Abstract: In view of the strong randomness and volatility inherent in wind and solar power generation—which render single forecasting models somewhat inadequate for handling their nonlinear and non-stationary characteristics—this paper proposes a combined short-term forecasting model for wind and solar power based on closed-loop complete ensemble empirical mode decomposition (CEEMDAN) and temporal convolution feature extraction. Firstly, CEEMDAN is applied to decompose the original wind and solar power generation data across multiple time scales, extracting high-energy intrinsic mode functions (IMFs) to effectively reduce the data’s non-stationarity and complexity. Then, a Temporal Convolutional Network (TCN) is employed to extract the temporal features of the wind and solar data, while a Bidirectional Gated Recurrent Unit (BiGRU) further captures the bidirectional dynamic characteristics of the time series; an Attention mechanism is also integrated to enhance the focus on key features, thereby constructing the TCN-BiGRU-Attention forecasting model. In addition, the Salp Swarm Algorithm (SSA) is introduced to optimize the model’s hyperparameters, and a feedback mechanism is used to dynamically adjust the decomposition parameters of CEEMDAN, further improving the forecasting performance. Finally, experiments conducted using wind and solar power generation data from a certain region demonstrate that, compared with other forecasting models, the proposed model achieves increases in R² of 2.26% and 2.79%, respectively, effectively enhancing the prediction accuracy.

Key words: power forecasting, closed-loop mode decomposition, salp swarm algorithm (SSA), temporal convolutional feature extraction, deep learning