GuideElectric Power & Natural Gas

How to forecast net load as distributed energy resources keep growing

As rooftop solar, batteries and electric vehicles spread, metered load stops describing what customers actually consume. Forecasting net load well means estimating the hidden generation, forecasting gross demand and solar separately, reconciling feeder and system forecasts so they agree, and scoring probabilistic outputs against the decisions they feed. This guide sets out a practical approach.

Reviewed 7 min read

On this page
  1. Why DER growth breaks traditional load forecasts
  2. Components of a net load forecasting pipeline
  3. Building a net load forecast, step by step
  4. Choosing a method to estimate behind-the-meter solar
  5. Forecast horizons and what changes between them
  6. Reconciling feeder, substation and system forecasts
  7. Scoring probabilistic net load forecasts
  8. Hypothetical example: residential feeders with fast rooftop solar growth
  9. Questions and answers
  10. Sources

Why DER growth breaks traditional load forecasts

Traditional load models learn how weather and the calendar drive metered demand. Behind-the-meter solar subtracts generation the utility cannot see, so metered load sags on sunny middays and climbs steeply as the sun sets. The California ISO's duck curve made this shape familiar at system level: net load, demand minus variable generation, dips in the afternoon and ramps hard into the evening2.

Three effects follow. The daily shape changes, with a deeper midday trough and a sharper evening ramp. Weather sensitivity flips, because a cloudy day now raises metered load. And the relationship drifts as installed capacity grows month by month, so last year's history describes a different system. Batteries and electric vehicles add behavior-driven components on top, shaped by tariffs, managed charging programs and aggregator dispatch. Energy demand forecasting and renewable integration both appear among the use cases ColdAI lists for power and gas operators1, and this is the method we would start from.

Components of a net load forecasting pipeline

01Weather and DER data02Estimate hidden solar03Forecast gross load04Combine to net load05Reconcile the hierarchy06Quantiles to operations
  1. Weather and DER data

    Irradiance, temperature and cloud forecasts, the DER registry and meter data.

  2. Estimate hidden solar

    Reconstruct behind-the-meter generation, past and forecast.

  3. Forecast gross load

    Model underlying demand on weather, calendar and customer mix.

  4. Combine to net load

    Subtract solar and add battery and EV adjustments.

  5. Reconcile the hierarchy

    Make feeder, substation and system forecasts coherent.

  6. Quantiles to operations

    Deliver forecast ranges to commitment, dispatch and planning.

Conceptual pipeline for net load forecasting. Stages can be merged in one model, but keeping them visible makes errors easier to trace.

Building a net load forecast, step by step

  1. Start from the decisions

    List what the forecast feeds: unit commitment, reserve and energy procurement, DER dispatch, feeder planning. Each sets a horizon, resolution, spatial level and the uncertainty measure it needs, such as a high quantile for reserves.

    Output
    Forecast requirements sheet
    Owner
    Operations and planning

  2. Assemble DER and weather data

    Gather the interconnection registry with capacity, location and commissioning date, interval meter data, irradiance and cloud forecasts from weather models and satellite nowcasts, temperature and humidity, holidays, and enrollment in EV and battery programs.

    Output
    Joined, time-aligned dataset

  3. Reconstruct gross load history

    Estimate behind-the-meter solar for every past interval and add it back to metered load. The model then learns demand drivers without the masking effect of solar output.

  4. Model gross load and solar separately

    Forecast gross load from weather and calendar, and solar from forecast irradiance and installed capacity at that date. Subtract solar and apply battery and EV adjustments; benchmark against a single model of net load with solar features.

  5. Reconcile across the hierarchy

    Forecast at feeder, substation and system level, then reconcile so the parts sum to the whole, as described below.

    Output
    Coherent forecasts at every level

  6. Produce ranges, not single numbers

    Generate quantiles or scenarios with quantile regression, gradient boosting under a quantile loss or weather-model ensembles, so operators can size reserves for the evening ramp.

  7. Evaluate by decision and watch for drift

    Back-test by season and weather regime, score ramps and peaks as well as averages, and retrain on a schedule tied to growth in installed capacity.

    Owner
    Forecasting team

Choosing a method to estimate behind-the-meter solar

  • If

    You hold a DER registry with capacity and location for most systems.

    Then

    Use a physical model: convert irradiance forecasts to output per system with open tools such as pvlib, then scale by registered capacity3.

    It needs no metered generation and responds immediately to new installations.

  • If

    A sample of rooftop systems has separately metered generation.

    Then

    Upscale: learn how that sample performs per unit of capacity and apply it to the registered total, by area and orientation where known.

    Real measurements capture soiling, shading and inverter limits a physical model misses.

  • If

    You only have net meter data and an incomplete registry.

    Then

    Disaggregate: separate solar from consumption using the correlation between net load and clear-sky irradiance, and flag premises where unregistered systems appear.

