A distribution planner in the Southwest opened the day-ahead forecast on a bright May morning and saw a system peak of 3,200 MW, comfortable, well within reserve margins. What the number did not show was that 800 MW of rooftop solar was actively suppressing the visible load, and that 120 MW of residential batteries were charged up and ready to discharge. By late afternoon, as clouds rolled in and EV charging began ramping up, the net load climbed 600 MW in 90 minutes. The AI model had seen the gross load, sort of. What it had not been explicitly taught was the difference between what the meter saw and what was actually happening on the other side of the service entrance.

Gross Load Versus Net Load: Why the Distinction Defines Modern Forecasting

The foundational concept in DER-adjusted forecasting is the distinction between gross load and net load. Gross load is the total electrical demand served by the utility system, including all distributed generation that flows through the utility's meters. Net load is what the utility actually sees at the substation level: the gross load minus any behind-the-meter generation that self-consumes or exports to the local feeder without showing up on the transmission-level meter. The gap between these two numbers is the invisible load, and in 2026, that gap is growing fast in every solar-penetrated territory.

Why this matters: an AI load forecasting model trained on utility meter data is trained on net load. It learned the patterns in what the meter saw, not in what customers were actually consuming. When behind-the-meter solar was small relative to system load, this distinction was manageable. When rooftop solar, battery storage, and EV charging add up to 15 or 20% of peak load, the gap between gross and net is large enough to create serious forecasting errors if the model is not explicitly designed to account for it.

The duck curve, familiar to operators in California and increasingly visible in the Southwest, Northeast, and Mid-Atlantic, is the visible signature of the gross-versus-net problem. The curve's midday dip is caused by behind-the-meter solar reducing net load during peak solar hours. The steep evening ramp, where net load jumps by hundreds of MW over a few hours, is caused by solar output declining just as residential demand rises and EV charging begins. An AI model trained only on net load data will learn the duck curve shape, but it will not know why it exists, and it will not be able to correctly project how the curve will change as DER penetration continues to grow.

The Double-Counting Trap in DER-Adjusted Forecasting

The double-counting trap is the most technically subtle error in DER-adjusted load forecasting. It occurs when behind-the-meter resources are counted twice in the same forecast: once because their effect is already embedded in the net-load training data, and once more when an explicit DER adjustment is applied on top.

Here is how it happens in practice. An AI model is trained on five years of hourly net-load data from the utility's advanced metering infrastructure (AMI). During those five years, solar penetration increased steadily. The model learned the evolving shape of net load, including the growing midday dip. It has implicitly absorbed the solar effect into its pattern recognition. Now the planning team, aware of solar growth, decides to add an explicit solar generation layer to the forecast: they subtract projected solar output from the model's output to arrive at a DER-adjusted forecast. The result is that solar has been subtracted twice: once from the net-load meter readings the model trained on, and once from the model's output. The adjusted forecast underestimates actual system load by double the solar amount.

The antidote is to understand exactly what the model was trained on before applying any adjustments. If the model was trained on net load (meter readings that already reflect solar export and self-consumption), adding an explicit solar subtraction is double-counting. If the model was trained on gross load (for example, using AMI data supplemented with behind-the-meter production estimates to reconstruct gross consumption), then explicit solar subtraction in the forecast step is appropriate. Ask the model vendor or your internal data science team this specific question: is this model trained on gross load or net load? The answer determines whether a DER adjustment is additive value or additive error.

Battery Storage: The Harder Double-Counting Case

Battery storage creates an even more complex double-counting risk. A behind-the-meter battery that charges during the day (increasing net load during solar hours) and discharges in the evening (reducing net load during the ramp) will leave both signatures in the net-load training data. If an AI model has learned this pattern from historical data and then the planning team also adds an explicit battery dispatch model to the forecast, the discharge effect is counted twice. This can make the evening ramp forecast too flat, understating the true net-load increase when batteries are fully dispatched and unable to provide additional flattening.

For large storage deployments, the correct approach is to explicitly model the battery state-of-charge trajectory over the forecast day and ensure the forecast model is not also embedding that trajectory in its output. This requires close coordination between the load forecasting function and the DER management team.

EV Load: The Emerging Ramp Problem

Electric vehicles are adding a new dimension of uncertainty to both gross load and net load forecasting. EV charging is largely demand-driven (when people arrive home) and partially controllable (smart chargers and managed charging programs can shift demand to off-peak hours). An AI model trained on historical data from a period of low EV penetration will underestimate the evening ramp in a high-EV-penetration service territory, because the charging load that now begins at 5 to 7 p.m. did not exist at that scale in the training data.

