BatterySwapAI 2026 — survival-model RUL + expected-cost work-order planner

Approach

1. Label reproduction. bsw/daily.py (reference) / bsw/daily_fast.py (polars, bit-identical) aggregate the hourly series per device-day and reproduce the official EOL criterion exactly (smooth_series() voltage < 2.40 V; 461/461 train devices match).
2. Remaining-useful-life models. Two complementary forecasters are blended in logit space (bsw/blend.py):
   • bsw/features.py + bsw/model.py + bsw/ensemble.py: per-device features at cut time T (only data before T: smoothed-voltage level/slopes/extrapolations, raw-voltage statistics, temperature and seasonality, temperature sensitivity, age/replacement history) feed a feature-view-diverse ensemble of multi-horizon gradient-boosted survival models (HistGradientBoostingClassifier, horizon as input, monotone in horizon and voltage) that output P(EOL <= T + h) on a grid of horizons.
   • The re-entry component uses static C5 device calibration, guarded by a target-free compositional certificate. It preserves the full C5 effect unless either (a) supported pre-/post-T-42d phases have opposite signs, or (b) the device sign opposes the median of at least three other supported nonzero devices in its current building. A contradiction abstains exactly to the fleet base; missing views never veto.
   • bsw/physics.py: a discharge-curve forecaster on the temperature-compensated voltage (population prior dU/dt = r(U) fitted on the train split, device-specific rate multiplier, seasonal room-temperature forecast from the device/building/fleet history before T, explicit modelling of the label's temperature filter), giving a second survival CDF.
   • Blend weights and logit-scale calibrations were determined by building-grouped 6-fold CV on the train split (analysis/cv_ensemble.py): OOF planner cost 1416 h/scenario (single GBM: ~1800; oracle: ~101).
   • Serving uses a causal support guard after two missed weekly reporting intervals: when a device has supplied no observation for at least 14 days, its CDF is set to zero before the unchanged optimizer replans. This avoids extrapolating supervision selected toward streams that later resume to permanently silent devices; an observation stamped exactly at the current cut explicitly cancels the guard.
3. Planner. bsw/sim.py is an exact, fast re-implementation of the cost accounting of batteryswap_public.evaluate.evaluate_plan (time, overtime, daily/weekly limits — including the travel-home carry-over), verified against the official evaluator. bsw/planner.py computes, for every device and candidate day, the expected early/late penalty (using the evaluator's conventions for unobserved EOL, per-device data end and emergency visits) and runs a multi-start local search over swap-day assignments minimising expected penalties + exact routing/overtime/limit costs. Devices not worth swapping are scheduled after the window.
4. Submission. script.py loads the stored artifacts (artifacts/) and produces submission.csv for all splits via batteryswap_public.utils.make_submissions. Each scenario is planned from its own inputs only (data up to the scenario start, locations, travel costs, settings).

Reproduction

python -m venv venv && . venv/bin/activate
pip install -r requirements-train.txt     # exact versions used (Python 3.10.x)
python train.py --dataset_path /path/to/dataset   # writes artifacts/ (GBM ensemble, physics prior, meta.json); ~10 min on 16 cores
BATTERYSWAP_DATASET_PATH=/path/to/dataset BATTERYSWAP_SPLITS=train python script.py
BATTERYSWAP_DATASET_PATH=/path/to/dataset BATTERYSWAP_SPLITS=train BATTERYSWAP_SUBMISSION_PATH=submission.csv python -m batteryswap_public.metric

Files

• script.py — competition entry point.
• train.py — trains the model and writes artifacts/.
• bsw/ — solution package (see above).
• artifacts/ — survival_model.joblib (GBM ensemble, scikit-learn 1.7.2), physics_prior.joblib (population prior + base CDF), meta.json (features, knots, members, blend constants).
• requirements.txt — the competition runtime requirements (unchanged); requirements-train.txt — pinned versions used for training.
• THIRD_PARTY.md — third-party dependencies and licenses.