A Homogeneous Meta-Learning LSTM-RNN Ensemble Method for Electric Vehicle Battery State of Charge Estimation

Several currently popular data-driven estimators such as Long Short-Term Memory-Recurrent Neural Networks (LSTM-RNN) rely on a single strong model which produces accurate predictions at the expense of lengthy training times. However, as lithium-ion batteries (LIBs) age, these models must be retrained with new data to learn the updated battery characteristics, either by transfer learning or by retraining from the ground up. This poses a problem for model retraining on low-power edge computing devices found in electric vehicle (EV) battery management systems (BMS), as model training is typically computationally intense and time consuming. This work proposes an ensemble meta-learning model consisting of several LSTM-RNNs that aims to significantly reduce model training time. To achieve this, multiple weak base models were trained using randomly sampled subsets of the training dataset, and the outputs were used as training inputs for a meta-learner. The proposed meta-learning method outperformed a simple average and an MAE-based weighted average ensemble fusion method – both of which were found to be not more accurate than the best-performing base model of the ensemble. Using similar training and testing datasets, the proposed model achieved a training time 2.6-3.5 times less than equally accurate (1.4% MAE) conventional shallow and deep LSTMs (DLSTM). Future research directions will be targeted at reducing the prediction times of the proposed method for increased practicality.