With the rapid development of smart grid technology, the speed, accuracy and stability of short-term power load forecasting are required. Aiming at the problems such as strong randomness of load in intelligent electricity environment, poor accuracy of short-term power load prediction and long calculation time when the amount of data is small, a combined prediction method based on random walk (RW) and improved sparrow search algorithm (SSA) to optimize generalized regression neural network (GRNN) was proposed. The model uses multiple inputs and single outputs. The input is load data and meteorological information, and the output is the predicted hourly load values. The RW was introduced to disturb the location of the sparrow, avoid falling into the local optimal value, and further improve its global search ability. The improved SSA was used to optimize the smoothing factor of the GRNN, and improve the self-learning ability, stability and accuracy of the model. The actual load data of a branch line in Xi'an, Shaanxi province is used for forecasting and verification. The results show that the improved algorithm has better convergence ability, and the model has better prediction accuracy.