Dynamic temporal scale decoupling and variable interaction modeling for photovoltaic power forecasting

Accurate photovoltaic (PV) power forecasting is essential for smart-grid operation and renewable energy integration, but existing models often struggle to capture short-term fluctuations, long-range nonlinear dependencies, and changes in meteorological conditions at the same time. To address these challenges, this paper proposes IBWO-STDiT, a hybrid deep learning framework for PV power forecasting. First, Pearson correlation analysis and Isolation Forest are integrated to screen relevant variables and remove anomalous records, improving the reliability of real-world PV inputs. The forecasting backbone then performs temporal-scale learning through a dual-branch CNN-xLSTM architecture, where the CNN branch captures local short-term fluctuation patterns and the xLSTM branch models long-range nonlinear dynamics using improved gating and memory structures. To balance branch contributions under different input conditions, a Dynamic Gated Fusion (DGF) module reweights the two representations at the sample, time-step, and variable levels. The fused representation is further transformed into compact variable-level tokens and processed by an encoder-only interaction module to model relationships among meteorological variables and historical PV power. Finally, an Improved Beluga Whale Optimization (IBWO) algorithm is employed to optimize key hyperparameters. Experiments on a real-world PV dataset covering four seasonal subsets show that IBWO-STDiT consistently outperforms the comparison models within each seasonal subset. Specifically, the proposed model achieves aggregate relative RMSE and MAE reductions of 38.04% and 43.74%, respectively, compared with the strong xLSTM–Encoder Part baseline, while obtaining R 2 values higher than 0.990 in all four seasonal subsets. Additional optimizer comparisons, fusion ablation studies, and variable interaction analyses support the value of hyperparameter search, adaptive feature fusion, and variable-wise interaction modeling for PV forecasting under different seasonal conditions.