What 90 days of pilot data showed about reserve cost reduction

Reserve procurement fell approximately 18% across 8 medium-voltage feeders during 90 days on a live distribution network in northern Portugal. We examine the underlying data, the sources of the reduction, and the parts of the pilot that fell short.

The measurement setup

The operator manages roughly 340 MW of peak demand across residential, light-industrial, and agricultural feeders. Before the pilot, its process used a 24-hour point forecast from a persistence model, manually corrected by the duty engineer each morning. By 09:00, the operator used that forecast to make next-day reserve procurement decisions.

We connected the eZap forecasting pipeline to 8 feeders using 15-minute smart-meter aggregates and ECMWF temperature forecasts from the Copernicus Climate Data Store. It produced P10, P50, and P90 probabilistic demand curves per feeder, refreshed every 15 minutes across a 48-hour horizon. The dispatch team ran both systems for weeks 1-4, then made eZap the primary input for weeks 5-13.

We tracked three measures: MAPE for the day-ahead P50 against actual consumption, standby generation volume contracted versus volume dispatched, and the operator's post-hoc assessment of scheduling comfort on high-demand days.

The reserve procurement outcome

Across 8 feeders, standby procurement volume dropped approximately 18% during weeks 5 to 13 versus the same calendar weeks in the prior year, adjusted for demand growth. Source: internal beta data from a 90-day closed pilot on one distribution network.

The 18% result varied by feeder. Residential-dominant feeders improved by 22-28%, mixed feeders by 14%, and agricultural feeders by 8%. The agricultural result has a specific explanation, which we address below.

A tighter P90 forecast band gave the duty engineer room to commit a smaller reserve margin while remaining confident that the obligation would be covered. The prior persistence model recorded around 8% MAPE on stable days and 14-16% during warm spells, when agricultural irrigation pumping produced demand spikes without a direct precedent in the prior day's curve. eZap's P50 MAPE was 4.2% on stable days and 7.8% on high-demand days over the same measurement period.

Where savings appeared, and where they did not

Most savings came during the day-ahead procurement window, through lower use of what the operator called "precautionary reserves". These were generation capacity contracts made 24 hours ahead for worst-case scenarios that seldom occurred. A more precise P90 estimate meant a smaller buffer was required.

Intraday corrections followed a different pattern. That adjustment process stayed unchanged during the pilot, so intraday re-procurement partly offset the day-ahead savings when actual demand exceeded the eZap P50 and the intraday forecast did not flag it in time. We estimate that closer coupling to the eZap 15-minute refresh could recover a further 6 to 9 percentage points of procurement efficiency.

The agricultural feeders underperformed for a clear reason. Irrigation scheduling partly follows water pressure decisions made by a separate infrastructure operator, and those decisions had little correlation with the weather signals available to us. We are not claiming eZap can forecast demand for which it has no inputs. This is not a modeling problem, it's a data availability problem, and data-sharing arrangements outside this pilot could address it.

Unexpected findings

Two results surprised us. By week 6, duty engineers had stopped running the old persistence forecast for the 8 pilot feeders and were using the P10 to P90 band as the main basis for reserve sizing discussions with their commercial team. The range matched their existing risk decisions better than a single point: the question became the worst plausible case to cover, rather than what demand would be.

The second surprise concerned SCADA integration. Writing forecasts into the operator's Energy Management System for its existing visualisation tools took longer than bringing data into eZap. The EMS vendor API had only been tested with single-point forecasts, so we restructured the probabilistic band into three separate time-series objects. We covered the SCADA integration specifics in a separate post.

A candid assessment at this stage

An 18% reduction in standby procurement across 8 feeders over 90 days is a genuine signal, not a number we engineered for the headline. We do not claim it will transfer directly to every network topology or operating context. Load mix, data quality, and intraday process maturity all affect the result.

For residential-and-mixed distribution feeders with reasonable smart-meter coverage and access to a good temperature forecast, probabilistic day-ahead forecasting delivered measurable procurement savings against a persistence-based baseline. The mechanism is well understood, and the pilot result aligns with what forecasting literature would predict.

We are now running a second cohort on 15 feeders across a second network with a different load mix. We will publish that cohort's results when the data is in, using the same methodology and the same candid account of where the model falls short.