Limmen-batteryswapai2026

BatterySwapAI 2026 — Team limmen

Censoring-aware remaining-useful-life prognosis for CR2477T IoT sensor batteries, coupled to a decision-theoretic replacement scheduler under worker-hour and travel constraints.

Everything needed to reproduce the stored model and the submitted plans is in this repository. All code here is authored by the participant and MIT licensed (see LICENSE).

---

1. Method

Four stages, each addressing a specific structure in the problem:

1. Censoring-aware survival model. 82.2% of the 461 training devices never fail inside the observation window. A Weibull AFT (lifelines) is fitted on all of them, censored rows included, rather than on the 82 observed failures alone. Supplies the beyond-window survivor tail.
2. Physics- and DSP-informed features. Distance to the 2.4 V EOL knee, temperature-compensated voltage (measured within-device coefficient 0.00483 V/°C), window minima, local volatility and trend, computed on the in-band smoothed series that mirrors batteryswap_public.utils.smooth_series.
3. Calibrated discrete failure CDF. Six monotonic gradient-boosted classifiers at 7/14/21/28/35/42 days, each isotonic-calibrated on out-of-fold predictions grouped by building, since the scored splits use buildings never seen in training. Monotonicity constraints encode physics sign-priors (lower voltage cannot reduce failure probability).
4. Expected-cost decision rule. Per battery, the realised evaluator cost of deferring (failure probability × late penalty × expected late days, plus a priced emergency round trip) against swapping (survival probability × early penalty × expected early days, plus wrench time). Buildings are visited only when the summed savings of their due batteries exceed the shared trip cost. Swap timing is the newsvendor critical fractile q* = c_early/(c_early + c_late) = 0.5/10.5 ≈ 0.048 — derived from the cost constants, not tuned.

2. Environment

python -m venv venv && source venv/Scripts/activate
pip install -r requirements.txt

Developed on Python 3.11.9. requirements.txt is the organizers' fixed competition set, copied verbatim; the submission adds no dependency beyond it.

3. Data

python download_data.py       # from the project repo; or use huggingface_hub directly

Fetches batteryswapaichallenge/BatterySwapAI-2026-Public into data/: devices.csv, eol_times.csv, battery_metrics.parquet, scenarios.json. No other dataset is used.

4. Preprocessing → training table

python scripts/build_dense.py --step=2 --start=2025-06-01 --jobs=18 \
    --out=artifacts/training_table_era_v6.parquet

One row per (snapshot, battery), features computed strictly from data at or before the snapshot and labels measured after it. Two deliberate choices:

• --step=2 — snapshots every two days rather than only the 48 shipped scenario starts, which the competition documentation explicitly invites.
• --start=2025-06-01 — restricted to the evaluation era. Training on the full 2022–2026 history scored 2124 against 2065: the older period has a 0.9% in-window failure base rate against 2.3% in the scored era, and isotonic calibration inherits the training base rate.

Runtime ≈12 min on 18 cores. Deterministic.

5. Training

AQUILA_MEMBERS=5 python scripts/train.py --table=artifacts/training_table_era_v6.parquet

Runtime ≈15 min. Deterministic given the table (SEED = 20260818; no RNG in featurisation, AFT fitting is deterministic).

6. Stored-model generation

python scripts/make_submission_repo.py --safety=1.0 --quantile=0.05

Bundles the fitted planner into aquila/planners/best.pickle — the artifact shipped in this repository — together with the aquila package, script.py and the Dockerfile.

7. Inference

Exactly as the competition harness runs it:

BATTERYSWAP_DATASET_PATH=../data BATTERYSWAP_SPLITS=train python script.py

script.py unpickles aquila/planners/best.pickle and calls Planner.plan(timeseries, locations, travel_costs, settings) per scenario, writing submission.csv. CPU only, single process. Measured 3m37s for 48 scenarios against the 30-minute limit; peak memory well under 32 GB.

The serve path degrades in tiers rather than raising, because one uncaught exception or a wall-clock overrun would void an entire run: full plan → voltage-threshold plan built from the raw series alone → defer-all, which is valid by construction. A time guard (AQUILA_TIME_BUDGET, default 1500 s) switches remaining scenarios to the fallback rather than overrunning. All tiers are verified by fault injection.

8. Dependencies, versions, licenses

No pretrained models and no external datasets are used. Every package is from the organizers' fixed environment; versions are those used in development.

All permissive (MIT / BSD / Apache-2.0), consistent with the competition's licensing requirement.

9. Commit history

Each tagged version below is a scored submission, kept intact for verification. Public-leaderboard total_cost:

v9 and v10 bracket the swap-count axis from both sides and confirm v6 as a local optimum: removing swaps costs +90.4 late per swap, adding them saves only −35.7. Negative results are documented rather than discarded.