Aiming at the problem that the load data of residential areas with fewer users is too sharp and steep, and the model is difficult to fit and the accuracy is low in short-term load forecasting, a load forecasting method based on improved feature processing and hybrid neural network. Firstly, a periodic feature mapping method is proposed, and the periodic feature of load data are obtained by this method. Then, the high-frequency noise in load data is filtered by using the improved filtering algorithm based on the sand cat swarm algorithm (SCSO) to solve the problem that the load data is too sharp and steep. Finally, the SCSO-MHA-GRU (SMG) hybrid prediction model is built, and the multi-head self-attention mechanism is introduced into the GRU network to improve the ability of the model to capture multiple input feature relationships. In this paper, the real power load data of a district in Jiangsu Province is used for experiment. The results show that the proposed method has stronger fitting ability and higher prediction accuracy than other comparison models in short-term load forecasting of residential districts with fewer users.