Where summer peak persistence forecasts break

A moderate overnight followed by a morning heat front can make yesterday's load curve a poor guide to today's peak. On these transition days, MAPE above 15% is common for residential-heavy feeders, and the error usually points in one direction: underestimating demand when cooling load rises faster than expected.

How persistence forecasting works and when it holds

The basic persistence model uses the assumption that demand at t+24h = demand at t. Some implementations also account for day-of-week effects, since one Saturday's profile may not be a useful guide when the comparable future period is shaped by weekday activity. Others add a seasonal adjustment. Even with those additions, the central idea remains the same: recent demand should provide the strongest signal for near-future demand, provided conditions change gradually enough for a 24-hour lag to remain useful.

That approach performs well when demand follows stable patterns, such as working-hour schedules, overnight heating cycles, and slow seasonal changes. It becomes unreliable when quickly changing weather drives load through a non-linear response.

The summer peak failure mechanism

In Iberian distribution networks, air conditioning is the main driver of summer residential peaks. Temperature does not translate into cooling demand in a straight line. Below approximately 22 to 24 degrees Celsius, most residential AC is not running. Above 28 degrees, the relationship is approximately linear. Between 24 and 28 degrees, the response varies widely according to building stock age, orientation, and occupancy patterns.

The most difficult case is a transition day. Overnight conditions are moderate, and the previous day's peak sits within the normal range. During the morning, however, a heat front moves in faster than forecast. The persistence model carries forward yesterday's familiar residential mid-evening shape, while today's demand begins climbing from noon as cooling systems switch on. The peak arrives in the mid-afternoon, well above the previous day's maximum, and remains high into the evening because buildings retain heat.

There is no temperature signal in a pure persistence forecast to anticipate that change. It reproduces yesterday's curve even when yesterday looked nothing like the day ahead. On these transition days, MAPE above 15% is typical for residential-heavy feeders, with the error skewed to underestimation.

The reserve procurement consequence

A 15% underestimate on a feeder carrying 180 MW of peak residential load creates a 27 MW forecast error. In day-ahead procurement, that gap must be covered by the reserve margin the duty engineer places above the persistence forecast. If the margin was set for normal-day error at 5 to 8%, the system is short by roughly 12 to 18 MW of committed generation capacity when the actual peak arrives.

Operators with experience of a network's summer behaviour often compensate manually. When the forecast indicates a hot day, the duty engineer may review the weather report, see that it will reach 35 degrees tomorrow, and add an intuition-based buffer. That can help, but the adjustment depends on individual experience. It is not systematic, and it does not quantify the uncertainty surrounding the correction. Someone new to the role, or covering another engineer's shift, may not have the same calibrated judgement.

What a temperature-aware model does differently

A model that receives temperature forecasts can learn the relationship between temperature and demand from historical data. It can capture the non-linear zone between 24 and 28 degrees, then apply that relationship to tomorrow's forecast temperature. The uncertainty in the weather input remains, so the approach does not remove the forecasting challenge. It does, however, replace an intuition-based manual adjustment with a quantified probabilistic estimate.

In our testing, a gradient-boosted model using temperature, day-of-week, and calendar features achieved MAPE around 6 to 7% on summer transition days for residential-heavy feeders. A pure persistence baseline produced 14 to 16% MAPE. The largest improvement appeared on the high-amplitude events that persistence handles poorly. On stable summer days, when yesterday's temperature was close to today's, the two approaches performed comparably.

The temperature-aware model's P90 output is especially useful during transition days. Since the model includes uncertainty in the temperature forecast, its P90 band widens when the temperature forecast has high variance. A wide P10-to-P90 range on a forecast-hot day communicates that both the eventual temperature and the amount of cooling load that will activate remain uncertain. That gives the operator more information than a single point estimate carrying an unstated buffer.

The limits of what any model can do with summer peaks

We are not arguing that temperature-aware models eliminate the summer peak forecasting problem. They reduce the error significantly and they represent the uncertainty more honestly. They still depend on the quality of the temperature forecast input, which degrades beyond 48 to 72 hours. They also remain vulnerable to novel demand patterns, such as a heat event occurring alongside a large public event that leaves residential buildings empty during the day, or a grid-connected industrial customer shedding load during peak heat periods in a way that is absent from its historical pattern.

Persistence should not be discarded because it performs poorly on summer peaks. For 300 of 365 days in a typical year, persistence is a reasonable approximation. The practical task is to identify the days when it is likely to fail. That signal comes from the temperature forecast, not from the historical load pattern alone.