Abstract: Accurate short-term load forecasting is an important guarantee for achieving lean operation and management of the power grid. However, there are difficulties in precise forecasting such as short-term load variability and selecting key factors of load forecasting. Variational mode decomposition is used to decompose the original power load data into multiple sub-sequences, to mine short-term load change characteristics while avoiding mode aliasing problems, and a complex variable selection algorithm is proposed to analyze and screen the key factors affecting load changes, effectively eliminating undesired data and further simplifying the complexity of the prediction model. Each sub-sequence is predicted through the long and short-term memory neural network that takes into account the short-term and long-term dependence of data, and merges the prediction results of each sub-sequence to achieve the final short-term load forecast, and a short-term load forecasting method is built accordingly based on variational modal decomposition and selection of complex variables. The verification results of the actual data of Changsha City selected for the entire year of 2019 show that the algorithm proposed here can accurately select the key influencing factors of load forecasting under complex external influence factors. Compared with the traditional forecasting models, the proposed model structure is simpler and the prediction accuracy is higher.