Abstract: Considering the problems of insufficient online prediction capability of the residual life of Li-ion batteries and the poor learning capability of the extreme learning machine based network model for small sample training data, a method of constructing continuous health factors is proposed to perform multi-step prediction of the residual life of Li-ion batteries using generalized regression neural network fused with an improved ELM. Firstly, the iso-voltage drop discharge time of Li-ion batteries is extracted as the health factor, and the discrete battery operating parameters are reconstructed into continuous health factors using the interpolation complementary method. Then, the generalized regression neural network is used to predict the early remaining life of Li-ion batteries, and the gravitational search algorithm is introduced to optimize the limit learning machine to build a mid-late RUL multi-step Li-ion battery multi-step with real-time update capability Finally, the model is validated based on the data of CS2-35～CS2-38 and CS2-7 batteries from the Maryland advanced life cycle engineering center. The experimental results show that the online RUL multi-step prediction model for Li-ion batteries based on the fusion of GRNN and GSA-ELM has high prediction accuracy and better robustness and applicability compared with other models.

Key words: equal voltage discharging time, general regression neural network, gravity search algorithm, extreme learning machine, multi-step online prediction