To address the challenges posed by the randomness and high volatility of multiloads in integrated energy systems, existing load forecasting methods often struggle to achieve high accuracy and stable prediction performance. To overcome this issue, this paper proposes a short-term load forecasting method for IES based on modal decomposition and multi-model fusion. First, the maximum mutual information coefficient is used for feature selection, aiming to effectively identify key factors closely related to load variation. Next, sample entropy combined with mutual information is employed as the fitness function, and the exponential triangular optimization algorithm is applied to obtain the optimal parameter combination for variational mode decomposition (VMD), enabling effective decomposition of IES loads into multiple intrinsic mode functions. Then, permutation entropy is used to filter the decomposition results and extract low-frequency and high-frequency components that reflect the load variation characteristics. Finally, a BiLSTM network is used to predict the low-frequency components, while a BiTCN-LPTransformer-BiGRU model is applied to forecast the high-frequency components. The final load prediction is obtained by aggregating the predictions of all components. Verification using actual load data, specifically for spring electricity load, shows that the model achieves an RMSE of 118.394 kW, an R2 of 0.991, and an MAPE of 0.351%. Compared to traditional models, this approach significantly improves prediction accuracy, validating the effectiveness of the proposed method.