There is an issue of decreased accuracy in load forecasting due to the complexity of load characteristics, so a short-term load forecasting model based on GWO-VMD and cascaded MCNN-MMLP residual networks is proposed. First, the original load data is processed using variational mode decomposition (VMD) optimized by grey wolf optimization (GWO) to reduce the complexity of the original load data. Secondly, a dual residual neural network combining multiscale convolutional neural networks (MCNN) and a multi-layer perceptron (MLP) is employed for transfer learning training and prediction of each mode. A multi-head attention mechanism is introduced into the MLP network to address the information bottleneck issue. Lastly, the MCNN-MMLP double residual model is used to predict and correct the errors in the preliminary prediction, thereby further improving the accuracy of the forecast. The proposed model outperforms the traditional and intelligent load forecasting models. It achieves a mean square error of 5.024 MW2, root mean square error of 2.241 MW, mean absolute percentage error of 0.160%, and coefficient of determination of 0.996 based on actual load data.