Short-Term Net Load Forecasting Under High PV Penetration Based on ICEEMDAN-BO-BiGRU

To address the strong non-stationarity, coupled multi-scale fluctuations, and insufficient parameter adaptability of net load forecasting models under high photovoltaic penetration, an ICEEMDAN-BO-BiGRU combined short-term net load forecasting model is proposed. First, the net load sequence is constructed from the actual load and photovoltaic output, and then decomposed into eight intrinsic mode function components and a residual component using improved complete ensemble empirical mode decomposition with adaptive noise. Second, a bidirectional gated recurrent unit forecasting submodel is developed for each component, and Bayesian optimization is employed to adaptively optimize the key hyperparameters. Finally, the forecasts of all components are linearly summed to obtain the final net load forecast. Experiments are conducted using the CN03 sample from the HEEW dataset. The results show that, compared with CEEMDAN-BiGRU, the proposed method reduces the MAE and RMSE by 37.84% and 35.73%, respectively, over the complete February 2022 test period, while achieving a mean R2 of 0.9783, indicating lower observed forecasting errors under the reported experimental configurations.