The magnitude of the EV forecasting challenge is growing quickly. In high-penetration areas, EV charging can add 50 to 150 MW of new evening load on a system basis, and significantly more at the individual distribution feeder level where EV adoption is concentrated in specific zip codes. A level-2 charger at 7.2 kW and a thousand vehicles charging simultaneously in a residential feeder zone represents 7.2 MW of load that materializes within a 30-minute window as commuters arrive home. At a system level, the ramp can exceed 1,000 MW per hour in high-penetration states after solar output drops and residential EVs connect simultaneously. That combined ramp shape (solar falling, EVs rising, residences still cooling) is steeper than anything a model trained on 2019 actuals has ever seen.

EV charging demand is also correlated with temperature in ways that differ from traditional residential load. Cold weather increases charging demand as battery efficiency drops and cabin heating draws from the battery pack, adding 15 to 30 percent more energy demand per vehicle on very cold days. An AI model that correctly handles traditional residential temperature sensitivity for space heating may not model the combined effect of residential heating load plus increased EV charging demand under cold conditions. This creates a winter peak forecasting gap in high-EV-penetration territories that has not historically been a significant risk but is growing as fleet sizes scale.

For planning purposes, the two questions to ask about EV load in any AI forecast are: first, whether the model includes explicit EV penetration estimates as a forward-looking input rather than assuming EV adoption is embedded in historical net load; and second, whether the model has been calibrated against actual EV charging profiles from the specific service territory, not from a national average. EV charging patterns vary significantly by housing type (apartment versus single-family), charger level (level 1 at 1.4 kW versus level 2 at 7.2 kW versus DCFC at 50+ kW), and local TOU rate structure. A model trained on a territory with a strong TOU peak-avoidance rate will have a very different EV ramp shape than a model trained on a flat-rate territory.

How to Spot a DER Forecasting Error: The Five Tells

A load forecaster reviewing an AI output can check for DER-related errors by looking for five specific patterns.

Tell 1: Midday Dip That Is Too Deep or Too Shallow

In a solar-penetrated territory, the midday (10 a.m. to 2 p.m.) net-load forecast should show a dip proportional to expected solar output given the day's cloud cover forecast. If the dip looks too deep relative to the solar forecast (suggesting double-counting of solar) or too shallow (suggesting solar is not being accounted for at all), this is a DER calibration signal. Cross-reference the forecast midday load against a simple calculation: take last year's comparable sunny day load, add any new load growth, and subtract today's projected solar output. The result should match the AI forecast midday to within a few percent.

Tell 2: Evening Ramp That Is Too Gradual

The evening ramp (typically 4 p.m. to 8 p.m.) in a solar and EV territory should be steeper than historical patterns because solar is declining and EV charging is beginning simultaneously. If the AI forecast shows a gradual, smooth evening ramp that looks like a pre-solar historical day, the model may not be capturing the combined solar decline and EV ramp correctly. Compare the forecast ramp rate to the last 30 similar evenings in actuals. If actuals are consistently ramping faster than the model predicts, you have a DER adjustment gap.

Tell 3: Weekend Versus Weekday DER Pattern Inversion

In commercial solar-heavy territories, weekday midday load may actually look higher than weekend midday load because commercial buildings' air conditioning demand exceeds the behind-the-meter commercial solar offset. On weekends, reduced commercial load combined with high residential solar can push midday net load to unusual lows. If the model treats weekend and weekday DER patterns symmetrically, it may produce errors on both day types.

Tell 4: Seasonal Solar Effect Not Tracking DER Growth

As solar penetration grows from year to year, the midday dip should be getting progressively deeper in each successive summer. If the AI model's midday forecast is not showing a year-over-year deepening that tracks actual solar capacity additions in the service territory, the model is not updating its solar penetration assumptions appropriately. This is a retraining and calibration issue.

Tell 5: Missing Cloud Cover Sensitivity

On heavily overcast days in a high-solar-penetration territory, behind-the-meter solar output drops sharply, and net load rises proportionally during midday hours. If the AI forecast does not show a significant midday load increase on overcast days relative to sunny days with the same temperature, the model is not correctly modeling the solar-cloud interaction. This can lead to underforecasting on overcast days when behind-the-meter solar fails to offset demand.

Building a Defensible DER-Adjusted Forecast

A defensible DER-adjusted forecast requires explicit accounting of the sources and their uncertainty, rather than treating DER as a residual. The following approach works for most utility planning contexts.

Start with the model output, understanding whether it was trained on gross or net load. If trained on net load with an older solar penetration baseline, estimate the incremental solar added since the training period (new residential and commercial installations, documented in your DER registry or with your interconnection data). Apply a solar correction: the incremental solar's expected midday output on the forecast day, expressed as an hourly MW profile. Do not subtract total solar, only the increment above what the model already knows about.

