Online state of health estimation of lithium-ion batteries based on Long Short Term Memory network for hybrid and electric vehicles

State of health (SOH) is a key parameter of lithium-ion battery. As the battery ages an accurate SOH estimation plays a important role in battery management system (BMS). This thesis refers to a dataset from a laboratory test campaign conducted at LIM to train a LSTM network for estimating the SOH variation. The dataset is composed by 55 charging-discharging cycles and each of them has 43 different usage profiles. In order to age the battery, the entire test cycle lasted for several months. SOH decayed from initial 100% to 82%. In order to test the robustness of the network it uses three different profiles: charging profile, pulse profile and polarized random walk profile. To catch the SOH degradation information, several features are computed from original voltage and current profile. They are: state of charge (SOC), variation of state of charge (dSOC), state of energy (SOE), variation of state of energy (dSOE) and variation of voltage (dV). Move sliding window approach is used for feature extraction. This approach can smooth the data and reduce the total amount of data while preserving the data information, thereby increasing the training speed. LSTM is selected neural network for this thesis. Three different profiles were trained by LSTM and all achieved high accuracy. The last section is focused on the implementation in Simulink. In order to better match the real working conditions, PRW profile is selected for Simulink modeling. Finally, the developed SOH estimation algorithm is combined with SOC estimator. The SOH and SOC estimator can interact with each other and use each other's data as input values. The model gets accurate results on PRW profile and charging profiles. The overall accuracy is almost equal to 1.99% for SOH estimation.