Research On SOC Prediction Of New Energy Vehicle Power Battery Based On Stacking Algorithm

At Action Plan for Carbon Dioxide Peaking Before 2030 promulgated by the State Council in 2021 clearly proposes to vigorously develop new energy vehicles and constantly promote the intellectualization and networking of new energy vehicles.The development of new energy vehicles plays a key role in coping with climate change and promoting green development,and the breakthrough and innovation of power battery management system（BMS）technology,which is the core of automotive energy management,is an important research field at present,of which the State Of Charge（SOC）of the battery is one of the important parameters of BMS,and it is also the basis for the charge and discharge control strategy and battery balancing work of the entire vehicle.But due to the complexity of the battery structure,SOC cannot be measured directly.It is also difficult due to the influence of temperature,battery aging and many other factors.Therefore,how to predict SOC conveniently,quickly and accurately is of great significance to improve battery utilization,prolong battery service life and continuously optimize battery technology.Based on the actual operation data of new energy vehicles,this thesis constructs the prediction model of power battery SOC through data mining technology in order to achieve more accurate prediction of SOC.The main research contents of this thesis are as follows:First,based on the statistical basic analysis and processing of the actual operation data of new energy vehicles,this thesis carries out the feature engineering in combination with the relevant knowledge of power battery and then selects eight features to participate in the model construction.Considering the non-linear characteristics of power battery and the interpretability of prediction results,this thesis selects the mainstream random forest RF,xgboost and lightgbm model in the Ensemble Learning Algorithm to build the prediction models of power battery SOC respectively.Finally,it is found by comparative analysis that xgboost model has the best ability to predict SOC through several model evaluation indicators.Secondly,in order to improve the predictive performance of the model,the respective advantages of RF,XGBoost and LightGBM are combined,and the three single models are fused with model fusion.Specifically,in view of the shortcomings of stacking fusion algorithms,two different improvements have been made using the idea of linear weighting:One side in order to reflect the performance of K test sets,this thesis uses Sum of Squares of Error（SSE）and Weighted Average in place of Equal-weighted Average to form the test set of the primary learner.On the other side The input of primary learner is the combination of the original feature set with the output of the primary learner.The combination aims to change the input characteristics of the second layer of the Stacking Model and to increase the complexity of the secondary learner so as to improve the prediction performance of the model.Through experimental comparative analysis,it is demonstrated that the improved Stacking Model has smaller MAE and RMSE and higher R~2 than XGBoost and traditional fusion model on the performance of power battery data set.It can effectively and accurately predict the power battery SOC.