Study of machine learning algorithms to state of health estimation of iron phosphate lithium-ion battery used in fully electric vehicles 2018-36-0178

State of Health (SOH) is an important parameter in Battery Management Systems (BMS) because it avoids the failure of a battery that could lead to reduced performance, operational impairment and even catastrophic failure, especially in electric vehicles. However a reliable battery state estimation management system in electric vehicles greatly depends on the validity and generalizability of battery models. This paper presents a generic data-driven approach for lithium-ion battery health management that eliminates the dependency of battery physical models for SOH estimation. In this work, iron phosphate Lithium-ion batteries were used. They were repeatedly submitted to charge-discharge cycles based on standard IEC and ISO profiles. The tension, current, charge, cell temperature and ambient temperature were constantly monitored in this period, and one big data set was created and stored. This data was then used to explore Machine Learning algorithms to estimate State of Charge (SOC) and State of Health (SOH) without explicitly delving into the physical modeling. Many of them produced rather poor results (Decision Trees, Support Vector Machines, Regression models), showing errors up to 50%. The best result was obtained using a Neural Network model called NARX (Nonlinear Auto-Regressive network with Exogenous inputs), which is an appropriate model to deal with time series. The 99% confidence interval obtained for the SOC estimation was -6.57% to 5.01%; for SOH, the 99% confidence interval was -0.64% to 0.34%.