(59z) Combined Use of Recursive Neural Network (RNN), Convolutional Neutral Network (CNN), and Attention Mechanism on Cycling Data of Lithium Ion Battery for Lifespan Prediction

Predicting the lifespan of lithium-ion batteries (LIBs) is essential for their safe and optimal use. The end-of-life (EOL) of a LIB is reached when the state of health (SOH) drops to 80% [1]. To achieve more accurate predictions with fewer data, i.e., data from a small number of initial charge/discharge cycles, the concept of knee-point, which represents the point where rapid degradation starts, has emerged. Data-driven approaches are prevalent, but previous methods have not explicitly considered the cyclic data patterns resulting from repeated charge/discharge operations.

In this work, we propose three methods to extract inter- and intra-cycle features from the initial cycle data and test them using a benchmark dataset [2]. The benchmark dataset contains the voltage, current, temperature (VIT) and SOH measurements resulting from different charging and universal discharging policies. As a rapid charging-discharging policy is utilized for all three batches of the dataset, knee points are seen in the data for all cells [3]. After identifying the knee-point for each dataset, we randomly shuffle the cell data to split them into the train, validation, and test sets since the life patterns of the three batches are similar due to the similar operating conditions. We utilized zero padding to realign all the profiles to a cycle length of 60 minutes. We also selected a time interval of 0.5 minutes for the data.

Then, we arrange the data into an array along the time and cycle dimensions to apply the convolutional neural networks (CNNs) and recurrent neural networks (RNNs). Since CNN primarily extracts geometrically close features (i.e., in the data matrix), we apply the CNN kernel along the cycle and time axes to capture key footprints from cycle-to-cycle behavior and the individual profile. We propose three combinations of CNN and RNN: 2D CNN alone, RNN + 1D CNN, and RNN + 2D CNN. For the RNN models, we apply three types of RNNs: the conventional RNN, LSTM, and GRU. The results show that the proposed methods outperform linear or nonlinear models that do not explicitly consider the inter-cycle behavior of the data profiles. The RNN + 2D CNN combination primarily gives the best results regardless of the RNN type with only 2 or 3 CNN layers.

In addition, attention mechanisms (AM) are incorporated into both the variable and time dimensions to enhance the performance of the prediction models and improve their interpretability. The proposed time-attention-based RNN + 1D CNN allows us to identify the significant time points in each cycling profile. Also, it helps eliminate the problem of vanishing gradients. The proposed variable-attention-based RNN + 1D CNN lets us gain insights into each degradation variable's importance changes as the cycle runs iterate. To sum up, the experimental results demonstrate that attention mechanisms improve the models' efficiency and interpretability.

References

[1] Hu X, Xu L, Lin X and Pecht M. Battery lifetime prognostics. Joule 2020;4:310-46.

[2] Severson KA, Attia PM, Jin N, Perkins N, Jiang B, Yang Z, et al. Data-driven prediction of battery cycle life before capacity degradation. Nat Energy 2019;4:383-91.

[3] Fermín-Cueto P, McTurk E, Allerhand M, Medina-Lopez E, Anjos MF, Sylvester J, et al. Identification and machine learning prediction of knee-point and knee-onset in capacity degradation curves of lithium-ion cells. Energy and AI 2020;1:100006.