Research on Physically-Constrained Continuous-Time Dynamical Models and Precise Smartphone Battery Life Prediction

Abstract

Addressing the challenge of balancing prediction accuracy and interpretability in smartphone battery life forecasting, this study constructs a physically-driven continuous-time dynamical model and a dynamically calibrated Time-to-End-of-Life (TTE) prediction system. The study first decomposes battery output power into external loads and internal losses based on the law of energy conservation, coupling it with a first-order lumped-parameter thermal model. To address user behavior randomness, a continuous-time Markov chain simulation mode switching is introduced, utilizing an Unscented Kalman Filter (UKF) for online parameter calibration. Empirical results demonstrate that this model accurately describes the evolution trajectory of the state of charge (SOC) under various load and temperature conditions. Specifically, the framework achieves precise runtime estimation across diverse scenarios—predicting TTE from 10.69 hours in compute-heavy modes to 23.46 hours in light-load modes—and identifies CPU operations as the primary energy consumption driver, while revealing GPS impact is negligible. This study provides a systematic quantitative solution for energy efficiency management and high-precision runtime prediction in mobile devices.