A remaining useful life estimation method based on long short-term memory and federated learning for electric vehicles in smart cities.

In modern society, environmental sustainability is a top priority as one of the most promising entities in the new energy sector. Electric vehicles (EVs) are rapidly gaining popularity due to their promise of better performance and comfort. Above all, they can help address the problem of urban air pollution. Nonetheless, lithium batteries, one of the most essential and expensive components of EVs, have posed challenges, such as battery aging, personal safety, and recycling. Precisely estimating the remaining useful life (RUL) of lithium battery packs can effectively assist in enhancing the personal safety of EVs and facilitating secondary trading and recycling in other industries without compromising safety and reliability. However, the RUL estimation of batteries involves many variables, and the operating conditions of EV batteries are highly dynamic as they change with the environment and the driving style of the users. Many existing methods exist to estimate the RUL based on batteries' state of health (SOH), but only some are suitable for real-world data. There are several difficulties as follows. Firstly, obtaining data about battery usage in the real world takes work. Secondly, most of these estimation models must be more representative and generalized because they are trained on separate data for each battery. Lastly, collecting data for centralized training may lead to a breach of user privacy. In this article, we propose an RUL estimation method utilizing a deep learning (DL) approach based on long short-term memory (LSTM) and federated learning (FL) to predict the RUL of lithium batteries. We refrain from incorporating unmeasurable variables as inputs and instead develop an estimation model leveraging LSTM, capitalizing on its ability to predict time series data. In addition, we apply the FL framework to train the model to protect users' battery data privacy. We verified the results of the model on experimental data. Meanwhile, we analyzed the model on actual data by comparing its mean absolute and relative errors. The comparison of the training and prediction results of the three sets of experiments shows that the federated training method achieves higher accuracy in predicting battery RUL compared to the centralized training method and another DL method, with solid training stability.