AI-MLPythonTensorFlowKerasLSTMpandas
Battery Degradation Forecast (LSTM)
Single-layer LSTM that forecasts smartphone battery percentage from telemetry, then rolls its own predictions forward for a 24-step outlook.
A time-series deep-learning study forecasting smartphone battery degradation from real device telemetry. A single-layer LSTM learns one-step-ahead battery percentage from four channels (battery %, CPU usage, running apps, and voltage), then feeds its predictions back in to generate a recursive 24-step forecast.
Data & model
- 385,429 sensor readings from a single Samsung SM-A910F (Android 8.0.0, 5,000 mAh Li-ion), 15 raw channels reduced to 4 model features.
- Chronological 20% test split, no shuffling, so the model is judged on genuinely future timesteps.
- Held-out test MSE below 10⁻⁴; final validation loss 5.55×10⁻⁷ after 30 epochs.
Notes
- Every figure on the results dashboard comes from the reference notebook run in the repo.
- Strong short-horizon accuracy, with the honest caveat that a single well-behaved device limits real-world variability.