Remaining useful life prediction of Lithium-ion batteries based on data preprocessing and CNN-LSSVR algorithm

• The CNN-LSSVR algorithm is proposed for limited battery history data in RUL prediction. • Capacity and internal resistance are jointly selected as a health factor to evaluate battery RUL. • The MRSVD method is for solving the problem of data noise due to unknown causes. • The validity and accuracy of RUL prediction methods are analyzed in terms of publicly available data and multi-type algorithms. Lithium-ion batteries are now widely available in power and energy systems. Targeting the thorny issues of limited battery historical cycle data and the impact of uncertainty in the data collection process in practical applications, this study proposes a Remaining useful life (RUL) prediction method for lithium-ion batteries based on the data preprocessing and the joint convolutional neural network (CNN)-least squares support vector regression (LSSVR) algorithm. Based on the performance degradation characteristics of the battery, the method proposes new RUL assessment indexes and corresponding health factors. The innovative Multi-Resolution Singular Value Decomposition (MRSVD) method is implemented to reduce the interference caused by noise and error. Eventually, the CNN-LSSVR algorithm and mutant particle swarm optimisation algorithm are utilised to solve the mapping regression and hyper-parameter optimisation problems, respectively, to achieve a complete prediction of RUL. In this work, the feasibility of the method is verified using publicly available datasets and compared with other common noise reduction and prediction algorithms after noise reduction and prediction experiments. The results show that the available capacity and internal resistance of the battery as health factors can effectively achieve degradation performance prediction. Compared with other traditional algorithms, the proposed RUL prediction method can reduce the mean absolute error and root mean square error by at least 37% and 61%, respectively, and has better stability. The RUL prediction method provided pave the new way for accurate prediction of battery data with limited number of samples and high noise characteristics. The fast and accurate battery RUL prediction method proposed in this work is highly beneficial for enhancing the stable and economic operation of power and energy systems.