Energy-balance module improves transformer frost forecasts without leaf sensors
Original title: A plug-in energy-balance freezing module for transformer frost nowcasting without leaf-sensor inputs
Without leaf sensors, PET-E reached a lead-mean CSI of 0.516 versus 0.478 for LightGBM across 12 hourly leads.
Researchers add a transformer-compatible module that embeds leaf energy-balance physics in freezing-probability forecasts through a temperature reference, learned residual, and a multiplicative probability cap. They prove conditional monotonicity and an alarm boundary, then compare output-pathway and attention placements. On 19 Korean stations over five frost seasons, the module cut Brier scores and warm-surface alarm fractions for a time-token transformer and iTransformer without a significant CSI drop at any lead. Held-out stations without static or leaf-sensor inputs also beat LightGBM at all 12 leads.
- Evaluation covered 12 hourly leads at 19 Korean stations.
- In 2025-2026 without leaf sensors, warm alarms at 2°C or above fell from 0.087 to 0.051.
- Held-out CSI was 0.468 and 0.467 versus 0.437 for LightGBM.
- The temperature pathway cut the alarm fraction in all four time-token settings.
- Products
- PET-EiTransformerLightGBM
- Countries
- South Korea
From the original article
1 min read at arxiv.orgFrost protection requires local warnings where leaf-sensor inputs are unavailable. In the frost literature we surveyed, physics entered machine-learning models through inputs, numerical-model coupling, or losses, rather than model structure.
By Juyeon Kim · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.