Research On Two-dimensional Temperature-field Sparse-reconstruction Of Lithium-ion Battery Pack Based On Deep Neural Network

A large number of studies have shown that high temperature,low temperature or uneven temperature distribution will seriously affect the performance of lithium battery packs.It is very important to monitor the temperature of battery pack quickly and accurately.Direct temperature measurement may have the problem of redundant layout sensors,which makes the cost higher.Indirect temperature measurement often needs to establish electrochemical model or equivalent circuit model for thermal simulation.Simple model may lead to limited accuracy,while complex model may lead to difficult solution.Based on the above analysis,this paper proposes a two-dimensional temperature-field sparse-reconstruction technology of lithium battery pack,which combines direct measurement and indirect measurement.Firstly,a 128-channel temperature measurement platform was built,then a battery pack composed of 18650 lithium batteries was taken as a sample to collect 40089 temperature-field data in 5 discharge cycles.The data were visualized to analyze the characteristics of the temperature-field.LR,SVR and DNN were respectively built as the temperature-field reconstruction algorithms.The influence of different reconstruction algorithms,different DNN optimization algorithms,different numbers of real thermal sensors,and different sensor layout positions were discussed around the temperature-field reconstruction accuracy.Finally,a software of two-dimensional temperature-field sparse-reconstruction of lithium battery pack was developed,which integrated the functions of temperature-field analysis,modeling,training and online prediction.The results showed that:(1)DNN can accurately predict the temperature under different numbers of real thermal sensors,the maximum MAE is 0.6468℃(SCR=0.78%),the minimum MAE is 0.080℃(SCR=50.00%),and the average prediction time of single data is only 0.19ms.(2)Drop out can effectively improve the reconstruction accuracy of DNN.(3)When there are few real thermal sensors,DNN has obvious advantages in reconstruction accuracy compared with LR and SVR,which means that DNN can effectively reduce the number of sensors under the same reconstruction accuracy requirements.(4)Distributed and equidistant the layout of RTSs can effectively improve the temperature-field reconstruction accuracy.In this paper,a two-dimensional temperature-field sparse-reconstruction technique of battery pack based on DNN was proposed for the first time,which can provide accurate,robust and real-time temperature prediction results without complex knowledge of thermal characteristics,thermal generation or thermal boundary conditions of the battery.Through this technology,the minimum number of RTSs and the optimal layout position under a specified accuracy requirements can be calculated,which has certain guiding significance for the thermal management of lithium battery packs.