Yujie Zhang

Smartphone time-to-empty (TTE) prediction is hindered by the interplay of cell electrochemistry, heterogeneous peripheral power draw, and variable user workloads. Existing methods address these factors in isolation: statistical approaches fail under non-stationary loads, data-driven methods suffer distributional shift, and physics-based models neglect the screen, CPU, and radios that dominate runtime discharge. To bridge these gaps, we propose a hybrid framework that couples a second-order Thevenin equivalent circuit model with component-level power decomposition and temperature/aging modulation. Backtesting on three devices and nine scenarios shows that the proposed model achieves the lowest TTE MAE in seven of nine cases (4.0%–43.1% reduction over the best baseline) and ranks first or second across all scenarios. Subsequent ablation reveals workload-dependent subsystem contributions: temperature correction aids mixed-use but introduces noise under stable thermal conditions, while ECM dynamics benefit pulsed loads but add variance under near-constant-current discharge. Further sensitivity analysis identifies nominal capacity, screen and CPU coefficients, and baseline draw as dominant parameters, with GPS and audio coefficients being negligible.