Privacy-Preserving Federated-Learning-Based Net-Energy Forecasting

Abstract

Energy forecasting not only enables infrastructure planning and power dispatching but also reduces power outages and equipment failures. To preserve the customers’ privacy, federated learning (FL) can be used to build a global energy forecasting model where customers train local models on their data and only send the models’ parameters to the utility server. However, FL may still leak customers’ data privacy because revealing the model’s parameters enables adversaries to launch attacks such as model inversion and membership inference. Moreover, most existing works only focus on load forecasting while energy forecasting for net-metering systems has not been well investigated. In this paper, we address these limitations by proposing a privacy-preserving FL-based energy forecasting model for net-metering systems. First, based on the analysis of real power consumption and generation readings, we design a hybrid deep learning (DL)-based energy forecasting model to provide an accurate prediction. Then, we develop an efficient data aggregation scheme to preserve the customers’ privacy by encrypting their models’ parameters during the FL training. Our extensive experiments’ results demonstrate that our predictor is accurate and our data aggregation scheme provides privacy preservation with high communication efficiency.