Hybrid data-based modeling for the prediction and diagnostics of Li-ion battery thermal behaviors

Lithium-ion battery (LIB) has been deployed for the electrification of the transport sector as a key strategy for climate change mitigation and adaptation. However, it has significant technical challenges such as thermal runaway, requiring a good understanding and accurate prediction of the LIB thermal behavior (heat generation rate). In this study, a novel hybrid approach using an Artificial Neural Network (ANN) is developed for predicting the heat generation rate with discharge current, output voltage, ambient temperature, cell surface temperature, and Depth of Discharge (DOD) as the feature vectors (inputs); where the DOD is estimated with an Extended Kalman Filter (EKF), and direct Coulomb Counting (CC) method, respectively. A shallow neural network utilizing the Marquette-Levenberg algorithm is designed and calibrated using over 8000 cases of the testing data. It is shown that the predicted heat generation rate of LIB agrees well with the experimental results with an accuracy of R > 0.995. Further potential of this hybrid data-based model is evaluated by simulating a thermal management system control and by introducing voltage and current sensor faults for diagnostic purposes. It is shown that, when compared to the experimental value, the relative error of the total heat output generated is less than 2% when there is no sensor fault, and greater than 50% and 25%, respectively, with an induced failure of the current and voltage sensor, demonstrating the ability to build accurate models relying solely on LIB discharge data for sensor diagnostics. This study highlights the combination of using battery thermal behavior with machine learning for real time battery system monitoring, controls and field diagnostics.