Zhang Caiping

Effects analysis of model parameters uncertainties on battery SOC estimation using H-infinity observer

The accurate estimation of the state-of-charge (SOC) of battery is the basic premise for the effective energy management and the important guarantee for the safe and efficient operation in electric vehicles. To improve SOC estimation accuracy and robustness, the paper analyzes the effects of different initial SOC errors and parameters variation on SOC estimation accuracy and robustness with H-infinity observer. A model in Matlab/Simulink is established to make calculation process come true, which is based on a new battery with nominal capacity of 90Ah. The simulation and experiment al results indicate that the H-infinity observer based SOC estimation can converge to the true values quickly even if at the set maximum initial error and its steady error can be controlled within 2%. The effects of model parameters change resulted from battery degradation including SOC-OCV relationship, capacity and internal resistance on H-infinity observer are investigated. It is concluded that SOC estimation accuracy with H-infinity observer largely depends on the accurate of the curve of SOC-OCV (Open circuit voltage, OCV), which provide a foundation for battery management.

Capacity estimation of large-scale retired li-ion batteries for second use based on support vector machine

Capacities of li-ion batteries are difficult to estimate quickly and accurately when second use batteries are in large scale with dispersed parameters. This may result in costing too much time and money to reuse batteries. By analyzing the capacity and resistance characteristics, there is no functional relationship between them. In order to solve this nonlinear problem, a SVM model with 3 parameters (i.e. penalty coefficient, kernel function parameter and loss function parameter) is established. The inputs are five calculated values of resistance and the output is the capacity value of batteries. Data of 70 battery modules are adopted to train the model while data of other 30 battery modules are used to test the model. In this process, two parameters optimization methods using genetic algorithm have been proposed. By comparison, the coefficient of determination (COD) value of method 2 is higher than method 1 both in training model and testing model. The average error of method 2 between measured values and estimated values is 0.67% whereas that of method 1 is 1.35%. In method 2, 90% of the estimation errors are under 2.5%. The results provide valuable insights for large-scale retired li-ion batteries into second use.