An integrated federated learning algorithm for short-term load forecasting

Accurate power load forecasting plays an integral role in power systems. To achieve high prediction accuracy, models need to extract effective features from raw data, and the training of models needs a large amount of data. However, data sharing will require the disclosure of the private data of the participants. To address this issue, we combined variational mode decomposition (VMD), the federated k-means clustering algorithm (FK), and SecureBoost into a single algorithm, called VMD-FK-SecureBoost. First, we used VMD to decompose the original data into several sub-sequences. This enabled us to extract the implied features to separately predict each sub-sequence to improve the prediction accuracy. Second, we use FK to recombine the sub-sequences into several clusters with common characteristics. Finally, with SecureBoost, we use clustering results to realize federated learning with privacy protection. We calculated the prediction values by accumulating the prediction results of the sub-sequences. The results for the examples in the US and Australia showed that the prediction performance of VMD-FK-SecureBoost was better than those of XGBoost and SecureBoost. Particularly, the MAPEs of one-step-ahead forecasting in the Texas and Newcastle CBD from our proposed method are 0.209% and 2.127% respectively, which are the lowest of all the algorithms.