Artificial Neural Network Methods for Lithium-Ion Battery Behavior Predictions

Abstract

The lithium-ion battery is a significant component in systems where electrification has been applied to achieve sustainability goals, such as electrical vehicles fulfilling emissions requirements. To control the lithium-ion battery to attain the safety, reliability, and performance demands of the electrified devices while maintaining design specifications, accurate predictions of the battery behaviors are essential. Data-driven methods with the recent technological breakthroughs in machine learning are regarded as solutions to characterize and simulate battery behaviors such as lithium-ion battery voltage and temperature. To predict the lithium-ion cell voltage at a low temperature and to estimate the temperature distribution of the lithium-ion battery pack, the Long short-term memory (LSTM) architectures are designed and trained with lithium-ion battery cell and pack test data. This data-driven method shows high accuracy, convergency, and robustness against external variables such as ambient temperature variation. However, the highly data-oriented nature of the LSTM method struggles with increased prediction error when local data deficiency occurs. To solve this drawback of the data-driven method, the physics-informed neural network (PINN) is developed to provide an additional source, physics law, for training the data-driven method. The PINN model to predict the battery cell temperature is developed by applying the battery thermal model to the loss function, implementing adaptive coefficients to the loss function, and modifying the neural network architecture with the pre-layer and connection layer, reflecting the analytical solution of the battery thermal model. The developed PINN model is superior in predicting the lithium-ion battery cell temperature with limited data size and an unidentified battery thermal model. This PINN model was applied to the aforementioned LSTM architecture to enhance the battery pack temperature estimation. The proposed LSTM-PINN hybrid model has more accurate predictions than the LSTM model proposed when data scarcity occurs. Furthermore, the LSTM-PINN hybrid model shows robustness against the incompleteness of the battery thermal model.