What it does
GridCast forecasts electricity demand for every site 48 hours ahead and says how sure it is: quantile outputs for every 15-minute step give the dispatch team a band rather than a line, so reserves can be planned against it. A nightly job retrains and backtests the model and promotes a new version only if its rolling error and interval coverage beat the current one.
What you get
- The Temporal Fusion Transformer training and serving stack, with connectors for 15-minute SCADA readings, weather forecasts and day-ahead prices.
- The dispatch dashboard with per-site forecasts, uncertainty bands, anomaly flags and attention-based explanations.
- The nightly retraining and backtesting pipeline, including the promotion rule and its report.
- API, CSV export, documentation, an onboarding session and a direct line to the engineers.
Deployment and licensing
Forecasts for 14 sites and 96 horizons take under two seconds on a CPU, so no GPU is required. The one-time licence is perpetual, runs in your infrastructure and includes twelve months of updates; the subscription adds monitored nightly retraining and can be cancelled monthly. Licensed per site tier.
Proven in production
For the regional grid operator it was built for, the model averages 2.8 % MAPE at 24 hours ahead against 7.4 % for the previous seasonal baseline, with a 90 % interval that covers 91.3 % of actual readings. The case study linked above has the details.