Federated Rank Learning with Dimensionality Reduction and Clustering for Electricity Load Forecasting

Power load forecasting is crucial for power companies’ planning and power dispatching. With the development of machine learning, power forecasting has adopted artificial intelligence techniques based on machine learning. In this paper, we propose a novel forecasting scheme, FRLDRC, which combines the UMAP dimensionality reduction method, the K-means clustering algorithm, and ranking-based federated learning techniques. This approach allows us to obtain a forecasting model while ensuring data privacy, as the data does not leave its domain. To validate the effectiveness of the proposed model, we design experiments using over two million real household electricity consumption data points spanning four years. The experimental results demonstrate that data clustering with dimensionality reduction improves the performance of the baseline model. Additionally, the federated learning-based approach ensures data security, and the ranking federated technique further reduces communication overhead.