Idea
An adaptive online forecasting model for battery degradation that improves accuracy with real-time data for energy and automotive industries
Research Paper
Core Innovation
This paper introduces iFSNet, a modified FSNet model that performs single-pass incremental multistep forecasting using pseudo targets generated by a linear regressor. Unlike prior offline models, it adapts continuously to new data distributions without retraining delays. The approach combines associative memory and adaptive structure to improve prediction accuracy on complex degradation patterns.
Market Size (TAM)
$20–50B TAM for battery management and prognostics; $2–10B SAM from electric vehicles and energy storage sectors. Driven by increasing EV adoption and renewable energy integration.
Potential Customers & Pain Points
- Battery manufacturers needing real-time degradation monitoring
- Electric vehicle companies requiring accurate battery life predictions
- Energy storage operators seeking to prevent failures
- Industrial equipment managers wanting adaptive maintenance scheduling
Business Model
Licensing the iFSNet model as an API or SDK to battery manufacturers and fleet operators; offering custom integration and support services.
Competitive Landscape
- NREL Battery Prognostics
- Tesla Battery Management
- CATL Battery Analytics
Implementation Challenges
- Integration with existing battery management systems
- Handling highly irregular degradation patterns
- Scaling to diverse battery chemistries and use cases
Validation Strategy
- Pilot deployment with battery manufacturers for real-time forecasting
- Benchmarking against existing offline models on diverse datasets
- Iterative improvement based on field feedback and new data
Research Paper Overview
Incremental Multistep Forecasting of Battery Degradation Using Pseudo Targets
Summary
This paper proposes iFSNet, an incremental Fast and Slow learning Network for online multistep battery degradation forecasting. It uses pseudo targets generated by a linear regressor to update the model sample-by-sample, enabling adaptation to changing data distributions without waiting for large retraining datasets. The model leverages associative memory and adaptive structure mechanisms to improve forecasting accuracy on both smooth and irregular degradation trajectories.