Cerebella Model Articulation Controller Based Method For The Battery State-of-Charge Estimation

Hybrid-Electric Vehicle (HEV) plays an increasing important role throughout the world in aspects of reducing motor vehicle exhausts and energy consumption. However, the power batteries, with short service life and high production costs, have become the bottleneck of HEV's large scale application, of which the key obstacle is the accurate estimation of battery state-of-charge (SOC).Owing to the strong nonlinearity that SOC shows during the charging and discharging process of battery, artificial neural network technologies have been widely promoted in the estimation of SOC. Based on existing technical literature, considering the time consuming for the training and learning process of general neural networks, this thesis induces Cerebella Model Articulation Controller (CMAC), which has fast rate of convergence and well local generalization ability, into the estimation of Mh-Ni battery SOC. The simulation results show that, comparing with BP (Back Propagation) neural network model, the train time consuming of CMAC model has greatly shortened, but the output error is obviously increased.Furthermore, aiming to improve CMAC output accuracy, this thesis deeply analyses the influence of generalization parameter and the standard deviation of receptive field function to single testing error, finding that the CMAC model with fixed parametric form has totally different output errors between intervals that have different rate of target output changing. Therefore, achieving to real-time adjustment of standard deviation or generalization parameter, a Parametric Receptive Function CMAC network is proposed in this thesis, and the updated algorithm of weights is also improved. The simulation results show that the output errors of PRCMAC model remain at relatively low stage, overcoming the weakness of traditional CMAC neural network which has low output precision.Finally, based on dynamic Gaussian receptive field function, this thesis establishes an approximate PRCMAC model to estimate the SOC of MH-Ni battery, and cyclic tests this model on Visual Studio platform to determine the best internal parameters with discharging sample data of Mh-Ni power batteries.The simulation results demonstrate the better convergence rate and high precision of the PRCMAC model, which can learn more sample data than other neural network models. So, the applicable scope of PRCAMC model is more extensive.