Battery performance prediction techniques based on machine learning (ML) models and lithium-ion battery (LIB) data collected in the real world have received much attention recently. However, poor extrapolation accuracy is a major challenge for ML models using real-world data, as the data frequency distribution can be uneven. Here, we have investigated the extrapolation accuracy of the ML models by using artificial data generated with an electrochemical simulation model. Specifically, we set a lower open circuit voltage (OCV) limit for the training data and generated data limited to the higher state of charge (SOC) region to train the voltage prediction model. We have validated the root mean squared error (RMSE) of the voltage for the test data at several lower OCV limit settings and defined the average + 3 standard deviations of them as an evaluation metric. Eight representative ML models were evaluated, and it was found that the multilayer perceptron (MLP) showed an accuracy of 92.7 mV, which was the best extrapolation accuracy. We also evaluated models with published experimental data and found that the MLP had an accuracy of 102.4 mV, reconfirming that it had the best extrapolation accuracy. We also found that MLP was robust to changes in the data of interest since the accuracy degradation when changing from simulation to experimental data was as small as a factor of 1.1. This result shows that MLP can achieve higher voltage prediction accuracy even when collecting data for comprehensive SOC conditions is difficult.
Loading....