to jointly address multiscale power fluctuations, unstable meteorological relevance, and model-parameter sensitivity. Successive variational mode decomposition (SVMD) first separates the original PV power series into components with different fluctuation characteristics, after which wavelet packet decomposition (WPD) further refines the reconstructed high-frequency component. A stability-driven feature-selection strategy then combines correlation-based redundancy screening with bootstrap perturbation and permutation importance to retain meteorological predictors with persistent predictive relevance. An improved whale migration algorithm (IWMA), incorporating fast random opposition-based learning and normal cloud mutation, is subsequently employed to optimize the hyperparameters of the Transformer-BiLSTM predictor. Second, a stability-driven feature selection strategy combines correlation-based redundancy screening with bootstrap-based stability evaluation to retain meteorological predictors with persistent predictive relevance. Third, an improved whale migration algorithm (IWMA), incorporating fast random opposition-based learning and normal cloud mutation, is used to optimize the hyperparameters of the forecasting model. Finally, the reconstructed multiscale components and selected meteorological features are fed into a Transformer-BiLSTM predictor, where self-attention captures long-range temporal dependencies and BiLSTM models bidirectional sequential dynamics; the component forecasts are subsequently aggregated to obtain the final PV power prediction. The framework is evaluated using real PV datasets under different seasonal and regional conditions. Comparative, cross-regional, capacity-normalized, and ablation analyses consistently show that the proposed framework improves forecasting accuracy, seasonal robustness, and generalization capability.