When AEMET issues red heat alerts for Zaragoza and Valencia, as it has for 7 July with forecasts touching 44°C, somewhere in every Spanish utility a model is being asked a deceptively hard question: not whether demand will rise, but by exactly how many megawatts, in which cities, at what hour. Getting that answer wrong in either direction is expensive. Under-forecast, and the shortfall must be covered in the volatile intraday market. Over-forecast, and money is committed to generation nobody needs. This article looks at how modern load forecasting models are built, which weather features actually improve them, and how to judge whether a forecast is any good.
What "Dynamic" Load Forecasting Means
Traditional load forecasting produced a handful of static curves: a typical summer weekday, a typical winter Sunday. Dynamic load forecasting replaces those templates with a rolling prediction that updates continuously as new weather data arrives, an intraday view for the next few hours, a day-ahead view for market bidding, and a week-ahead view for scheduling and procurement. Each horizon has a different tolerance for error and leans on different weather inputs: the intraday model cares about the cloud bank arriving at 15:00; the week-ahead model cares about whether the heat episode breaks on Thursday or persists into the weekend.
The Weather Features That Actually Matter
Raw temperature is the obvious input, but mature models rely on engineered features that capture how buildings and people actually respond:
Degree days and thresholds. Demand doesn't respond to temperature linearly, it barely moves between 15°C and 22°C, then climbs steeply as air conditioning switches on. Cooling and heating degree days, computed against locally calibrated base temperatures, capture this hinge behaviour far better than the raw reading.
Lagged temperature. Buildings have thermal mass. The demand peak on day three of a heat episode is higher than on day one at the same air temperature, because walls and interiors have been soaking up heat for 48 hours. Models that include yesterday's and the day before's temperatures consistently outperform those that only see today's forecast.
Humidity and "feels like" temperature. Thermostats are set by people, not thermometers. A humid 32°C drives more cooling load than a dry 35°C in many climates, which is why apparent-temperature features often carry more predictive weight than dry-bulb temperature.
Solar irradiance. With rooftop solar widespread, what the grid observes is consumption minus self-generation. A bright afternoon suppresses measured demand; a passing cloud field makes it rebound. Irradiance has quietly become a demand-side feature, not just a generation-side one.
Calendar interactions. The same 38°C afternoon produces a different load shape on a school day, a Saturday, and a public holiday. The predictive power comes from the interaction of weather with the calendar, not from either alone.
Training Honestly: The Look-Ahead Trap
A subtle but common mistake is training models on historical weather observations. In live operation, the model will never see observations of the future, it will see forecasts, with all their imperfections. A model trained on perfect hindsight learns relationships it cannot reproduce in production, and its backtest results flatter to deceive. Best practice is to train on archived forecasts as they were issued at each decision point, paired with the demand that actually materialised. This teaches the model not only how weather drives load, but how forecast uncertainty at each horizon should temper its confidence.
Judging a Forecast: Beyond Average Error
Mean absolute percentage error is the headline metric, and well-run systems achieve low single digits day-ahead. But averages hide what hurts. An error of 2% spread evenly across a mild day costs little; the same 2% concentrated in the evening peak of a red-alert day can force emergency purchases at the day's highest prices. Serious evaluation therefore weights errors by when they occur - peak-hour accuracy, accuracy on extreme days, and the timing error on the daily maximum, because those are the moments when the forecast is actually being paid for.
Watching the Drivers in Real Time
However good the model, operators still need to see its inputs moving. The OpenWeather Energy Dashboard gives operations teams a live view of the demand drivers themselves: temperature, humidity, and irradiance across a service territory, updated continuously, so that when reality starts to diverge from the morning's forecast, the team sees it in the weather data first, hours before it fully shows up in the metered load.
The Payoff
Load forecasting is one of the rare areas of grid operations where a modest technical improvement compounds daily. A model fed with well-engineered weather features, trained honestly on archived forecasts, and evaluated on the hours that matter will quietly save money every single trading day, and earn its keep many times over on the handful of days each summer when AEMET's maps turn red.