    It works with what you have and doubles as a check on registry completeness.

  • If

    Several of these sources are available.

    Then

    Combine them, using the physical model as the backbone and metered samples to calibrate it.

    Each method's error sources differ, so blending reduces them.

Forecast horizons and what changes between them

AspectIntradayDay-aheadWeek-ahead
Main decisionsReal-time dispatch, DER dispatch, ramp managementUnit commitment, market bids, reserve sizingMaintenance windows, fuel and procurement planning
Most useful weather inputSatellite cloud nowcasts and latest measurementsNumerical weather prediction, ideally ensemblesEnsemble forecasts and climatology
Dominant error sourcePassing clouds and data latencyCloud cover and temperature forecast errorWeather uncertainty itself
Typical model familiesPersistence corrected by recent errors, fast ML on lagged dataGradient boosting or neural networks on weather featuresStatistical models with weather scenarios
Update cadenceEvery few minutes to hourlyAs each weather run arrivesDaily

Boundaries between horizons vary by market and utility; the point is that inputs and error sources shift, so one model rarely serves every horizon well.

Reconciling feeder, substation and system forecasts

Forecasts made independently at different levels rarely add up, and planners notice when feeder forecasts imply a system peak the operations team does not expect. Bottom-up aggregation keeps local detail but inherits noisy feeder errors; top-down allocation is stable but blind to local solar growth.

Reconciliation methods forecast every level and then adjust them to be coherent. The minimum trace approach finds the combination that minimizes total forecast error variance given how errors correlate across the hierarchy, and reconciled forecasts are often more accurate than unreconciled ones, especially at sparse levels4. An open textbook explains the methods with worked examples5.

Scoring probabilistic net load forecasts

Mean absolute percentage error, the habitual load metric, misbehaves once net load approaches zero or turns negative on sunny afternoons: small absolute misses become enormous percentages. Use absolute or squared error in megawatts for point forecasts, normalized by peak where comparisons are needed.

For ranges, pinball loss scores each quantile, and was the headline metric in the probabilistic load track of the Global Energy Forecasting Competition 20146. Pair it with interval coverage, checking that a nominal ninety-percent band contains about that share of outcomes, and with ramp and peak-timing errors, which matter more to operators than average accuracy.

Hypothetical example: residential feeders with fast rooftop solar growth

Questions and answers

What is the difference between gross load and net load?

Gross load is what customers actually consume, regardless of where the electricity comes from. Net load is what remains for the grid to supply after subtracting variable generation, especially behind-the-meter solar. Utilities measure net load at substations and meters, so as solar grows, gross load has to be estimated rather than observed, and forecasting both gives better results.

How can a utility estimate behind-the-meter solar without meters on every system?

Combine an interconnection registry, irradiance data and a physical model of solar output, then calibrate it with a smaller set of systems that do have generation meters. Where the registry is incomplete, net meter data can reveal unregistered systems through their midday pattern. Satellite-derived irradiance helps on feeders without nearby weather stations.

Why is MAPE a poor metric for net load forecasts?

Mean absolute percentage error divides each error by the actual value. When net load nears zero or turns negative on sunny afternoons, the denominator collapses and small misses produce enormous or undefined percentages. Absolute errors in megawatts, pinball loss for quantiles and interval coverage give a far more honest picture of forecast quality as solar grows.

How should electric vehicle charging be handled in a net load forecast?

Treat it as its own component once it is large enough to see in feeder data. Unmanaged home charging follows arrival times and tariffs, while managed charging follows the program's control signals, so enrollment data and tariff schedules become forecast inputs. Where vehicle counts are growing fast, scenario forecasts for planning are more useful than history alone.

Sources

  1. Electric Power & Natural Gas: energy demand forecasting and renewable integration — ColdAI
  2. What the duck curve tells us about managing a green grid — California ISO · checked 10 October 2026
  3. pvlib python documentation — pvlib project · checked 10 October 2026
  4. Optimal Forecast Reconciliation for Hierarchical and Grouped Time Series Through Trace Minimization — Journal of the American Statistical Association · checked 10 October 2026
  5. Forecasting: Principles and Practice, chapter on forecasting hierarchical and grouped time series — OTexts · checked 10 October 2026
  6. Probabilistic energy forecasting: Global Energy Forecasting Competition 2014 and beyond — International Journal of Forecasting · checked 10 October 2026
  7. IEEE 1547-2018: Standard for Interconnection and Interoperability of Distributed Energy Resources with Associated Electric Power Systems Interfaces — IEEE · checked 10 October 2026

More in Electric Power & Natural Gas

Back to Electric Power & Natural Gas

Next step

Send us a feeder where your forecasts miss on sunny days

Share the feeder's interval data, the DER registry extract and your current forecast. We will reply with where the error comes from and whether a net load approach would fix it.

Discuss net load forecasting