DecarbAI Demands

Predicts 8760-hour heating, cooling, electricity, and DHW demand profiles for Swiss buildings using a per-city, per-building-type XGBoost + PCA ensemble, served through an interactive Dash dashboard.

Demo

Install

pip install decarbai-demands

Requires Python 3.11+.

Model download

Model artifacts are not bundled in the package. On the first prediction for a given city the package downloads {city}.tar (~100 MB) from a public S3 bucket and caches it at ~/.cache/decarbai/models/. Subsequent runs use the local cache.

Run the dashboard

After installing the package, launch the dashboard with the bundled CLI:

decarbai-demands dashboard

Or, from a source checkout:

uv run python -m decarbai_demands.dashboard

Opens an interactive Dash UI at http://127.0.0.1:8050/ for configuring a building and predicting its hourly heating, cooling, electricity, and DHW demand profiles.

Repo structure

src/decarbai_demands/
├── schema.py              Pydantic InputData + the 17-feature ordered vector
├── model.py               XGBoost + PCA ensemble loader (Model + load_artifacts)
├── pipeline.py            run_pipeline: feeds InputData through the per-city models
├── dashboard/             Dash app (subpackage)
│   ├── app.py             create_app() + run_dashboard() + __main__
│   ├── layout.py          UI tree (header, form sections, panels, modal hookup)
│   ├── callbacks.py       Prediction, CSV download, live range validation, modal toggle
│   ├── styles.py          Style dicts shared across the UI
│   ├── constants.py       Cities, building types, demand types, form defaults
│   └── ranges.py          TRAINING_RANGES + the info-modal builder
└── model_cache.py         S3 download + local cache at ~/.cache/decarbai/models/
                           Layout: {country}/{city}/{building_type}/{demand}.zst

Heating and cooling each have their own .zst model file; DHW and electricity share electricity_dhw.zst (first 8760 values = DHW, next 8760 = electricity).

Model training envelope

Predictions are most reliable when inputs lie within the ranges the model was trained on. Outside this envelope, the model extrapolates and confidence drops. The dashboard surfaces the same ranges inline next to each input and in the ⓘ info modal.

Direct inputs

Derived features (computed from geometry)

Credits

This tool was developed under the Innosuisse DecarbAI project, a UK/CH bilateral research collaboration.

• ML models & methodology — Dr. Hassan Bazazzadeh & Dr. Georgios Mavromatidis (Urban Energy Systems Laboratory, Empa)
• Building stock data & simulation inputs — BuildingAtlas, providing a scientifically grounded, data-rich foundation for Swiss and European building typologies
• ML deployment & dashboard — Youssef Sherif & Sympheny's Team (Sympheny)
• Funding — Innosuisse UK/CH bilateral grant