Buildings account for roughly a third of global energy use, and much of it is wasted quietly: an air-handling unit left running over a weekend, a lighting schedule that never switched back after a holiday, a chiller cycling against a failed sensor. A raw meter reading cannot say whether 280 kW is a problem — at 2 p.m. on a hot July weekday it is normal, at 2 a.m. on a Sunday in April it is not.
Wattnorm learns what a building should draw for the current time and weather, scores every 15-minute meter reading against that expectation as it arrives, and raises direction-aware alerts on a Grafana dashboard — too high = waste, too low = failure — with the avoidable energy expressed in kWh, CO₂e and cost so that a building manager, not a data scientist, can act on it. It was developed on six years of data from a university research building. Historical data is replayed through the live pipeline today; a real meter publishes the same message format tomorrow.
Two weeks of real meter data replayed through the live pipeline. Blue: metered demand · orange dashed: expected demand · shaded: 90 % normal range · red: alerts. Lower panels: residual in σ and the CUSUM that catches sustained deviations.
Excess energy on HIGH alerts, converted with the grid emission factor and tariff (both editable on the dashboard; the defaults are placeholders). A normal building reads zero.
make setup # uv sync + pre-commit
make test # unit tests
make up # mosquitto + timescaledb + mlflow + grafana + scorer (Docker Compose)
make train # train + calibrate in the stack → MLflow @staging
make eval # evaluate vs baseline on the held-out year; promotes @production on PASS
make backfill # load the whole 2020-01 → 2021-05 history fast (background)
make live # human-paced demo: one building day per ~15 min (START=… INJECT=sustained_offset)
make retrain # trailing-window retrain with promotion gate (AS_OF=YYYY-MM-DD)
make figures # regenerate docs/img/fig*.png and numbers.json from the stack's model and database
| Service | URL |
|---|---|
| Grafana | http://localhost:3000 (view without login; admin/admin to edit) |
| MLflow | http://localhost:5001 |
| Scorer health | http://localhost:8001/health |
Publish {"meter_id": "...", "ts": "<ISO-8601 with offset>", "kw": <float>} to
building/<meter_id>/demand (QoS 1) and weather to weather/<station_id>/obs. Nothing downstream
changes; the broker and database are .env settings.
data/raw/demand.csv (gitignored) is the 15-minute electric demand of Clark Hall, Cornell University,
in kW (the portal reports building electricity as power; the pipeline converts each slot to kWh as
kW × 0.25 h). The export's clock is UTC (site.demand_tz in configs/default.yaml); calendar
features are computed in America/New_York. The data are not redistributed with this repository.
Cornell University, Facilities and Campus Services — Energy and Sustainability. Energy Management and Control System (EMCS) Portal, Clark Hall electric demand, 15-minute interval, January 2015 – May 2021. https://portal.emcs.cornell.edu (accessed 2022 for the original study; re-used here). Data are the property of Cornell University and are used for research and educational purposes.
Weather is the Open-Meteo ERA5 archive for the building's location (make weather).
Wattnorm builds on my master's research at Western University, published as An ensemble learning framework for anomaly detection in building energy consumption (Araya, Grolinger, ElYamany, Capretz and Bitsuamlak, Energy and Buildings 144, 2017). The paper combined CCAD-SW, a pattern-based autoencoder classifier over sliding windows, with support vector regression and random forest predictors in a majority-vote ensemble that, on data from a building in Brampton, Ontario, raised CCAD-SW's sensitivity by 3.6 % and cut its false-alarm rate by 2.7 %. Wattnorm takes the prediction-based branch as its core and adds what a deployment needs: calibrated bands, adaptation to a building that changes, a promotion gate, and a dashboard.
docs/writeup.md— methodology, data work, evaluation, and the defects found along the waydocs/runbook.md— operating guide for building managersdocs/PLAN.md— design, phases and decisions
The Python package is smartbuilding (src/smartbuilding/).
MIT — see LICENSE. The Clark Hall data are Cornell University's and are not covered by it.