Improving Extrapolation Accuracy of Voltage Prediction Model for Lithium-Ion Battery Using Machine Learning

Abstract

Listen

Translate article

As battery electric vehicles (BEVs) become more widespread, collecting large amounts of time-series data on lithium-ion batteries in real-world settings has become possible. In recent years, there has been growing interest in methods that utilize collected battery data and machine learning (ML) models to predict battery performance accurately [1-4].However, there is a challenge in that prediction accuracy declines when the distribution of training data is uneven, and it has been a barrier to practical application. This challenge is due to real-world data often being unevenly distributed, and ML models generally have low extrapolation accuracy [5-6]. A specific challenge for BEVs is that the data in the low state of charge (SOC) region is insufficient because batteries are infrequently discharged to lower SOC due to users' anxieties about running out of electricity, resulting in less prediction accuracy.Here, we studied several representative ML models and their characteristics to achieve high extrapolation accuracy of voltage prediction. We artificially generated data for training and testing using an electrochemical simulation model. To evaluate the extrapolation accuracy of each model, we trained the ML models using data that excluded the low SOC region. We quantified voltage prediction accuracy for the entire SOC region using root mean square error (RMSE). Fig. 1 shows results for four ML models (“Random Forest,” “Gaussian Process,” “LSTM," and “MLP-ReLU"). The results confirmed that accuracy highly depends on the model and that MLP-ReLU (multilayer perceptron with ReLU activation function) has the highest accuracy. For further study, we set up several SOC ranges and comprehensively evaluated the RMSEs of eight representative ML models. Our results confirmed that the MLP-ReLU showed the best accuracy [7]. Furthermore, we examined four representative MLP activation functions (“ReLU,” “identity,” “tanh,” and “logistic”) and confirmed that the ReLU performed the best.This study shows that MLP-ReLU is suitable for building a model with high extrapolation accuracy even when training data is unevenly distributed. From these results, we concluded that we are close to realizing practical applications for predicting battery performance with high accuracy by utilizing real-world battery data and ML models.