Comparative Study of State-Of-Charge Estimation with Recurrent Neural Networks

The key to the safe and reliable operation of an electric vehicle is the precise knowledge of the state-of-charge (SOC)of its energy storage system. Since this state variable is not directly observable, it is derived from other quantities using model-based techniques such as Kalman filtering. More recently, data-based approaches using machine-learning have become feasible, promising the potential of reduced modeling effort and therefore reduced implementation time under similar or even higher accuracy. Many studies in this field, however, utilize single cell battery data and an extract of operating conditions that limits the reliability of the deduced conclusions for real-world applications. This paper provides an evaluation of recurrent neural networks with long short-term memory cells for SOC estimation in comparison to extended Kalman filtering based on equivalent circuit models. The analysis is conducted using the measurements of a battery module with 60 parallel connected cylindrical 18650 lithium-ion cells, a wide ambient temperature range between 0°C and 40°C and the deployment of a battery management system.