SOH Prediction Of Power Lithium-ion Battery Based On Gradient Boosting Decision Tree

Lithium-ion batteries are widely employed in electric vehicles and energy storage fields due to their superior performance.With the continuous use and aging,the state of health(SOH)and performance of battery will deteriorate,which endangers driving safety in serious cases.Therefore,the accurate SOH prediction is of great importance to the battery maintenance and safe driving of electric vehicles.The machine learningbased prediction method has attracted wide attention because it does not need to consider the complex reaction mechanism inside the battery.The Gradient Boosting Decision Tree(GBDT)algorithm has the advantages of strong generalization ability and high prediction accuracy.Therefore,this thesis employs the GBDT algorithm to achieve promising prediction for SOH.The main contents of which are as follows:This thesis firstly analyzes the change of the battery charging voltage curve under different cycles,and selects and constructs effective aging characteristics from the charging voltage curve in the case of partial charging as the input of the SOH prediction model considering the battery actual use.The experimental results demonstrate that the selected aging characteristics are reasonable and can be employed to predict SOH under actual battery usage.After the effective aging features are selected,they are employed as the input of GBDT algorithm and the SOH prediction model is established.To reduce the learning time of GBDT,solve problems such as insufficient sample learning,this thesis improves the GBDT from three aspects: the selection of the initial learning machine,the construction of slack variables to set the learning rate,and a Chebyshev-Pade approximation of the tanh(x)function,thereby an Improved Gradient Boosting Decision Tree(IGBDT)is proposed.Meanwhile,aiming at the complex and changeable characteristics of the actual driving conditions of electric vehicles,combining with the Hoeffding tree and other ideas to design incremental learning strategies,an Online Study-Improved Gradient Boosting Decision Tree(OS-IGBDT)algorithm is proposed,which makes the predictive model to have online learning capabilities to adapt to changes in working conditions.The experimental results manifest that the IGBDT can reduce quite a few learning times,and further improve the learning ability and prediction accuracy of GBDT.Meanwhile,OS-IGBDT can realize effective online incremental learning and obtain superior SOH online learning and prediction effect.Although the proposed OS-IGBDT algorithm is an effective online learning method,it still lacks a clear update mechanism and has problems of low learning efficiency.The idea of concept drift detection is introduced,and a Drift Detection based on Bernstein Inequality(BI-DD)is proposed.Finally,an online learning and model updating strategy based on BI-DD is proposed.The experimental results demonstrate that the BI-DD detection algorithm is effective and the online learning and model update strategies based on the concept drift detection results can make OS-IGBDT learn more efficiently.Finally,the proposed GBDT related algorithm is employed to the State of Charge(SOC)prediction.In view of the fact that the battery's actual use conditions and iteration errors can impress the joint estimation method,resulting in the inaccurate SOC and SOH prediction results,the proposed GBDT related algorithms-based SOH prediction method is used to update the SOC estimation model.The experimental results show that the SOH prediction method based on GBDT related algorithms can effectively update the SOC estimation model due to the accurate SOH prediction value,and achieve promising SOC prediction results under different cycles.