Definition
Forecast error decomposed by weather regime (clear sky, partly cloudy, overcast, cloud-edge transition for solar; steady, ramping, lull, storm approach for wind) rather than reported as a single aggregate. Reveals regime-specific bias that an aggregate accuracy figure averages away.
The Measurement Gap
Every renewable energy operation with merchant or semi-merchant exposure depends on production forecasts to bid into energy markets. The standard industry practice is to measure forecast accuracy as a single aggregate number: mean absolute error across all hours, all conditions, all seasons. That number is nearly useless for decision-making because it masks the regime-specific biases that actually cost money.
A solar forecast model might show 4% aggregate error, which looks acceptable. But inside that average, the model over-forecasts by 12% during cloud-edge transition conditions (the hardest regime to predict) and under-forecasts by 3% during clear-sky conditions (the easiest). The cloud-edge over-forecast generates imbalance penalties every time it occurs. The clear-sky under-forecast leaves revenue on the table. The aggregate metric sees 4% and calls it good. The trading desk sees cash flow volatility and does not know where it comes from.
FRBI decomposes forecast accuracy by weather regime and measures directional error, not just magnitude. This distinction matters because a model that is consistently 5% high in one regime can be corrected with a simple bias adjustment. A model that is randomly scattered by 5% in all directions cannot be corrected without fundamental retraining. The directional decomposition tells you which problems are fixable today and which require model architecture changes.
The decision FRBI supports is whether the energy forecast model is reliable for all weather conditions the trader is bidding into, or whether systematic over-forecasting in some regimes is generating avoidable imbalance penalties while under-forecasting in others is leaving revenue uncaptured.
Formula
Classify each forecast period into a weather regime. For solar, the regimes are clear sky, partly cloudy, overcast, and cloud-edge transition. For wind, the regimes are steady wind, ramping, low-wind lull, and storm approach. Within each regime, calculate the mean signed error, not absolute error. A positive result means the model over-forecasted (the operator bid too much and pays penalties). A negative result means the model under-forecasted (the operator bid too little and left money on the table).
Formula (Notation):
FRBI(r) = (1/N_r) times the sum of (Forecast_i minus Actual_i) / Actual_i, for all periods i in regime r
Companion Metric (Regime Financial Impact):
RFI(r) = sum of (Forecast_i minus Actual_i) times Imbalance_Price_i, for all periods i in regime r where Forecast exceeds Actual
Variable Definitions:
| Variable | Definition |
|---|---|
| r | weather regime classification |
| Forecast_i | forecasted energy production for period i, in MWh |
| Actual_i | actual energy production for period i, in MWh |
| N_r | number of forecast periods classified as regime r |
| Imbalance_Price_i | the real-time imbalance settlement price for period i |
Edge Cases and Measurement Boundaries:
Curtailed periods should be excluded or flagged separately, since actual production was constrained by grid operator orders rather than forecast error. Regime classification boundaries should use meteorological definitions (cloud cover percentage, wind speed thresholds) rather than subjective labels. Regimes with fewer than 30 periods per quarter should be aggregated into broader categories to avoid small-sample-size noise.
Proposed formula, pending validation pilots.