To address the issue of insufficient prediction accuracy caused by the high volatility and strong randomness of power load under the coupling effect of multiple factors, a short-term load forecasting model was constructed that integrates optimized VMD(variational mode decomposition) secondary recombination decomposition and a TCN-BiGRU-TPA(temporal convolutional network-bidirectional gated recurrent unit-temporal pattern attention) hybrid architecture. First, the maximum information coefficient was used to evaluate the correlation between weather factors, electricity prices, and power load, and the factors with higher correlation were selected as input features. Second, the complete ensemble empirical mode decomposition with adaptive noise was employed to initially decompose the load sequence into multiple subsequences, and the sample entropy of each subsequence was calculated. Subsequences with similar entropy values were recombined to form new sequences. The recombined high-frequency sequences were further decomposed using variational mode decomposition optimized by the slime mould algorithm, so as to fully extract the multi-scale and time-frequency characteristics hidden in the load data. Then, the selected feature variables and the decomposed components were fed into the TCN-BiGRU-TPA model for prediction. Finally, the Spanish Valencia regional-level dataset was used for experimental verification. Compared with the optimal benchmark model, the proposed model shows significant reductions in MAE (mean absolute error), RMSE(root mean square error), and MAPE(mean absolute percentage error)while the R² value is increased by 0.003. The experimental results demonstrate that the proposed model achieves higher forecasting accuracy.