Research On State Estimation Of Li-ion Battery Based On Neural Network Method

In order to deal with the increasingly serious environmental pollution and the gradual reduction of fossil energy on a global scale,the new energy industry,especially electric vehicles,is increasingly being valued by governments.As the power source of electric vehicles,the research of power batteries has also become the top priority of the entire electric vehicle industry.Lithium-ion batteries are widely used in energy supply units of electric vehicles due to their unique advantages in terms of energy density,power density,cycle life,self-discharge rate,etc.However,behind the rapid development,there are many key issues that remain unresolved.For example,the internal state of lithium-ion battery cannot be directly measured,which requires a specific method to estimate it.The data-driven method directly starts from the test data of the battery,is simple to use and understand,and has a good application prospect.The neural network model based on the simple recurrent units overcomes the problems of gradient disappearance and gradient explosion.It has strong learning ability on medium and long-term data,and has a simple threshold and fast training speed.It is suitable for estimating the state of the battery.To this end,this article has done the following:Firstly,the ternary lithium-ion battery is used as the research object,the internal working mechanism and external characteristic parameters of the battery are analyzed,and the power battery is charged and discharged based on the built power battery test platform,and the external characteristic test data of the battery is obtained.These data provide detailed data support for the next step of estimating the state of lithium-ion battery.Secondly,in view of the complex structure and difficult training of the traditional recurrent neural network model,a simple recurrent units model was used to estimate the battery SOC.In this method,each update step only depends on the input at the current moment,which decouples the dependence on the output of the hidden layer at the previous moment,and the updates at different moments can be executed in parallel.Thus greatly reducing the computational complexity.Based on the characteristics of the model,the data suitable for training and verification are selected and used as input data after normalization and other processing.The experimental results show that,compared with the traditional neural network algorithm,the SRU has better accuracy and faster training speed when the hyperparameters are set the same.Then,the influencing factors of battery SOC estimation were analyzed and the SOC values were verified experimentally under different working conditions.Firstly,the effects of ambient temperature and aging on the battery capacity were analyzed from the aspect of working mechanism.The charge-discharge characteristics and capacity characteristics of the battery at different temperatures were analyzed and the characteristics of retired batteries,etc.Then,based on the experimental data,the neural network algorithm is used to estimate and compare the SOC of the battery based on different temperatures and decommissioned battery data.Finally,the simple recurrent units neural network model is applied to estimate the battery state of health.According to the characteristics of battery health state,its essence and influencing factors were firstly analyzed.Based on this,indicators suitable for battery health state estimation were selected,which mainly included average charging and discharging voltage,constant current charging time,voltage increment per unit time and capacity increment.Based on this data set,the battery SOH is estimated and briefly compared.