Effect of variable temperatures on machine learning battery SoH estimation for auxiliary aircraft batteries

In aircraft applications Nickel-based batteries are increasingly being replaced by Lithium-based technologies, particularly for their auxiliary unit, which requires a detailed study of their aging for management purposes. This research examines the influence of temperature variations on the identification and extraction of battery features, assessing their suitability as indicators for estimating the battery State-of-Health (SoH). The effect of temperature variations is also studied for machine learning models, both benchmark ones (kernel-based and neural network) and a proposed TCN - Neural ODE one. Optimized kernel-based models through Bayesian Search Cross-Validation and Grid Search Cross-Validation show improved results. Two neural network models, Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), are tested, but show a worse performance, probably due to the smaller dataset. The hybrid model presented, a TCN - Neural ODE improves upon the neural-network model, showing a better performance with smaller datasets and short and long-term dynamics. The effectiveness of the algorithms is determined using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and computing time. • Auxiliary batteries in aircraft applications are subjected to varying temperature profiles. • Time and differential features give rich information, even if they can be sparse. • Bayes optimization for kernel-based models required less training time. • Proposed TCN-Neural ODE outperformed neural networks. • This hybrid model is notably efficient under real-cycle and sparse data conditions.