For battery storage, obtain the current installed capacity and expected dispatch profile from your DER management system or virtual power plant operator. If significant behind-the-meter storage is expected to discharge during the evening peak, apply a corresponding upward adjustment to the net-load forecast for those hours, because the model may be expecting storage discharge that no longer exists in the expected magnitude.

For EV load, if your territory has seen material EV adoption growth since the model's training period, apply an incremental EV load profile. This profile should be built from managed charging program data, time-of-use meter samples, or published adoption curve estimates for your region. A typical residential EV charging profile peaks around 6 to 8 p.m. on weekday evenings. The incremental magnitude depends on fleet size and charger power level.

Document each DER component separately: what was adjusted, the basis for the adjustment quantity, the uncertainty range on that quantity, and the source of the data. A DER-adjusted forecast that cannot show its component adjustments separately is not defensible in an IRP proceeding or a planning review.

The difference between a DER-aware forecast and a DER-adjusted forecast is whether you explicitly know what you are adding or subtracting, and why. Treating DER as a residual produces the double-counting trap. Treating each resource as an explicit, documented component is the professional standard.

A Worked Example: The May Afternoon Miss

Return to the Southwest planner from the opening story. Here is how a DER-aware verification workflow would have caught the problem before the afternoon ramp.

The AI forecast showed a 3,200 MW afternoon peak. The planner's first verification step is to check the DER component. She queries the DER registry: there are 800 MW of installed rooftop solar and 120 MW of behind-the-meter battery storage. The model's training data ends in September 2024. Solar capacity in the service territory has grown from 600 MW to 800 MW since then, an increment of 200 MW that the model does not fully know about.

She runs the DER check: on a typical sunny May afternoon, the 800 MW of installed solar generates about 600 MW. The model (trained on 600 MW of solar) was expecting about 450 MW of solar offset. The 150 MW gap (200 MW new solar at roughly 75% capacity factor) should have been added back to the net-load forecast during midday hours but subtracted from the gross load during evening hours when that solar stops producing. Instead, the model is already accounting for 450 MW of solar in its midday dip. If she were to subtract 600 MW of total solar from the model's output, she would double-count 450 MW.

She applies the correct incremental adjustment: 150 MW midday reduction (the new solar contribution the model does not know about) and a 150 MW evening increase (when that solar ramps off). She also notes that 120 MW of batteries are expected to discharge between 5 p.m. and 9 p.m., which the model treats as persistent net-load reduction. She checks the battery state-of-charge with the DER management team and learns that the batteries are fully charged after three sunny days. She adds 120 MW back to the 6 to 8 p.m. net-load hours because the model was built on data when batteries did not exist at this scale in the territory.

The revised forecast for the afternoon peak rises to approximately 3,470 MW. Operations is alerted to the steeper-than-expected evening ramp. When the clouds arrive that afternoon and EV charging begins, the actual peak reaches 3,490 MW: a 20 MW difference from the DER-adjusted forecast, versus a 290 MW difference from the original AI output. The planning team was prepared. They were not scrambling.

Key Takeaways

• Net load is gross demand minus behind-the-meter generation. An AI model trained on net-load meter readings already has DER effects embedded; applying an additional explicit DER subtraction without understanding this creates the double-counting trap that systematically underestimates load.
• The double-counting trap is most severe when incremental solar or battery capacity has been added since the model's training period and the planning team applies an adjustment for the total DER fleet rather than the increment the model does not know about.
• EV charging creates new evening ramp patterns that models trained on pre-EV-adoption data will systematically underestimate, particularly during peak summer or winter days when EV battery efficiency and home energy management add correlated load.
• Five DER tells signal a forecasting error in the output: a midday dip that is too deep or too shallow, an evening ramp that is too gradual, an incorrect weekend-versus-weekday DER pattern, seasonal solar growth not showing in successive forecasts, and missing cloud cover sensitivity on high-solar days.
• A defensible DER-adjusted forecast treats each resource component (incremental solar, battery storage, incremental EV load) as a separate, explicitly documented adjustment with its own uncertainty range and data source.
• The professional standard is to ask the forecasting model vendor one critical question before applying any DER adjustment: was this model trained on gross load or net load? The answer determines whether a DER adjustment is additive value or additive error.
• DER-related forecast errors are not random; they are structural and directional, and they become more consequential as DER penetration increases. Building explicit DER verification into the standard pre-peak workflow is not optional in any territory with more than 10% DER penetration relative to peak load.