Versatile and high-accuracy seq2seq model with self-attention mechanism for Li-ion battery characterization

In this paper, we innovatively propose a seq2seq model that incorporates a self-attention mechanism for modeling the external characteristics of Li-ion batteries. During training, an input sequence of shape (4, time_steps) is formed by concatenating the battery voltage at each time step with the external stimuli such as current, temperature, and cumulative capacity at the next time step. The voltage variation at next step will be used as the target sequence of shape (1, time_steps). During inference, the input sequence is iteratively updated by adding the voltage variation obtained at each step to the previous voltage value, allowing us to predict the voltage variation over the entire time series. The seq2seq model employs two heads that utilize different sequence masks, ensuring attention to both historical data and current stimuli. This approach achieves top-tier prediction accuracy on an open dataset. Furthermore, our model demonstrates outstanding versatility and transferability, as it is not constrained by battery type, operating conditions, temperature, or data upload frequency. Lastly, the characteristics of aged batteries can be represented through fine-tuning.