| Lithium-ion batteries are widely used due to their high energy density and long life.Battery Management System(BMS)plays an important role in maintaining the efficient and reliable operation of the battery.State of Health(SOH)prediction is the core function of the BMS.Only by accurately predicting the battery state can the battery be managed rationally to improve its utilization and range and extend its service life.The thesis investigates SOH prediction methods for lithium-ion batteries using the NASA and University of Maryland Center for Advanced Life Cycle Engineering(CALCE)lithium-ion battery datasets in three main areas of work:(1)For the two datasets studied,12 health factors characterizing the degradation of battery performance are extracted from the charging curves.The current and voltage of the charging process are analyzed in depth,and the health factors are mined from both measured and calculated variables,and outliers and noise are processed to provide a basis for subsequent battery SOH prediction.(2)In order to improve the accuracy of SOH prediction for Li-ion batteries,an complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)is constructed to decompose the data and reduce the dimensionality with kernel principal component analysis(KPCA).Then it is input into a combined Gaussian process regression(GPR)and long-short term memory network(LSTM)model.The effect of denoising,reducing data redundancy and improving prediction accuracy is achieved.The superiority of the proposed model was tested by comparing it with other models.(3)In order to improve the training efficiency of the model,a data processing method of random forest for feature selection and partial least squares for feature dimensionality reduction is proposed.Then,in order to solve the self-discharge during the degradation of lithium-ion batteries,which leads to the uncertainty of SOH prediction,a model based on the combination of LSTM and quantile regression is proposed to obtain the interval prediction of SOH and the probability density prediction using kernel density estimation.The results are compared by different training strategies and show that the proposed model has higher coverage and lower mean width with more concentrated probability density distribution at the same confidence level,which provides better prediction capability for practical applications. |
