Distilled one-shot federated learning for heterogeneous residential load forecasting

Abstract

Load prediction is the foundation of power system operation and analysis. It also plays a decisive role in the safety, stability, and economic operation of the smart grid. Traditional load forecasting methods are too difficult to meet the needs of the modern smart grid. As technology evolves, federated learning (FL) offers a new solution for load forecasting. However, in the training of global model, FL multiple iterations are needed to converge, which largely depends on the repeated transmission of model parameter updates between client and server. When the global model structure is complicated or the local data heterogeneity of each client is serious, more iterations are needed to converge the model. Therefore, we designed a load prediction method based on distillated one-shot federated learning to solve these problems. The proposed method distills the training data through a dataset distillation algorithm and uses the distilled data instead of gradients for transmission, greatly reducing communication costs. In addition, we propose a double-weight self-checking mechanism to avoid each client updating the global model of the server on average and improve the accuracy of load forecasting. The results of the British example show that the prediction performance and privacy of distillated one-shot federated learning are better than those of FedAvg and FedDM. Particularly, our proposed method requires only 0.28 MB of communication consumption during the training process, which is much lower than the above two algorithms.