Load forecasting for distribution systems has been a solved problem, in rough terms, for several decades. The fundamental approach is regression-based: historical load data, weather inputs (temperature, humidity, cloud cover), day-of-week and calendar effects, and sometimes economic indicators. The models produce reasonably accurate short-term (day-ahead, week-ahead) and medium-term (seasonal, annual) load projections that planners and operators use for capacity planning, procurement, and day-of grid operations.
The problem is that those models were trained on historical data from a grid where load was driven by consumption and conventional generation was bulk-delivered from transmission. Distribution feeders had no generation. The meter at the service entrance measured load in one direction. The AMI interval data that trained the next generation of forecasting models also captured this structure: all intervals are positive load values, with occasional near-zero values during low-consumption hours.
Feeders with significant DER penetration break that structure. The AMI intervals for a customer with rooftop solar may show negative values (export to grid) during midday hours. The feeder aggregate shows non-monotonic behavior: load drops sharply at solar ramping (9 to 11 AM on clear days), recovers partially as air conditioning load increases, then drops again if battery storage shifts midday solar generation into the afternoon. A feeder load model trained on pre-solar historical data will treat these patterns as anomalies rather than structural features, and will produce systematically biased forecasts.
The shape of the bias
The forecasting bias from unmodeled DERs is not random noise that averages out. It is structured and predictable in its direction if not its magnitude, which makes it especially dangerous for operations planning.
On high-solar days (clear sky, summer, high angle of incidence), traditional models will overestimate midday feeder load because they don't subtract the solar generation that will offset customer consumption. The overestimate in midday peak load can lead to over-procurement of capacity resources or over-commitment of demand response capacity for an event that won't actually be needed at the forecast level. That over-procurement has a cost, but operationally the risk is manageable: you committed more than you needed, not less.
The underestimate case is operationally more dangerous. On high-solar days with late-afternoon battery discharge, traditional models will also underestimate the load recovery ramp that occurs as solar generation falls after 3 to 4 PM and battery systems that exhausted their afternoon discharge cycle come back online as loads rather than generators. The load recovery ramp in high-DER-penetration circuits can be steep (100 to 200 kW per 15-minute interval on a feeder with 500 kW of solar-plus-storage) and longer-lived than the traditional afternoon demand peak that the conventional model was calibrated on. Operations teams running traditional load forecasting models are frequently surprised by how sharp and persistent the late-afternoon ramp is on days when solar generation was high.
This is sometimes called the "duck curve" problem at the transmission level, but it manifests at the distribution feeder level in a more granular and harder-to-manage form because feeder-level DER penetration varies widely across the service territory. The same utility may have feeders where DER penetration is 5% (traditional model works fine) and feeders where DER penetration exceeds 30% (traditional model is systematically wrong), and the aggregate load forecast that covers the entire service territory masks the feeder-level errors with regional averaging.
Why behind-the-meter generation makes the net load problem harder
The AMI data problem is more subtle than it first appears. For a customer with rooftop solar but no battery, the AMI meter at the service entrance reads net consumption: grid consumption minus solar generation, plus any solar export. This net metering architecture means the AMI data the utility sees for solar customers has already subtracted the customer's solar generation, so the utility doesn't see the gross consumption or the gross generation separately. It sees only the net.
This creates a problem for disaggregation models that try to estimate behind-the-meter solar generation from AMI data. The common approach is to compare the solar customer's AMI interval data to a statistical profile of non-solar customers with similar characteristics (same rate class, similar climate zone, similar building size estimate). The difference between the solar customer's interval data and the non-solar profile is attributed to solar generation. The method works at a portfolio level when the samples are large enough, but the error per individual customer is high, and the error aggregates across feeders with high solar penetration in ways that compound the forecasting error.
For utilities that want feeder-level load forecasting accuracy in high-DER-penetration territory, the net metering data structure is a fundamental limitation. Accurate feeder-level DER impact modeling requires direct knowledge of behind-the-meter generation, not net consumption data. That means inverter-level telemetry: actual generation data from each solar inverter, separate from the net consumption reading at the AMI meter. Without that data, feeder-level forecasts in high-DER areas are bounded in accuracy by the quality of the statistical estimation of behind-the-meter generation, which is lower than the accuracy achievable with direct measurement.
Battery storage adds a scheduling dimension the statistical model can't learn
Solar forecasting, while complex, is largely a function of irradiance and temperature: given weather forecast inputs, the generation profile of a solar fleet is partially predictable. Battery dispatch is a different problem. Battery SoC transitions are driven by economic optimization logic, program enrollment status, DR event signals, and the owner's manual overrides. These are not functions of weather; they are functions of incentive structures and decision logic that change as rate designs change and programs change.
A statistical model trained on historical battery behavior will learn the patterns that existed when the training data was collected. When the utility changes a TOU rate structure, or launches a new demand response program, or adjusts the DR event trigger thresholds, the battery fleet's collective behavior changes as enrolled home energy management systems and BEMS controllers adapt to the new incentive structure. The statistical model's learned patterns become stale at the next program change, which in a dynamic grid environment can happen multiple times per year.
DER-aware load forecasting that handles battery storage requires an explicit model of battery behavior, not a statistical approximation learned from historical data. The explicit model takes as inputs: enrolled capacity and SoC distribution by feeder, DR program dispatch schedule and event history, TOU rate structure, and historical dispatch patterns from the battery management platform. It outputs a load impact estimate for battery charge and discharge activity by feeder and hour. This is a different data requirement than the weather-driven statistical approach: it requires a live connection to the DER management platform's enrollment and dispatch data, not just AMI historical data and weather inputs.
What DER-aware forecasting looks like in practice
The forecasting architecture that handles DERs correctly is a layered model, not a single-model replacement for the traditional approach.
Layer one is the traditional statistical load model, which continues to handle bulk load behavior driven by weather and calendar effects. Layer two is a DER generation offset model that estimates solar generation by feeder using inverter telemetry data (where available) or statistical methods calibrated against direct measurement samples. Layer three is a battery dispatch model that projects battery charge and discharge activity by feeder for the forecast horizon based on current SoC distribution and scheduled program events. The three outputs are combined into a net load forecast that accounts for all three drivers.
This architecture requires investment in the data infrastructure that feeds layers two and three: inverter-level telemetry for solar generation (not just net AMI reads), current SoC data from the battery management platform, and programmatic access to the dispatch schedule for enrolled programs. Where that data infrastructure exists, feeder-level forecast accuracy in high-DER-penetration circuits improves substantially. Where it doesn't, the forecast remains structurally bounded by the statistical estimation methods, which may be adequate for planning purposes but are too coarse for real-time operations decisions on feeders where DER behavior is material to the load shape.
Getting that infrastructure in place before the forecasting errors become operationally expensive is the challenge. The utilities that will navigate the transition to high-DER-penetration service territories best are the ones that recognized early that the net load forecasting problem requires DER data, not just historical AMI data, and started building the integrations while there was still operational headroom to do it without urgency.