For the Forecast Review Analyst
Does this demand forecast need closer review?
Review electricity demand forecasts across 51 US balancing authorities. Identify the forecast shown, check its observed error, and review demand, weather, and data-quality warnings before deciding what to investigate.
Forecast review workflow
Start with the forecast you are reviewing
-
Identify the source
Read the forecast name and published default. A selected model comparison may differ from the forecast currently published.
-
Check the matching evidence
Inspect observed error for that series, including the forecast horizon, measurement window, and scoring coverage. Keep training results and data limitations in view.
-
Decide what to investigate
Review the next demand peak, forecast curve, weather, and flagged observations to decide whether the forecast needs closer review.
Why it exists
What you can inspect
A forecast curve alone does not explain whether it needs attention. Review its source, observed error, demand history, and weather together, with update times and data warnings in view.
Demand and weather
Compare reported demand with forecasts and review the weather inputs used by the models.
Forecast error
Check observed error for the displayed series at a named forecast horizon and window, with scoring coverage. Horizon offsets run from forecast start; training evaluations are separate.
Weather alerts
Review active National Weather Service alerts and flagged demand observations for the selected region.
Weather scenarios
Adjust temperature, wind, or solar assumptions and inspect the resulting demand estimate. Each result labels its calculation method.
Dashboard views
What each view is for
Overview starts with the Forecast Review Analyst's decision. Renewables, trading, and data-science contexts offer other starting views and historical summaries using the same underlying data.
Methods and results
How results are checked
Read the model methods, data-quality rules, published evaluations, and failure analysis behind the dashboard. Model comparisons include XGBoost, Prophet, SARIMAX, and Ensemble.
Design decisions
The design record explains model weighting, data-quality rules, and the validation required before a retrained model can serve forecasts.
ADR trail in PRD.md →Forecast accuracy
Holdout backtests predict 168 hours recursively: each prediction feeds into the next. Results are published for all 51 balancing authorities. A separate live benchmark compares named model series with each operator’s day-ahead forecast on matched observations. The benchmark may measure a different series from the published default and does not score the fallback baseline.
In the published 2026-08-07 recursive holdout, the evaluated ensemble has a median per-BA error of 4.35%, compared with 3.69% for XGBoost alone. The ensemble has lower error on 21 of 51 authorities. Combining models does not consistently improve accuracy; the per-authority results show where it helps and where it does not.
30-day model-comparison benchmark (~24h lead), across 46 operators: GridPulse model median error 4.94%; operator forecasts 3.63%. GridPulse’s median error is 1.31 percentage points higher.
See the head-to-head benchmark → Ensemble results → Backtest results →Data sources and revisions
EIA publishes provisional readings that can revise heavily before settling. GridPulse measures each feed's revision behavior, classifies it live, and discloses exclusions in plain language on the dashboard.
The anchor-conditioning study →Forecast failure analysis
The published failure analysis explains a forecast drop and the validation check added afterward.
The forecast-dive diagnosis →Project notes
GridPulse is an independent project using public data.
Developed with AI assistance.
Practitioner validation of the workflow and its operational value is the next step.