An Investigation of Battery Degradation Model under Different Cycling Conditions using Probabilistic NN-GA

Predicting battery State of Health (SoH) is crucial in assessing its performance degradation to ensure safe operation and minimize the risks of failure in electric vehicle application. Different predictive models have been established as a stand-alone or optimization algorithms for model-based approach. Existing works have shown that Neural Network (NN) coupled with optimization algorithms such as Extended Kalman Filter (EKF), Genetic Algorithm (GA), etc., as well as probabilistic NN have better prediction accuracy. This study aims at exploring the potential of GA as optimization algorithm while leveraging on the statistical memory of probability distribution of the solutions in predicting the battery degradation. Two types of NN architecture were constructed where the hyper-parameters of each were (1) based on the existing work and (2) optimized by GA, known as NN-GA. Both architectures were evaluated under varying operating temperatures and multiple discharging profile. It is found that NN-GA outperforms NN under all operating temperatures with Mean Absolute Percentage (MAPE), Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) of 0.68%, 0.0137 and 0.01 respectively. For cold (4°C) and hot (43°C) operating temperature, NN-GA predicts the degradation with MAPE of 0.98% and 1.24%, an improvement by 0.63% and 0.74% respectively.