Data-Driven Battery Capacity Estimation in Electric Vehicles: Insights from Large-Scale Real-World Data

Abstract Global road transport is responsible for 16% of worldwide $$\hbox {CO}_2$$ emissions. The use of EVs can mitigate health issues caused by tail-pipe emissions and reduce the energy requirement by up to 50% per kilometer. Ensuring the longevity of the lithium-ion batteries often used in EVs is crucial in making EVs a viable and environmentally friendly alternative. The battery management system regulates the use of these batteries and is responsible for estimating the battery capacity. Optimizing the performance of these systems ensures safe operating conditions of the battery. This can reduce unnecessary battery wear. This study aims to explore the real-world performance and implications of a data-driven machine and deep learning approach for estimating the battery capacity in EVs. A comparison between ARIMA(X), XGBoost, LSTM and TCN using a large-scale real-world dataset is presented. The LSTM network achieves the lowest RMSE (1.42), MAE (1.09) and MAPE (2.70%). It showcases its ability to accurately model the data, achieving stable performance for vehicles with different mileages. The LSTM model performed worse for cars with a mileage between 0 and 50,000 kms achieving a 2.17 RMSE, this is partially caused by a limited number of high-capacity training labels. These results demonstrate the ability of data-driven models to accurately predict battery capacity for real-world data. However, further identifying the precise factors at play is crucial to understanding the LSTM behaviour and data patterns.