A Data-Driven Neural Network for State-of-Health Estimation in Lithium Ion Batteries

Abstract

With the imminent proliferation of renewable energy technologies, the need for better batteries has become increasingly important. There is a rapidly increasing number of batteries being deployed on the road in electric vehicle packs and on the ground in standalone and grid-connected energy storage units and this number will continue to increase as demand for renewable technologies escalates. There has been an acceleration of battery innovation in recent years in preparation for this energy revolution. While a lot of research has been focused on the manufacturing of batteries and making them less expensive, more efficient and longer lasting, other areas of research focus on battery management. This paper will present an empirical method for estimating State-of-Health, which is a measurement of the overall health and remaining life of the battery. The neural network used for this task will be trained using battery aging data from the NASA Prognostics Data Repository. The model will have inputs of voltage and temperature, as well as, average voltage and current over ξ precedent time steps. The model will output State-of-Health estimations in the form of max capacity over each cycle. The model exhibited errors as low as 2.16% MAE and 3.30% RMSE. Better State-of-Health Estimation will improve battery management processes and will reduce economic costs in battery manufacturing and operation due to uncertainties about life expectancy.