A hybrid approach based on decomposition algorithm and neural network for remaining useful life prediction of lithium-ion battery

Abstract

Aiming at the problems of non-linearity, non-stationary and low prediction accuracy of the original capacity degradation data for lithium-ion battery, a novel remaining useful life prediction approach is proposed. First, complete ensemble empirical mode decomposition adaptive noise is employed to achieve complete adaptive decomposition of the original data to prevent effective information about the capacity regeneration part from being eliminated. Next, fused high and low frequency parts are obtained through zero-crossing rate and new fusion rules, which can reduce the number of input network components and lighten operating costs. Then, the low frequency part is predicted using deep neural network; the high frequency part is predicted by self-designed improved Res2Net-Bidirectional Gated Recurrent Unit-Fully Connected (IRes2Net-BiGRU-FC). Finally, optimal results selected according to the criterion of minimum absolute error contains respective advantages of two high frequency fusion rules. Two sets of data from NASA under different charging and discharge conditions are used for simulation experiments and compared with other methods. The results show that our approach is feasible regardless of whether it is based on the battery data obtained in the constant voltage and current mode or the current random mode.