Capturing Temporal and Spatial Patterns in Battery SOH Prediction with a Dynamic T-GNN Model

Predicting the state-of-health (SOH) of Lithium-ion batteries is crucial for Prognostic and Health Management systems. SOH prediction has been extensively researched, with several methods, including gated recurrent units and Long Short-Term Memory (LSTM) networks, being widely used. However, these methods often face challenges with explainability and weak generalization, especially for long-term forecasting near the end of life, where degradation patterns become more complex to predict. To address those challenges, we propose the Temporal Graph Neural Network (T-GNN) model that integrates an LSTM with a dynamic GNN architecture, combining their strengths to enhance prediction accuracy and model interpretability for battery SOH estimation. The T-GNN is designed to capture degradation patterns over time while handling both temporal and spatial dynamics effectively. By incorporating a dynamic structure, the TGNN model continuously responds to changes in battery cell operation conditions and learns adaptive deep relationships between battery cells for enhanced explainability and generalization for long-term forecasts. The proposed model's performance is evaluated on a publicly available dataset and compared to three benchmark models. The results demonstrate high predictive performance, as evidenced by low root-meansquared error values, highlighting its potential to improve real-time battery management and support more reliable long-term forecasting in practical applications.