As the scale of offshore photovoltaic power generation continues to expand, the power output is significantly influenced by the complex marine environment, exhibiting strong non-linearity and non-stationarity. In response to the challenges encountered in current methods for predicting offshore photovoltaic power, this paper proposes a prediction model based on CEEMDAN-LSTM-XGBoost. This model performs preprocessing of Clear Sky Index (KPV) data and constructs three-dimensional features from meteorological, temporal, and historical power data. It utilizes Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) for multi-scale decomposition, builds a long short-term memory network to capture long-term dependencies in time series, and incorporates Extreme Gradient Boosting (XGBoost) to address nonlinear features and short-term fluctuations. Through a three-tier fusion architecture, the model further enhances prediction performance. The experiments, based on the 2023 measured data from an offshore photovoltaic power station in Jiangsu, compare the proposed model with benchmark models such as LSTM-XGBoost, CEEMDAN-LSTM, and CEEMDAN-XGBoost. The results show that the proposed model consistently performs the best across all seasons, with R2 values exceeding 0.99, which holds significant practical implications for power system scheduling and renewable energy integration.
Li et al. (2026) studied this question.