All case studies
Aurelis Energy, Energy and utilities
One load forecast the grid planners actually trust
Nine regional forecasting spreadsheets became a single model service, retrained weekly and monitored in production alongside the systems that consume it.
- lower day-ahead forecast error
- A third
- planning cycle time
- 5 days to 1
- Platform
- Snowflake, Apache Airflow, Kubernetes, PostgreSQL, Terraform
The situation
Nine regional teams each maintained their own day-ahead load forecast in a spreadsheet. When two regions disagreed, the dispute went to a weekly meeting rather than to a model, and procurement hedged against the gap.
What we did
- Consolidated fourteen years of interval meter data, weather history and outage records into a single governed store, with the regional definitions reconciled before any modelling started.
- Built one forecasting service with regional parameters rather than nine separate models, so an improvement in method reaches every region at once.
- Put the service behind the same monitoring as the rest of the platform: input drift, prediction distribution and downstream consumption are all alerted on.
- Ran the new forecast in parallel with the spreadsheets for a full quarter before anyone was asked to switch.
Where it landed
Planners now review one forecast and spend their time on the exceptions it flags. The spreadsheets were retired, not banned, because the model earned the swap during the parallel run.