Predicción de la trayectoria de degradación de la capacidad de la batería para dispositivos médicos portátiles con...

Accurate prediction of battery capacity degradation is critical for the continuous operation of wearable medical devices and patient safety. However, batteries in these devices often show complex, nonlinear degradation patterns caused by irregular charging cycles and intermittent high loads. To address this, we propose a novel framework that integrates deep temporal decomposition with autoregressive fine-tuning. Specifically, a multi-scale feature extraction network is designed to capture both local fluctuations and global trends. A dedicated temporal decomposition module then separates long-term trends from short-term noise, effectively mitigating noise arising from wearable battery operating conditions. Additionally, an autoregressive residual correction mechanism is introduced to correct iterative prediction errors, thereby reducing cumulative bias in long-term forecasts. Extensive experiments on three diverse datasets demonstrate the improved performance of the proposed method, maintaining the Mean Squared Error within 20% across the evaluated datasets and experimental settings.