Abstract: Aiming at the problem of load prediction for power users under the protection of data privacy, we propose a power load prediction method based on improved federated learning algorithm. Firstly, a multi-user power load prediction framework based on lateral federated learning is constructed. On this basis, the traditional federated learning algorithm is not accurate enough and is vulnerable to malicious attacks, thus an innovative FedSTA (federated similarity training and aggregation) algorithm based on cosine similarity to optimize the local model update process and global model weighted aggregation is innovatively proposed. The calculation example results using actual load data show that the global model trained by the framework proposed in this paper has considerable prediction accuracy and certain generalization ability. In addition, compared with the FedAvg algorithm and the FedAdp algorithm, the FedSTA algorithm proposed in this paper significantly improves the accuracy of the global model trained. Finally, this paper verifies the robustness of the FedSTA algorithm and its ability to identify attacked clients. The results show that the algorithm can accurately identify the attacked clients and assign them a smaller aggregation weight. Compared with the FedAvg algorithm, the impact of the global model prediction accuracy is significantly reduced.