Virtual experiments for battery state of health estimation based on neural networks and in-vehicle data

To ensure the safety, performance, and warranty of electric vehicles, it is crucial to monitor the evolution of the state of health of lithium-ion batteries. Estimators for the state of health are often based on costly, time-consuming, and predefined testing procedures under laboratory full cycling conditions. In contrast, automotive operating conditions are highly volatile and thus cannot be interpreted by laboratory feature extraction methods. Given a rapidly growing fleet of electric vehicles and a limited number of battery test facilities, the need for alternative and scalable methods to determine state of health is essential for future developments. In this paper, we present a novel data-driven approach for battery state of health estimation based on the virtual execution of battery experiments. Therefore, an LSTM-based neural network learns the electrical behavior of an automotive battery cell based on in-vehicle driving data. This LSTM model is then used to simulate the electric response during capacity testing, incremental capacity analysis, and peak-power testing, which are explicitly designed for automotive lithium-ion batteries and adapted to real-world customer usage. Results show state-of-the-art accuracy for state of health estimation in terms of internal resistance (1.77% MAE) and remaining capacity estimation (0.60% MAE). This virtual execution of battery experiments is scalable, saves laboratory effort and test facilities, and in return requires only operational driving data.