Federated Learning for Long-term Forecasting of Electricity Consumption towards a Carbon-neutral Future

Abstract

In this paper, we propose an approach for long-term forecasting of electricity consumption based on federated learning. Basically, federated training was conducted on electricity consumption forecast models of several regions simultaneously, which can not only enrich training samples but also improve the generalization ability of the forecast model. More specifically, long short-term memory neural network (LSTM) is adopted as the local model for federated learning, in which carbon emission is one of the input features, so that the electricity consumption forecast results are more consistent with the carbon-neutral development path. In this study, we forecast electricity consumption of a certain area in China from 2022 to 2035, and experiment results verify the effectiveness of the proposed method compared with traditional time series method.