vaborg/BatteryComp2006

BatterySwapAI 2026 — EML symbolic-regression RUL model

A submission for the BatterySwapAI 2026 challenge, built on the official example layout. The Remaining-Useful-Life model is a closed-form symbolic expression discovered with SymbolicRegression.jl over the EML primitive eml(x, y) = exp(x) - ln(y), ported 1:1 from the Julia pipeline and wired into the competition RULModel / Planner interfaces.

• batteryswap_example/train.py — the EML feature pipeline + fitted expression, EMLRULModel (implements RULModel), OrderedPlanner (implements Planner), and the pickle builder. All classes referenced by the pickle live here so that script.py's from batteryswap_example.train import * resolves them at load.
• batteryswap_example/planners/best.pickle — the trained planner (generate it with the command below, then commit it).
• script.py, Dockerfile, requirements.txt — unchanged from the example.

The RUL model is deterministic and has no learned parameters, so fit is a no-op and the pickle can be regenerated without the dataset. RUL is returned in days relative to the planning instant (the fitted expression emits hours; predict divides by 24).

Setup

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt -r requirements.dev.txt

Generate the submitted planner

Run from the repository root so the pickled classes get the canonical module path batteryswap_example.train.* (this is how script.py imports them):

python -c "from batteryswap_example.train import main; main()"

This writes batteryswap_example/planners/best.pickle. If BATTERYSWAP_DATASET_PATH points at a local dataset, it also prints per-scenario scores as a sanity check. Commit the pickle.

Test the submission in Docker (recommended)

Mirrors the competition runtime exactly.

docker build -t batteryswapai-2026-eml .
docker run --name batteryswapai -v ./dataset:/tmp/data batteryswapai-2026-eml \
  bash -c "/app/env/bin/python3 script.py && /app/env/bin/python3 -m batteryswap_public.metric"
docker cp batteryswapai:/app/submission.csv ./submission.csv

Submit

Commit and push to your HuggingFace model repository, then use New submission in the competition application.

Notes / next steps

• OrderedPlanner.safety_margin_days shifts predicted EOL earlier — a first, blunt handle on the downtime-dominated asymmetric cost (newsvendor δ*). Set to 0.0 to reproduce the example baseline.
• Part 2 (fleet scheduling: co-location batching, travel costs, worker-day limits) replaces OrderedPlanner. The runtime ships ortools and lifelines.
• To produce genuine quantiles (for a newsvendor-conservative plan), replace the point prediction in EMLRULModel.predict with your Monte-Carlo RUL distribution and select the target quantile.