Machine Learning-Based Battery State-of-Health Prediction for Electric Vehicles

Abstract

Rapid electrification of transportation has increased the demand for accurate Battery State-of-Health (SOH) prediction; capacity loss, resistance growth, thermal stress and cycling uncertainty directly influence Electric Vehicle safety, range, maintenance cost and battery lifetime. Existing estimation methods often rely on isolated health indicators, restricted operating conditions, or models with limited temporal interpretability, reducing their reliability under nonlinear battery-aging conditions. A hybrid Long Short-Term Memory-attention architecture is developed to integrate sequential degradation learning with the adaptive identification of sensitive battery states. The analytical framework includes Coulomb counting, internal-resistance estimation, multidimensional feature construction, feature standardization, nonlinear state transition modeling, and a physics-guided objective function incorporating prediction-error, monotonicity, temporal-smoothness, and parameter-regularization terms. The MathWorks computational and dynamic-system simulation platform is used for model development and waveform assessments. When subjected to 200 charge-discharge cycles, the measured SOH decreases from 100% to 82%, and values predicted via the soft sensor closely follow the degraded trajectory. Such applications include battery management, predictive maintenance, fleet supervision, warranty assessment, residual-value estimation, intelligent energy scheduling and sustainable mobility.