A Research Of Adaptive Estimation Method Of Battery SOC Based On Data Driven Model

Because lithium battery has some special advantages,such as high power density and environmental protection,it has become the main energy storage medium of pure electric vehicles.The prediction of the state of charge(SOC)is an significant parameter of battery.The precise estimation of SOC is conducive to improve the energy conservation and avoid overcharge and over discharge,extending working life of the battery.Lithium battery is a highly complex nonlinear time-changed system which is heavily affected by environmental temperature.And in the process of driving electric cars,the acceleration and deceleration are repeated.The data distribution differences caused by these factors produce great difficulties for SOC estimation.Due to the change of battery and different operating conditions,there are great differences in data distribution.It will affect the reliability of SOC estimation model with different data distribution between training set and testing set.The dynamic changes of data distribution will lead to the performance degradation of SOC estimation model trained with previous data.To solve this problem,adaptive SOC estimation algorithms were studied to realize accurate SOC estimation under changing working conditions.Specific work list as follows:(1)The experiment focus was on the influence of temperature,charge-discharge rate and battery degradation upon battery capacity.The input features were determined based on the study of measurable parameters.The temperature and discharge rate were introduced into SOC calibration.A calculation method of the dynamic coulombic efficiency was proposed which was more in consistent to reality.(2)The adaptability of SOC estimation model to the changing of working conditions was improved.Incremental support vector regression algorithm was applied to SOC estimation which adapted to the change of working conditions through incremental learning.The sample forgetting mechanism was improved to make the efficiency of incremental learning better.Based on the principle of error driven and the method of weighted sampling,duplicate samples and random samples were eliminated.(3)Considering that incremental support vector regression algorithm would forget historical knowledge in the process of incremental learning,incremental LSTM algorithm was applied to SOC estimation.The forgetting gate of LSTM was used to compress the historical sequence data.And the particle filter algorithm was introduced to learn the network parameters.The compression information and the state of particle swarm would be used for subsequent learning.In order to solve the problem of sample exhaustion in parameter learning,classification resampling technology was also mainly studied.Finally,simulation experiments were carried out with the discharge data under different temperatures and working conditions.The results showed that the concept drift phenomenon in battery data distribution led to the degradation of SOC estimation effect of traditional models.And the incremental support vector regression model reduced the average absolute error from 1% to less than 0.5% by incrementally learning added data.Through the sample forgetting mechanism,the incremental support vector regression model not only improved the learning efficiency,but also further reduced the SOC average absolute error by 0.044%.The experiment also compared incremental support vector regression and incremental LSTM.It was proved that incremental LSTM could maintain better memory for historical data after continuously learning data under various working conditions.