Research On State Estimation And Life Prediction Methods Of Power Battery

Based on the current depletion of traditional energy and the environmental pollution caused by it,it is urgent to explore new green energy.As a typical battery type in power batteries,lithium batteries are widely used in smart devices,new energy vehicles and other fields with their advantages of high energy density,stable performance,low pollution and long cycle life.During the operation of the lithium battery,the battery performance will gradually degrade with the accumulation of by-products of the electrochemical reaction inside the battery,and the battery life cannot reach the ideal value.When the battery life expires and is not replaced in time,it will cause major safety accidents and threaten human life and property safety.Therefore,it is of great significance to predict the state of health(SOH)and remaining useful Life(RUL)of the battery.Both SOH and RUL are variables that characterize the deterioration of battery life.The article is mainly based on data-driven,using deep learning and transfer learning methods for SOH estimation and RUL prediction.The main research contents are summarized as follows:First of all,this article summarizes the domestic and foreign research status of lithium battery RUL from three different types of prediction methods,expounds the internal structure and principles of lithium batteries,and then builds a life decay test platform based on laboratory battery test instruments,analyzes the basic characteristics of lithium batteries through experimental data,and studies the influence of different discharge rates,temperature and other factors on the battery life degradation process.Then,this paper constructs a neural network model with Long Short-Term Memory(LSTM)structure as the basic processing unit to estimate the SOH of lithium batteries.The LSTM network structure has strong processing capabilities for long-term sequences,and can achieve SOH estimation by learning the correlation between the battery data sequence before and after,fitting the battery capacity degradation trend.In order to verify the performance of the model,this paper also uses the BP neural network to build a prediction model and conducts a comparative analysis of model performance based on the same prediction starting point.The experimental results prove that the LSTM model has lower errors.Finally,in view of the long time span of sample data leading to long training time and small number of samples leading to large neural network prediction errors,this paper integrates the concept of transfer learning on the basis of the LSTM model,and divides the data field based on the battery type: source domain and target domain,using the maximum mean discrepancy(MMD)measures the distribution distance between transfer learning domains,and adjusts the transfer model according to the distribution distance,so as to transfer knowledge between domains.Experiments have proved that the pre-training model obtained after the source domain data is trained has fully learned the capacity degradation trend,and can be merged with the target domain knowledge after migration,thereby improving the prediction performance of the model and the network training efficiency,and reducing the error rate.