VoltWatch

Inspiration

EVs generate massive battery telemetry, yet current monitoring systems remain unstable, mode-sensitive, and often ignore uncertainty. Estimates can jump between driving and resting states and sometimes contradict battery physics. As EV adoption accelerates, this instability becomes a financial and operational risk for OEMs and fleets. We were inspired to build a system that makes battery intelligence stable, interpretable, and decision-ready.

What it does

VoltWatch is a physics-informed, mode-aware battery health calibration and forecasting framework. It: Separately models driving and resting battery behavior Tracks capacity fade and internal resistance growth Estimates state of health (SOH) in real time Forecasts remaining useful life (RUL) and expected end-of-life (EOL) Provides Bayesian uncertainty bounds instead of single-point predictions The result is stable hourly monitoring and actionable daily lifespan forecasting.

How we built it

We designed a multi-stage neural architecture: Mode classifier separates drive vs rest telemetry. Degradation estimation network models latent battery states (capacity ↓, resistance ↑). Physics-informed calibration layer constrains unrealistic degradation trends. Forecasting network predicts RUL and EOL using time-aware inputs. Bayesian estimation layer outputs uncertainty-aware predictions. The system is structured for fleet-scale deployment via API and dashboard integration.

Challenges we ran into

Stabilizing predictions across drive ↔ rest transitions Preventing physics-inconsistent trends (e.g., resistance decreasing unrealistically) Designing uncertainty estimation without overcomplicating the architecture Balancing technical depth with computational efficiency for real-time monitoring Ensuring interpretability while maintaining predictive power was a core challenge.

Accomplishments that we're proud of

Building a fully structured, completely novel framework backed by reserarch, multi-stage architecture instead of a single black-box model Integrating physics constraints to enforce realistic degradation behavior Implementing uncertainty-aware forecasting Designing the system with enterprise deployment in mind Creating a closed-loop architecture that supports retraining and scaling

What we learned

Battery forecasting is not just a prediction problem — it’s a stability and trust problem. Mode-awareness dramatically improves signal consistency. Industry-grade ML systems require interpretability and uncertainty, not just accuracy. Bridging physics and machine learning leads to more reliable real-world systems.

What's next for VoltWatch

Integrate real-world BMS fleet telemetry Validate against maintenance and warranty records Deploy a controlled pilot across 50–500 vehicles Benchmark forecast stability and error metrics (MAE, variance reduction) Develop enterprise API endpoints and automated risk alerts