Capacity prediction of lithium-ion battery using an attention-based hybrid deep learning model

Accurate State of Health (SoH) prediction is a crucial indicator to the safe and reliable management of lithium-ion battery. To diagnose the health, data-driven schemes are used in both standalone and hybrid structures. However, traditional methods often face challenges such as vanishing gradients, raises issue to learn to preserve data over multiple timesteps and difficult to capture in long-term dependencies in time-series data. Such limitations are addressed with Nonlinear Autoregressive Model with External Input (NARX) Attention-Based Gated Recurrent Unit (NARXGRU-A) and NARX recurrent neural network to predict the capacity of the battery. These models are assessed using open-source datasets to estimate SoH for multi-cycle ahead capacity prediction based on average charge–discharge cycle data. Among the evaluated techniques, NARXGRU-A demonstrates superior performance and is employed for real-time data. However, real data acquisition is time-consuming. To address this challenge, synthetic data is generated by generative adversarial networks from limited real data. Findings show the proposed NARXGRU-A model exhibits better performance over other techniques to multiple data sets with error metrics of less than 1%. These findings highlight strong prognostic capabilities of the model for multi-step, long-horizon SoH prediction to the battery capacity.