EnerNet: Attention-based dilated CNN-BILSTM for state of health prediction of CS2 prismatic cells in energy systems

The energy storage devices, such as various types batteries, are widely used in energy storage systems. The State of Health (SOH) is an essential parameter for the battery, which can estimate the battery's functionality and durability. However, the accuracy of the SOH prediction should be enhanced further to meet the requirements of practical application. This research presents a novel deep learning (DL) framework EnerNet that combining Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (BILSTM), and attention mechanisms to predict the SOH. The innovation of EnerNet lies in its unique architectural integration where, the dilated convolutions are employed in the architectural design to achieve the long-term dependencies and error trends in the battery data without compromising computational complexity. The CNN and dilated CNN incorporating sliding windows to extract data features from data variables (Voltage, current, temperature and capacity) of charging/discharging cycles of multiple batteries in CALCE and NASA datasets. The BILSTM networks are able to model temporal dependencies and the attention mechanism helps to concentrate on the most informative sequences for increasing the SOH precision, which is important for EnerNet performance. The performance of EnerNet was assessed using the leave-one-out cross-validation (LOOCV) that showed EnerNet outperformed other models across different datasets. For CALCE dataset EnerNet has MAE of 0.0081, RMSE of 0.0130 and R² of 0.9961 on Dataset I. On the NASA dataset, EnerNet yielded an MAE of 0.45, RMSE of 0.86, and an R² of 0.994 on the B0005 battery. Compared with other DL base models including CNN, LSTM, CNN-BILSTM, and BILSTM-GRU, these results presenting EnerNet as an innovative solution for battery health monitoring and validate the proposed model enhanced capability in accurately predicting the SOH.