State‐of‐Charge Estimation of Second‐Life Batteries Using the Federated Learning

ABSTRACT The advancement in automobile electrification has been causing concern about retired batteries. Since the battery cannot store and deliver energy for a high‐speed vehicle when its state of health (SOH) is below 80%, this battery must have another destination. A second application for energy storage from solar or wind farms can be a solution. However, a pack of second‐life batteries (SLB) is built in some cases from batteries from different first applications, which introduces the problem of the pack's heterogeneity—this makes applying a standard Battery Management System (BMS) difficult. As a solution, data‐driven and Machine Learning (ML) models have been used to help BMS for SLB, learning from the data. However, these models are specific to a single package. Therefore, this work applies Federated Machine Learning (FML) to build a global State‐of‐Charge (SOC) for SLB. The main intention is to develop a global model based on several learnings. This way, the proposed idea comprises three steps using 32 different SLBs with a capacity ranging from 1300 to 2500 mAh, which introduces a significant difficulty to the process. In the first step, a single BMS comprises a Raspberry Pi 3, and some sensors monitor the voltage, current, and surface temperature of a single SLB. The data are used to build a Long Short‐Term Memory network to predict the SOC. The process was done on 20 different SLB batteries. The results indicated an R 2 above 99 and an RMSE below 3%. After that, each BMS sends the weight of each model to the server. In the second step, the server receives the weights and builds a global model using the ensemble learning approach. The model inserted an exclusion rule to avoid outliers in the decisions. The global model is tested with four different SLBs. The validation process shows an R 2 of 94 and an RMSE of 5%. The global model has been downloaded to eight SLB batteries with BMS for testing. As the model has never seen the data of the eight batteries, the R 2 was 96, and the RMSE was below 5.6% on average. The obtained results indicate that the proposed method is capable of achieving reliable SOC estimation under heterogeneous SLB operating conditions. This way, future investigations will involve inserting this idea into a pack of more than one battery.