Demand Forecasting Accelerator
ML-based hourly demand forecasting with automated MAE/RMSE/MAPE accuracy monitoring, horizon-bucketed analysis, and weekly retraining. Databricks Asset Bundle.
Product Preview
Try the interactive reader and demo tools below, or get the full product with all content unlocked.
⚙ Try Demo ToolsFile Structure 17 files
Documentation Preview README excerpt
Demand Forecasting Accelerator
ML-based hourly demand forecasting with automated accuracy monitoring and
weekly retraining, packaged as a deploy-ready Databricks Asset Bundle. Trains a
Gradient Boosting model on consumption + weather history, produces rolling
72-hour forecasts, and tracks MAE/RMSE/MAPE by horizon. Typical **10–30% MAE
improvement** over rule-based methods.
Fully generic and environment-agnostic: configure your own catalog/schema and
source tables. No organization-specific dependencies.
What's inside
- Databricks Asset Bundle — inference (every 2h), accuracy monitoring
(daily), and weekly retraining jobs with MLflow model registry.
- Training + inference notebooks — GBM with calendar, cyclical, degree-day,
and lag features; recursive multi-step forecasting.
- Accuracy + data-quality SQL — MAE/RMSE/MAPE by horizon bucket and
freshness checks.
- A dependency-free feature library (
lib/demand_features.py) that mirrors
the notebook feature engineering, with an offline test suite (6 tests).
Why degree-days
Heating and cooling load track degree-days (distance of temperature from a
comfort base), not raw temperature. That single transform is the most
predictive weather feature — see guides/demand-forecasting-methodology.md.
Quickstart
pip install pytest && pytest tests/ -v # validate feature logic offline
databricks bundle deploy -t dev -p <profile>
Provide hourly_consumption, weather_observations, and weather_forecasts
tables (schemas in guide/03_architecture.md).
License
MIT — see LICENSE.