With the development of smart grid technology and the advance of the power market, the complexity of power consumption patterns has gradually become prominent. This then makes higher demands on the accuracy and stability of short-term load forecasting. Given the lack of comprehensive consideration of time series data correlation and eigenvalues in traditional load forecasting methods, a combined forecasting model based on optimization of variational mode decomposition (OVMD), minimal redundancy maximal relevance (mRMR) and a long short-term memory neural network (LSTM) is proposed. First, the load sequence with high fluctuation is decomposed into a group of relatively stable modal components, in which the parameter of VMD is optimized by a sparrow intelligent algorithm. Secondly, the mRMR method is used to analyze the correlation between each modal component and the input feature set elements of the prediction model, obtain the optimal input feature set of each prediction model, and introduce the real-time electricity price into the load impact factor analysis. Finally, the LSTM method with different structural parameters is used to predict each component separately, and the predicted results are superimposed to obtain the final predicted value. An example is given to analyze the actual operational data of Australia, and compared with the conventional load forecasting method, the validity of the method is verified. This work is supported by the National Natural Science Foundation of China (No. 51667018 and No. 52067020).