Prediction of Battery SOH and RUL Based on Cooperative Characteristics in Voltage-Temperature-Time Dimensions

Abstract

In the prognostics health management (PHM) of marine power lithium batteries, the estimation of the state of health (SOH) and the prediction of remaining useful life (RUL) are of great importance to ensure the battery operational safety and efficiency. In this study, an improved multivariate dimensionality-reduction for Bayesian optimized bi-directional long short-term memory (IMD-BiLSTM) algorithm is proposed and applied to realize SOH estimation and RUL prediction of lithium battery. Specifically, based on the discharging data of lithium battery under specific operating conditions, several health indicators are proposed for the work. On this basis, a collaborative dimensionality reduction algorithm based on Pearson correlation and principal component analysis is proposed to further retain feature information and reduce input dimensionality. Then, the prediction model based on BiLSTM is established, in which the hyperparameters are optimized by Bayesian algorithm. Finally, the effectiveness of the proposed IMD-BiLSTM method is verified by experiments based on the NASA PCoE dataset, and the prediction accuracies of SOH and RUL are emphatically analyzed. Numerical simulation results show that the proposed IMD-BiLSTM-method can effectively extract battery health characteristics and achieve dimensionality reduction. In addition, the proposed IMD-BiLSTM-method significantly outperforms the compared state-of-the-art algorithms in SOH/RUL prediction accuracy and robustness.

Bidirectional Long Short-Term Memory Model of SoH Prediction for Gelled-Electrolyte Batteries under Charging Conditions

“…This paper utilized Matlab to construct the deep-learning architecture of the predictive model. Based on previous experiments [39,[42][43][44], the Adam optimizer was used to train our model with the batch size and epochs set to 32 and 100, respectively. To avoid gradient explosion, the maximum gradient was set to 2; the dropout rate of each LSTM function was set to 0.3 to avoid overfitting.…”