Research on Short-Term Photovoltaic Power Forecasting Based on ICEEMDAN–BiLSTM–Attention–IMSO

Abstract

Short-term photovoltaic (PV) power forecasting faces challenges from strong randomness and non-stationarity. This paper proposes a combined ICEEMDAN–BiLSTM–Attention–IMSO forecasting framework. First, ICEEMDAN decomposes the historical PV sequence into components, which are reorganized into three entropy groups based on sample entropy to reduce nonlinear interference. Second, a BiLSTM network with a self-attention mechanism extracts deep temporal features. Finally, a multi-strategy improved mantis shrimp optimization (IMSO) algorithm adaptively optimizes the hyperparameters for each sub-model. Case studies on the 2021 State Grid Corporation of China dataset confirm the model’s superiority. Compared with the fixed-hyperparameter baseline, the proposed framework reduces RMSE by 9.37% and MAE by 13.31%, outperforming combined models optimized by mainstream algorithms. Crucially, ablation experiments demonstrate that removing the sample entropy regrouping causes severe performance degradation, with RMSE surging to 8.6193 MW and R2 dropping to −0.2341. These findings, in conjunction with the comparative experiments, mutually demonstrate the substantial contribution and comprehensive effectiveness of the proposed framework for short-term PV power forecasting.