Why weather data quality sets winter peak forecast accuracy

A 5-kilometre forecast-point offset can create a 1 to 2 degrees Celsius error on calm, clear nights. That illustrates why winter peak load forecasts depend heavily on temperature inputs. Residential heating drives demand more than any other single factor, but forecast quality also differs substantially among weather sources available to distribution operators.

Demand temperature coefficients differ by feeder type

Temperature does not affect every feeder equally. Its coefficient of demand, meaning the MW load change for each degree Celsius change, varies considerably by feeder type. Recognising that difference helps operators decide how much to invest in weather data quality.

In northern Portugal, a mostly residential feeder can see peak demand rise 15 to 22% when average daily temperature falls from 12 to 5 degrees Celsius, as electric space heating switches on. On a mainly commercial-office feeder, the same decline may raise peak demand 8 to 12%, since better insulation and HVAC systems reduce weather-sensitive marginal consumption. Where industrial process load dominates, the coefficient may stay below 5% because process energy use is not weather-driven.

Therefore, better weather data becomes more valuable as a feeder's residential share increases. A mediocre temperature forecast has little effect on load accuracy for an industrial-heavy feeder. On a high-residential feeder in a cold region, temperature forecast quality is the main determinant of winter forecast accuracy.

Four weather sources tested

During winter 2024 to 2025, we tested four sources on pilot feeders in northern Portugal and the Minho region: hourly temperature forecasts from IPMA, Portugal's Instituto Portugues do Mar e da Atmosfera; the ECMWF medium-range forecast through the Copernicus Climate Data Store; Open-Meteo's free API, which combines several NWP models; and a commercial provider with hyperlocal forecasts at 1 km grid resolution.

Each source was judged by spatial resolution against the feeder catchment and by accuracy at 24-hour and 48-hour horizons, compared with actual weather station readings.

Results from the testing

IPMA provides free hourly forecasts with dependable national accuracy, yet its spatial detail limits feeder-level modeling. A forecast point may sit several kilometres from the feeder in another microclimate, especially in northern Portugal, where terrain produces sharp temperature differences between coastal and inland areas and between valley floors and higher ground. The gap matters little in flat cities, but in complex terrain a 5-kilometre offset can cause systematic errors of 1 to 2 degrees Celsius on calm, clear nights with strong inversion gradients.

ECMWF delivered excellent results for large-scale temperature patterns across 24-hour to 72-hour horizons and led the test set at 48-hour forecasting. Its approximately 9 km resolution in ERA5 products available through Copernicus suits most feeder zones. Access is the main drawback: the Copernicus CDS API needs registration, and its query queue can delay real-time updates. That delay is generally immaterial for day-ahead planning, when weather inputs need only one or two daily updates.

Open-Meteo was strong at 24-hour horizons in coastal and flat areas and supplied a free API with fast responses. It trailed ECMWF at 48-hour horizons, consistent with an architecture that blends shorter-range models which lose skill sooner. For operators seeking a free day-ahead option with reasonable quality, Open-Meteo is a defensible choice.

The commercial hyperlocal service supplied the finest spatial detail, but its 24-hour to 48-hour temperature accuracy did not consistently exceed ECMWF's despite the higher price. Its clearest benefit appeared in 2 to 6-hour forecasts during convective events. Since winter peak forecasting is mainly a day-ahead task shaped by synoptic cold fronts rather than local detail, most feeders did not justify the premium.

When the weather model misses

The key winter peak failure is a cold event that arrives sooner, proves stronger, or covers more area than forecast. Such events are uncommon, but their consequences are clear: demand is underestimated, reserve procurement falls short, and the operator must activate emergency reserves or buy in the intraday balancing market at elevated prices.

The consequence depends on the forecast error's size, not merely its presence. A 2-degree error when temperatures already exceed the heating threshold matters less than a 4-degree error right at that threshold. In northern Portugal residential zones, demand responds non-linearly through the 8 to 12 degrees Celsius activation transition. A model that handles colder conditions well but misses that threshold is operationally worse than one with a slightly higher average error but better transition accuracy.

In the eZap model, P10 temperature scenarios from the ECMWF ensemble feed the P90 demand scenario. The upper demand band therefore includes a plausible cold-tail weather outcome rather than only the median temperature forecast. Linking weather uncertainty with demand uncertainty makes the P90 band useful operationally when winter peak risk is greatest.