Implementation of generative adversarial network-CLS combined with bidirectional long short-term memory for lithium-ion battery state prediction

Highlights

• Combination of GAN-CLS and bidirectional-long short term (LSTM).

• Battery data generation from images using GAN-CLS.

• Bidirectional LSTM model training for predicting the battery internal state.

• Provision of the reduction time, cost and a more accurate state estimation.

Abstract

This study newly introduces a complementary cooperative algorithm considering generative adversarial network (GAN)-Conditional Latent Space (CLS) combined with bidirectional long short-term memory (BLSTM) for improved and efficient lithium-ion rechargeable battery state prediction. The GAN-CLS algorithm, which is an advanced method of GAN, can generate corresponding images from an input label description. Long short-term memory (LSTM) is a specific recurrent neural network (RNN) architecture that can predict sequences more accurately than conventional RNNs. In terms of battery state prediction, the combination of two methods (GAN-CLS and LSTM) surely provides more improved and efficient rechargeable battery state prediction in contrast to conventional state predictors. The procedure of this study is as follows. First, we propose methods to enhance the data from battery charge/discharge by converting prepared data to images; then, the GAN-CLS method is used to generate corresponding battery data from previous images. Subsequently, the generated data is used to train the BLSTM model. Finally, the trained model is used to predict the battery state. By various experiments and verification, it is concluded that the proposed study can be a good solution for rechargeable battery state prediction (reduction of the time cost 50 times in modeling and 20 times in train/test, provision of a more accurate prediction mean square error (MSE) smaller than 0.0025 and the average MSE less than 0.0013).

Introduction

The state-of-charge (SOC) of the lithium-ion (Li-ion) battery is an important evaluation index of a battery management system (BMS). Therefore, performing an accurate measurement of the SOC is the main concern in analyzing battery-pack performance. In general, the battery pack is composed of hundreds and thousands of single cells for high voltage and energy storage [1]. The characteristics of batteries are different after undergoing several charge/discharge cycles. The BMS test needs to be executed many times for desired model accuracy, even though the characteristics of batteries are slightly different under the same state-of health (SOH).

Coulomb counting is the basic method to measure SOC. In this method, the current is integrated over time, and the integral is divided by the capacity to obtain the SOC [2,3]. However, Coulomb counting methods highly depend on the initial SOC estimation and the bias in the current sensing. The conventional SOC estimation algorithm is represented by a Kalman filter [4,5]; however, the SOC-open circuit voltage (OCV) characteristics of a battery have strong nonlinearity at the beginning and end of the charge/discharge state [6], [7], [8]. The extended Kalman filter (EKF) [9], [10], [11], [12] is proposed to deal with the problem of predicting the SOC under the linear assumption of the Kalman filter. However, the performance of the Kalman filter and its derivative version depends on an accurate battery model, such as the conventional Thevenin model or second Thevenin model [13]. Both models require offline parameter identification, as well as measurement of the OCV, and a determination of the time elapsed of the model. Jian Chen et al. [14,15] proposed the use of the adaptive neuro-fuzzy and radial basis function neural network as a nonlinear observer. In the experiment of [14], the battery is discharged with a constant current in steps of 5% of capacity and rested; then, the corresponding resistance R can be calculated as the ratio of the instantaneous voltage drop U and current I. Then, the parameters of the RC networks at the corresponding measured SOCs can be determined. Finally, the neural network is used to fit the curve, and the Thevenin model is used to estimate the SOC. This method can yield an accurate estimation if the battery has the same SOH. Han et al. [16] used the genetic algorithm (GA) to identify the battery capacity and analyze the battery age model under different test temperatures. In [17], a radial basis function (RBF) neural network (NN) was used to learn the bounds of the uncertain dynamics of the battery equivalent circuit model. Then, the outputs of the RBF NN were used to estimate the SOC based on a robust sliding-mode observer (RSMO). However, this requires tremendous testing under different temperatures to identify the model parameters. Other SOC estimation methods are established on machine learning strategies, which encompass artificial neural networks (ANNs), fuzzy NNs, back propagation neural networks (BPNNs), and adaptive fuzzy NNs [18], [19], [20], [21], [22]. These data-oriented approaches treat the battery model as a black box, with requiring knowledge of the battery internal structure. The battery model is treated in the weights of NNs. A large amount of experimental data is required to train the NNs for high accuracy. In addition, it is difficult to balance the model precision and generalization performance. Furthermore, it necessitates tremendous computing workload and resource consumption. A summary of the lithium-ion battery state prediction methods is listed in Table 1.

In this study, a method involving the use of GAN-CLS (called as matching-aware discriminator) to generate the required data, rather than running experiments, is adopted [23]. In this method, measured data with different SOH are padded with zeros at the end, such that all data are of the same length. Then, the vectors are reshaped as 2-dimensional images, and convolutional neural network (CNN) is used to extract the rough and detailed features as needed. For the original GAN, the random data with fixed distribution were used to generate the desired distribution. However, battery SOC-OCV characteristics vary with different SOH, which means that the SOC-OCV characteristics exhibit a different distribution. In the GAN-CLS model, the discriminator in GAN receives the generated data and corresponding SOH, such that the input data includes the corresponding SOH information. The performance of the generated data was tested based on a significance test, using the best data to represent the experimental data. The generated data are used to train the bidirectional-LSTM (BLSTM) networks that operate on the input sequence in both directions to facilitate optimal decision-making for the current input [24,25]. As a result, the GAN-CLS generates high-quality pseudo-experimental data, which can be used for BMS or fault-detection simulation. The proposed methods reduce the time cost 50 times in modeling and 20 times in train/test. The results show a predicted mean square error (MSE) smaller than 0.002. The proposed methods significantly reduce the time and resource cost.