Federated learning-based prediction of electric vehicle battery pack capacity using time-domain and frequency-domain feature extraction

The rise of big data technology presents new opportunities for unified monitoring of the health status of electric vehicle (EV) battery packs. However, privacy, security, and the complexity of real-world data pose significant challenges. To address these, we propose a novel federated learning-based approach. First, domain knowledge is leveraged to extract labeled capacity data and key features that characterize capacity degradation trends from extensive real-world EV datasets. Next, we develop a hybrid forecasting model combining Convolutional Neural Networks (CNNs) and Fourier Neural Network (FNN) to capture both time-domain and frequency-domain features of capacity degradation. The model operates within a Federated Learning (FL) framework, ensuring data privacy by enabling local training of time series models at each node and central parameter aggregation using the Federated Averaging (FedAvg) algorithm. This collaborative setup avoids direct data sharing while effectively integrating global insights. The trained model is validated using charging data from 20 EVs, demonstrating superior performance and robustness compared to baseline and sub-models. The proposed method offers a promising solution for accurate, privacy-preserving battery capacity prediction, enhancing the management of EV battery health in real-world scenarios.