Physics-Informed Residual Graph Network for Temperature-Based Early Warning of Thermal Runaway in Lithium-Ion Batteries

Abstract Thermal runaway in lithium-ion batteries poses severe safety risks, and early detection is essential to prevent catastrophic failures. This study develops a physics-informed residual graph network (PINN–GNN) for temperature-based early warning using only historical temperature data. The model integrates a temporal graph convolutional encoder with a differentiable heat-balance integrator that reconstructs latent internal and external heat-generation trends from stable historical segments. By learning these hidden dynamics, the framework forecasts short-term temperature evolution and provides a quantifiable 30–60 s early-warning interval before the onset of rapid escalation. Physical consistency is enforced through non-negativity, smoothness, bias and slope supervision, and adaptive weighting for rising segments, while a residual multilayer perceptron captures unmodeled nonlinearities through bounded corrections. The method is evaluated on 52 Ah and 300 Ah LiFePO 4 cells subjected to electrical and thermal abuse, covering over 30 experiments per group with diverse thermal trajectories. Results demonstrate an average window-level prediction error below 2.5% and millisecond-level inference speed, enabling real-time monitoring. The proposed framework thus offers a physically interpretable, computationally lightweight, and data-efficient solution for thermal runaway early warning using temperature-only sensing, providing actionable lead time for intervention in practical battery systems.