Accuracy numbers are easy to publish and impossible to compare without context. When Ampgrove says 94 percent feeder load forecasting accuracy, that statement is only meaningful if you know the forecast horizon, the metric definition, the evaluation period, and the baseline being compared against. Here is exactly what we mean.
The number: 94 percent accuracy at the 4-hour horizon
Our 94 percent accuracy figure is measured as 1 minus the mean absolute percentage error (MAPE) of our 4-hour ahead forecasts evaluated against actual feeder load readings on held-out test data. A 6 percent MAPE means that on average, our 4-hour ahead forecast misses by 6 percent of actual load. On a feeder with a 1,000-amp peak, that is a 60-amp average error at the 4-hour horizon.
We calculate this metric at hourly intervals across the full evaluation period, not just at peak. Averaging across all hours means our accuracy figure includes easy-to-forecast overnight hours (where error is low) alongside the harder-to-forecast shoulder and peak periods. If we only reported accuracy at peak, the number would be worse. Reporting average-across-all-hours is the honest method.
What the baseline comparison is
The 94 percent figure from our pilot data compares against the same feeders' historical forecast accuracy using the utilities' existing forecasting processes. The average baseline MAPE across our three pilot utilities was 14 percent at the 4-hour horizon, meaning our models reduced average forecast error from 14 percent to 6 percent. The 94 percent is the new performance level; the improvement is from approximately 86 percent to 94 percent.
This is a meaningful but not extraordinary improvement. We are not claiming to solve a hard problem with magic. We are applying ML time-series techniques with good training data and feeder-specific model calibration to a problem that has historically been addressed with simpler, less data-intensive methods.
Where the 94 percent does not hold
Three conditions produce worse-than-average accuracy in our current models. First, feeders with very recent major structural load changes: a new large commercial customer, a new EV charging facility, or a significant demand response program that has been active for less than 90 days. The model has not yet trained on enough post-change data to have fully adapted. Second, extreme weather events that are outside the range of the training data. Third, data quality issues such as meter dropouts or calibration events during the evaluation period.
We report these conditions alongside our accuracy figures when we present pilot results. A utility evaluating Ampgrove should ask specifically about accuracy on their highest-complexity feeders, not just the average across the pilot group.
The number that matters operationally
MAPE averaged across all hours is a useful summary statistic, but the number that matters operationally is accuracy during the hours when forecast errors have consequences: peak load periods where a miss in the high direction means an overload event, and shoulder periods where a miss in the low direction means unnecessary capacity reservation cost. We track both of these separately and will share the hour-specific distribution on request.
How accuracy degrades across longer horizons
Load forecasting accuracy decreases as the forecast horizon extends. Our 4-hour ahead accuracy of 94 percent becomes approximately 90 percent at the 12-hour horizon and 87 percent at the 24-hour horizon, based on our pilot data. The degradation is driven primarily by weather forecast uncertainty accumulating over longer time windows. NWP models are substantially more accurate at 4 hours than at 24 hours, and that accuracy difference cascades into load forecast accuracy.
This is why we position 4-hour ahead forecasting as our core operational product. It is the horizon where weather uncertainty is still low enough to produce forecasts that support confident action, while still providing enough lead time for meaningful operational response. The 24-hour forecast we also generate is useful for planning purposes but we are explicit that it carries wider uncertainty bands and should inform softer decisions like crew scheduling rather than specific switching operations.
Feeder size and its effect on accuracy
Smaller feeders are generally harder to forecast accurately than larger ones. A large feeder with 2,000 customers has load diversity that dampens individual customer behavior variance: individual anomalies average out across the customer population. A small feeder with 150 customers is more sensitive to the behavior of a handful of large customers or to localized weather variation within the feeder's service area.
Our pilot data shows MAPE for feeders under 500 customers averaging around 8 to 10 percent at the 4-hour horizon, versus 4 to 6 percent for feeders over 1,000 customers. This is a known characteristic of distribution-level forecasting and is not unique to our models. We surface feeder size as one of the factors in our accuracy confidence displays so operators have appropriate calibration for forecasts on small feeders.
Comparing against vendor accuracy claims
When evaluating load forecasting vendors, the accuracy comparisons that matter are not comparisons between vendor-published numbers. Those numbers come from different evaluation methodologies, different feeder types, and often different definitions of what counts as the forecast horizon. The only valid comparison is a head-to-head evaluation on your specific feeders, evaluated against the same historical holdout period, using the same accuracy metric.
Ampgrove supports this kind of evaluation. We will provide our forecast output on a set of held-out feeders from your historical data, compared against your existing forecast outputs on the same feeders over the same period. The result is a direct comparison on your specific load profiles that cannot be obscured by methodological differences in how each vendor defines accuracy. If we do not outperform your current approach on your feeders, we will tell you that before you commit to a pilot, not after.
The broader point is that accuracy claims in this market need to be evaluated skeptically and with context. We publish our numbers with source qualifications because we believe that is the right practice for a company that is asking utilities to make real operational decisions based on our outputs. A forecast that is not honest about its limitations is not a tool you can trust.