Forecasting heat and electricity demand for a district utility
Hourly forecasts with calibrated uncertainty bands replaced spreadsheet estimates, so the dispatch team now commits plant and purchases balancing energy against a stated confidence level.
- Sector
- Energy and utilities
- Location
- France
- Duration
- 6 months
- -52%mean absolute forecast error
- 89%observed coverage of the 90% interval
- 48 hhourly horizon, retrained weekly
The challenge
Day-ahead demand was estimated by hand from last year’s profile and the weather forecast. Errors on cold mornings forced expensive balancing purchases and over-firing of peak boilers.
Meter, weather and plant data sat in three systems with different time bases and gaps.
What we did
A single hourly data foundation was built: reconciled meter readings, weather observations and forecasts, calendar effects and plant availability, with explicit handling of clock changes and missing intervals.
Forecast models were compared on rolling back-tests. The chosen model produces hourly point forecasts and prediction intervals for 48 hours ahead, retrained weekly.
Forecasts are delivered into the dispatch tool with the interval shown as a band, and a daily accuracy page lets the team see when the model is drifting.
The outcome
Mean absolute error fell by roughly half against the manual method, and the 90% interval covered 89% of observed hours over the first heating season.
A representative engagement. The figures are indicative of typical outcomes for this kind of work and are not drawn from a named client.
Contact
Tell us about the problem you need to solve
You may have a defined project, an underperforming process or a technical question that needs investigation. We can help assess the challenge, identify the information required and define an appropriate next step.