Li-ion battery State-of-Charge estimation using computationally efficient neural network models

Li-ion battery State-of-Charge (SOC) estimation is a complex challenge for battery management systems designers, due to the battery's non-linear behaviour at different operating conditions and ageing levels. As a possible solution, multiple machine learning models have been proposed for SOC estimation throughout the years. These provide an advantage over model-based methods, as they do not require a deep knowledge and study of the battery's internal behaviour. However, many of these proposed models could not be considered due to their complexity. The high number of required stored parameters and/or elevated memory consumption during estimation may pose challenges to the application of these methods. Therefore, in this paper, several feedforward neural network models are proposed for SOC estimation, with an efficient method for online input preprocessing and low parameter requirement in storage. These models are simulated and validated using battery data, taken at different temperatures with several driving cycles and charge cycles, achieving lowest estimation Root Mean Squared Error (RMSE) of 1.096% over the whole validation dataset.