Robust state-of-charge estimator for lithium-ion batteries enabled by a physics-driven dual-stage attention mechanism

The current deep state-of-charge (SOC) estimators face a challenge in extracting short-time-domain multi-scale features, which makes it difficult to capture important driving variables in multi-channel battery measurement features and invariant features in the time and frequency domains of the historical hidden states. This limitation leads to the tradeoff between the ability of the estimators to be trained at a lower cost and the higher error peaks and cumulative errors in the SOC estimation. A hybrid approach, which combines the physics-based characteristics of SOC changes with real-time measured data, is likely an effective solution to address this issue. In this work, we propose a novel SOC estimator, which is based on physics-driven dual-stage attention-based bidirectional recurrent neural network (PDA-BRNN). This estimator consists of two stages. In the first stage, the physics-driven input attention (PIA) mechanism synthesizes historical information and radial strains of the electrode particles of a battery, which are calculated via a physics-based analytical model, from numerous input channels. This synthesis helps identify the main driving variables and suppress local error peaks and cumulative errors in the estimation process. In the second stage, dual-temporal attention (DTA) mines, amplifies, and correlates effective time-varying information from different dimensions in the historical hidden states that benefits the current estimation process for reducing the error levels and improving the overall estimation accuracy. Experimental results demonstrate that the proposed PDA-BRNN is more accurate and robust than the state-of-the-art deep estimators under coherent multi-scale disturbances or Gaussian noise in measurements.