Good battery modeling is critical for energy management of electric vehicles and hybrid electric vehicles. Because of its simplicity and satisfactory performance, equivalent circuit models are widely used in this area. A frequently adopted equivalent circuit model is one that consists of an open-circuit voltage and a resistor in series with two sets of parallel resistor-capacitor combinations. This model performs well in describing battery transient behavior due to the dynamics of such physical phenomena as mass transport effects and double layer effects.
Generic methods for obtaining the parameters of this model involve analyzing the battery voltage behavior under step changes of load current. The fact that the model has two time constants places a challenge on parameter identification. Some most often used method makes use of the property that each of the two time constants plays a dominant role at different stages of the battery voltage response, and calculates the model parameters accordingly. For such method, the results are greatly influenced by the partition of faster and slower dynamics of the battery, and selection of the data points used for the calculation. Moreover, because majority of the testing data is not used towards calculating the parameters, the obtained model might not reflect the overall battery characteristics well and consequently might not give high-fidelity predictions. For other methods that use nonlinear curve fitting or genetic algorithm for parameter searching, the successful implementation greatly depends on the proper setting of initial values and searching space.
A novel method of parameter identification for the equivalent circuit model is presented in this paper. It makes use of a regression equation which is linear in variables that can be measured or calculated from the test. With this approach, all testing data during the relaxation period of a constant current pulse discharge or charge test is used towards obtaining the model parameters and the calculation is in the sense of least squared error. Application of the method to real battery testing data is presented. The example indicates that the method gives very good results with modeling error of less than 0.5%.