A forecast that moves as fast as your load does.
Prelko re-forecasts every site, feeder and portfolio at 15-minute resolution, from the next hour out to seven days. Each forecast carries a confidence band, so planners see not just the expected load but how sure the model is.
What goes into every forecast.
Load is a function of history, weather and behaviour. Prelko models all three, per site, and learns how much each one matters for your network.
Load history
At least 12 months of interval data is ideal; forecasting can start with 3. Gaps and meter faults are detected and repaired before training.
Weather
Temperature, humidity, heat index, cloud cover and wind, observed and forecast for each site's coordinates, refreshed with every numerical weather run.
Calendar & behaviour
Public holidays, festivals, school terms, industrial shift patterns, tariff time-of-day bands and daylight hours.
Operational signals
Planned outages, network reconfigurations and demand-response events, so the model does not learn from abnormal days.
Embedded generation
Rooftop and utility solar and wind output, so gross and net load are forecast separately and reconciled.
Hierarchy
Feeder, substation and system forecasts are reconciled so the parts always add up to the whole.
One model, every planning horizon.
Intraday dispatch, day-ahead scheduling and week-ahead procurement all read from the same forecast, so the numbers never disagree between teams.
| Horizon | Typical use | Refresh |
|---|---|---|
| 1-6 hours | Intraday balancing, real-time dispatch preparation | 15 min |
| Day-ahead | Scheduling, day-ahead market bids, storage plans | Hourly |
| 2-7 days | Week-ahead procurement, maintenance windows, staffing | Every 6 h |
An ensemble, not a single guess.
Prelko combines gradient-boosted models, which capture weather and calendar effects, with sequence models that track the shape of recent load. Their outputs are weighted by how each has performed on your data over the last few weeks.
Forecasts are probabilistic. Alongside the expected value (P50) you get P10 and P90 bounds, which is what peak detection and dispatch recommendations use to judge risk.
Clean
Interval data is validated, outliers flagged and missing reads filled.
Learn
Models train per site on load, weather and calendar features.
Forecast
Ensemble produces P10, P50 and P90 for every 15-minute interval.
Score
Each interval is compared to actuals and the weights update.
Load forecasting questions
How much history do we need to start?
Three months of interval data is enough to produce a first forecast. Accuracy improves as the model sees a full year of seasons, holidays and weather extremes.
What resolution does Prelko forecast at?
15-minute intervals by default, matching most scheduling and settlement periods. 30-minute and hourly outputs are aggregated from the same forecast.
Can we forecast individual feeders and the whole system?
Yes. Forecasts are produced at each level of your hierarchy and reconciled, so feeder forecasts sum to the substation and system totals.
Works best as part of the full loop.
Peak Demand Detection
Peak windows flagged with lead time and confidence
Learn moreDispatch Recommendations
Balancing actions proposed, approved by your operators
Learn moreRenewable Integration
Solar and wind variability built into every forecast
Learn moreGrid Performance Reports
Forecast accuracy and balance trends, tracked over time
Learn moreSee your own load, forecast.
Share a sample of interval data and we will walk through a forecast built on it, including where the model is most and least certain.