A Multiscale Entropy-Based Long Short Term Memory Model for Lithium-Ion Battery Prognostics

Data-driven methods for battery prognostics aim to capture information from past data to predict the future battery performance. Large amount of data required in these methods can be further complicated by noises. Coarse-graining technique may reduce the noises in the battery data. Entropy, from information viewpoint, can be used as an input variable for prognostics. Long Short Term Memory (LSTM) algorithm is a good candidate to fit the battery data since it is compatible with numerical dependent variables and yields great generalization ability. In this paper, we utilize the multiscale Shannon entropy of lithium-ion battery coarse-grained voltage sequences and the actual past capacities as input variables in a LSTM model for predicting the next-cycle capacities. As compared to the traditional data-driven approaches to battery prognostics, the proposed model results in high accuracy since it takes into account the multiple time scales inherent in voltage sequences and stores key information from capacities over charge/discharge cycles.

Automated summary

The study utilizes multiscale Shannon entropy of lithium-ion battery coarse-grained voltage sequences and past capacities as input variables in a LSTM model to predict next-cycle capacities. Compared to traditional methods, the proposed model achieves higher accuracy by considering multiple time scales inherent in the voltage sequences.