Abstract

An accurate value of a cell’s state (i.e. its State of Charge (SOC)) is very important in the field of Hybrid Electric Vehicles (HEVs), since in contrast to Electric Vehicles (EVs) the battery is constantly charged and discharged during use and losing track of the SOC would lead to serious problems within the battery management unit. Unfortunately, SOC cannot be measured directly and therefore has to be estimated. This leads to a point where system modeling plays a role. The variety of approaches to cell modeling is wide, ranging from simple statistical models to neural nets to complex, physics based models. Using a Kalman Filter (KF) to estimate a systems state has already been tested by Gregory L. Plett and his results working with lithium-ion polymer cells are promising. This modeling approach is based on state space representations that model behavior and dynamics of the system (i.e. the cell). Detailed insights of the cell are not necessary, dynamic effects like hysteresis and relaxation can easily be taken into account. By taking SOC as a system state, once the model and its parameters are determined, a KF (or an Extended Kalman Filter (EKF) in the case of nonlinear state space equations) is used to predict the state of the system while running and corrects the estimate by exploiting the actual measurements. This only requires the values of the previous time step, there is no need to store more history.

This thesis investigates whether using different kinds of lithium-ion cell chemistries results in similar outcomes. Data for system identification and testing was obtained from standard cell tests performed at Magna E-Car (Graz). The models proposed by Plett were used and nonlinear system identification was done by applying an EKF. It appears that the KF considered by itself works well but system identification is crucial. Accurate estimates of the system parameters are absolutely essential for satisfying state estimation. Thus carefully determined experimental data is necessary. Otherwise system identification leads to faulty parameters which in turn cause inaccurate state estimation of the model.