Edge-assisted Attention-based Federated Learning for Multi-Step EVSE-enabled Prosumer Energy Demand Prediction

Abstract

Energy demand prediction for the prosumer building, which is capable of playing the role of an electric vehicle (EV) charging station (EVCS) with installed EV supply equipment (EVSE), is currently of paramount importance for ameliorating energy efficiency and mitigating energy wastage. However, the time-dependency characteristics between successive energy demand data, the stochasticity of the number of EVs, and the randomness of the energy demand data of EVs and prosumers cause challenges in accurately predicting energy demand. Therefore, it is urgent to do energy demand prediction for prosumers. Nevertheless, energy demand prediction through centralized training is an extravagant process. This is because transferring energy data to a centralized machine for prediction will not only cause network bandwidth and energy consumption, but also cause communication delay. Thus, in this paper, an edge-assisted attention-based federated learning (FL) algorithm is proposed for multi-step energy demand prediction of prosumers, where the goal is to minimize the average forecasting loss. Specifically, since the attention mechanism has the advantage of detecting important features from inputs, to capture the temporal features and improve the prediction accuracy, the long short-term memory-utilized sequence to sequence model with the attention mechanism (LSTM-Seq2Seq-att) in FL setting is employed in each local edge server to train the global model collaboratively. The evaluation results clarify the effectiveness of the proposed method.