Day-ahead forecasts built unit commitment. Short-interval forecasts govern dispatch. Using the same model — or the same temporal resolution — for both creates a structural mismatch that shows up as balancing costs, not as error metrics.
The Two-Problem Confusion
Utility forecasting evolved around a specific operational workflow: commit generating units the night before, and manage deviations in real time with regulating reserves. The day-ahead forecast was designed to answer one question — how much generation to commit — and at that task, 2–4% MAPE is genuinely adequate.
The problem is that balancing authorities now face a different question in a different timeframe. When the ISO/RTO publishes 5-minute dispatch instructions and your actual net load deviates by 80 MW from your schedule, the relevant question isn't "what will load be tomorrow at 2pm?" It's "what will load be in 15 minutes, right now, with current weather and current grid state?"
A day-ahead forecast with 3% MAPE for an 800 MW peak system represents a possible error of 24 MW at the 15-minute level. In practice, short-interval errors compound because day-ahead models don't capture intra-hour load ramps, cloud cover transitions, or large commercial HVAC cycling. The 15-minute ahead error for a typical 24-hour model applied at short intervals often runs 8–12% — not 3%.
What the MAPE Number Hides
Mean Absolute Percentage Error aggregated over an entire day obscures the distribution of errors. A day-ahead model with 2.5% MAPE can produce 15-minute intervals where the error exceeds 15% and still average out fine over 96 half-hour periods. The problem is that FERC's real-time imbalance charges don't average out — they're assessed per interval.
Balancing authorities operating under NERC BAL-001-3 track Control Area Error (ACE) accumulated over 10-minute periods. A model that is accurate on average but produces concentrated errors in morning ramp periods — when HVAC loads surge and solar generation is rising — will generate ACE violations precisely at the times that matter most for frequency regulation.
Short-interval forecasting models trained specifically for 15-minute and 5-minute resolution address this by: (1) incorporating SCADA telemetry updated every polling cycle rather than only weather forecasts, (2) applying separate ensemble members calibrated for different weather regimes, and (3) weighting recent errors more heavily in the confidence interval calculation.
The Unit Commitment vs. Real-Time Dispatch Distinction
Unit commitment is a forward-planning problem with 12–36 hour horizons. The cost of a commitment error is a startup cost or forced shutdown — expensive but manageable. Real-time dispatch operates at 5-minute intervals with costs tied to imbalance energy prices, which in LMP-based markets can spike to $1,000/MWh during scarcity conditions.
This is why balancing authorities with variable renewable penetration above 15% of installed capacity need separate forecasting models for the two problems. The input features differ (day-ahead needs weather patterns; real-time needs cloud transmittance and SCADA deviation from previous interval), the training labels differ (next-day energy vs. next-interval MW), and the acceptable error profile differs entirely.
A study of ERCOT settlement data from 2022–2023 found that utilities relying on day-ahead models for real-time imbalance management paid an average of $4.20/MWh more in imbalance charges than those with purpose-built 15-minute forecasting systems — not because their day-ahead accuracy was worse, but because they were applying the wrong tool to the wrong problem.
Where Short-Interval Models Fail Too
Short-interval forecasting isn't universally better. For the purpose of generating unit commitment schedules submitted to the ISO/RTO 24 hours in advance, a gradient-boosted 15-minute model offers no advantage over a well-calibrated day-ahead ARIMA or neural network model. Running 96 successive 15-minute predictions to cover a 24-hour window magnifies error accumulation rather than reducing it.
The practical solution is not to replace day-ahead forecasting with short-interval forecasting but to maintain both: a day-ahead model for unit commitment and ancillary services scheduling, and a short-interval model that updates in near-real-time for dispatch decisions and DR trigger thresholds. As we discuss in our article on net-load forecasting with high distributed solar penetration, the real complexity arrives when behind-the-meter generation makes the "load" value you're forecasting a composite of two uncertain variables.
Implementation Considerations for Balancing Authorities
Moving from single-model to dual-model forecasting requires several operational changes that don't always appear in vendor pitch decks. First, the short-interval model needs a low-latency data feed from SCADA — polling intervals of 2 minutes or less. Models that update on hourly weather feeds are not short-interval models in the relevant sense.
Second, confidence intervals matter more for dispatch decisions than for unit commitment. A dispatcher deciding whether to issue a curtailment event needs to know not just the forecast MW but the probability that load exceeds a specific threshold. Point forecasts are insufficient for that decision — probabilistic forecasts with 80th and 95th percentile bands change the decision calculus substantially.
Third, the feedback loop between forecast error and DR dispatch performance requires logging at the interval level. Aggregate weekly reports obscure the specific weather conditions or load patterns where the model underperforms. Granular error logs tied to timestamp, temperature, and humidity enable targeted model recalibration rather than wholesale retraining.
The Operational Takeaway
Utilities that consolidated on a single day-ahead forecasting system made a reasonable decision when that was the best available technology. The emergence of affordable ML infrastructure and low-cost SCADA-to-cloud data pipelines changes the calculus: purpose-built short-interval forecasting is now within reach for mid-sized utilities without dedicated data science teams.
The question to ask of any forecasting system isn't "what is the aggregate MAPE?" It's "how does error distribute across the critical 15-minute intervals during morning ramps and afternoon peaks?" That's the number that appears in your settlement statements.