The forecast looks accurate until you price the errors
A wind forecast has small average bias but misses the most valuable interval. The analyst must explain why a single accuracy metric is not enough.
Permanent case linkYour task
Measure error, distinguish financial weighting from accuracy, and design an out-of-sample test before replacing the forecast.
Your role
Renewable forecasting analyst reviewing model performance
Practice data & worked solution
Bias can hide costly misses
Original synthetic four half-hour intervals for a 100 MW wind farm. Error is forecast minus actual MW. Spot prices are fixed. The value gap is forecast energy receipts minus actual spot receipts; it is not an imbalance penalty, a hedging P&L or evidence of forecast-caused price movement. No curtailment, loss factors or contract settlements are included.
| Interval | Forecast MW | Actual MW | Spot AUD per MWh |
|---|---|---|---|
| 1 | 40 | 30 | 50 |
| 2 | 60 | 70 | 100 |
| 3 | 80 | 50 | 300 |
| 4 | 20 | 30 | -20 |
Your analysis
- Calculate signed errors, mean bias, MAE and RMSE in MW, then express MAE as a percentage of nameplate capacity.
- Reconcile forecast and actual energy and spot receipts. Explain why the value gap is not a uniform price times the net energy error.
- Subtract 5 MW from every forecast. Recalculate bias, MAE, RMSE and value gap. Would these four rows justify deploying the corrected model?
Worked solution
Accuracy scorecard
Errors are +10, −10, +30 and −10 MW. Mean bias is +5 MW; MAE is 15 MW, or 15% of the 100 MW rating; RMSE is √300 = 17.32 MW. Normalisation is by nameplate here, not actual generation. Small average bias does not mean small individual errors.
Financial weighting
Forecast energy is 100 MWh; actual energy is 90 MWh. Forecast spot receipts are A$15,800 and actual receipts A$11,450. The A$4,350 gap equals the sum of each error × 0.5 hours × its price. The third interval contributes A$4,500 by itself; other errors partly offset it.
Correction and validation
Corrected errors are +5, −15, +25 and −15 MW. Bias becomes zero; MAE remains 15 MW and RMSE improves to √275 = 16.58 MW. Forecast receipts fall by A$1,075, so the value gap becomes A$3,275. Four retrospectively corrected rows are not a deployment test. Freeze the correction, test on unseen dates and compare horizons, weather regimes and constrained intervals.
Review your response
- Metrics: signed and absolute errors are not confused.
- Time: half-hour duration is included in energy and value calculations.
- Interpretation: financial weighting is separate from statistical accuracy and actual settlement.
- Validation: no training/test leakage or claim of proven improvement from four observations.
This checklist supports self-review. No tutor marking or independent credential is included.
Related topics 5
Sources & evidence
All figures are original teaching inputs, not observations or published forecasts. The official source supplies industry context, not this dataset. Source link checked 17 October 2026. Compare with AEMO forecasting guidance; actual analysis needs forecast issue time, target interval, availability, dispatch constraints and quality flags.
AEMO solar and wind energy forecasting