• Propose a ultra-short-term PV cluster power forecasting framework using causal feature clustering and deep ensemble learning. • Extract multi-scale PV power features via SVMD and build a dynamic causality matrix using nonlinear causality tests. • Develop a 3D-DTW-based dynamic K-Medoids clustering method with adaptive cluster number selection. • Build a CNN–BiLSTM–KAN hybrid model to improve forecasting accuracy across different seasons. The large-scale grid integration of photovoltaic clusters has introduced significant uncertainty and regulation pressure to power systems, making the improvement of photovoltaic power forecasting accuracy essential for the safe and stable operation of new power systems. To address the problems of large differences in meteorological conditions, complex spatiotemporal correlations, and strong output fluctuations within photovoltaic clusters, this paper proposes an ultra-short-term photovoltaic cluster power forecasting method based on causal feature clustering and a deep hybrid model. First, successive variational mode decomposition (SVMD) is used to extract multi-scale intrinsic modes from power sequences, reducing the influence of high-frequency random disturbances, and a dynamic causal matrix is constructed through nonlinear causality tests. Second, an improved K-Medoids clustering method based on 3D dynamic time warping (3D-DTW) distance is introduced to realize dynamic sub-cluster partitioning of photovoltaic stations, thereby enhancing the timeliness and physical consistency of the clustering results. Finally, a CNNBiLSTMKAN hybrid forecasting model is constructed, and the key parameters of each model are uniformly optimized. The model conducts deep feature modeling within clusters from three aspects: local feature extraction, temporal dependency modeling, and nonlinear mapping, and then aggregates the outputs to obtain the overall cluster forecasting results. Experiments are conducted on photovoltaic clusters in Jilin and Jiangxi Provinces. The results show that the proposed method achieves the best annual average forecasting performance in both provinces. Compared with traditional learning models, the RMSE of Jilin and Jiangxi Provinces is reduced by 16. 94% and 16. 71%, respectively, the MAE is reduced by 12. 88% and 25. 19%, respectively, and R 2 is improved in both cases. These results verify the effectiveness of the proposed method in improving photovoltaic cluster power forecasting accuracy under different regional and seasonal conditions, and further demonstrate the strong adaptability and generalization ability of causal feature clustering and deep hybrid modeling under complex weather disturbance scenarios.
He et al. (Fri,) studied this question.