Short-term residential load forecasting plays a basic role in economic power consumption and fault early warning, and its combination with federated learning has become the main trend of data sharing. Limited by the lack of sensitivity of deep learning methods to low-load data, load forecasting is difficult to improve the accuracy. At the same time, performance of some local models in traditional federated learning is poor. Firstly, a hybrid short-term residential load forecasting model (DCNN-AM-LSTM-AE) is proposed. Then, a personalized federated learning aggregation algorithm (FL-LF) is designed. By improving the local update method and equipping with aggregators, the FL-LF controls the global aggregation model to shift in a direction conducive to improving local prediction performance. The experimental results show that DCNN-AM-LSTM-AE is more sensitive to low load data, and the prediction results are better than other existing methods. The FL-LF can not only improve the prediction accuracy, but also ensure the model avail-ability of local users.