Energy demand forecasting

Planning ran on last year's averages. Peaks were missed and nobody could say how far off the estimates usually were.

• ✦Feature engineering on consumption, weather and calendar data
• ✦Gradient boosting and neural models compared with rolling backtests
• ✦Prediction intervals so planners see the uncertainty, not just a number
• ✦MLflow-tracked models deployed as a scheduled scoring job

A forecast with a known error margin, checked every day against reality.