On a sunny spring afternoon, demand is high yet thermal units are barely needed. Then the sun sets, solar output falls away and thousands of megawatts must be ready within hours. That swing is driven by net load, not load. This guide shows how to forecast net load with AI and use it to plan units, ramps and reserves. Start with a forecast review.
Net Load Forecasting for Power Plants with Wind and Solar
Forecast load minus wind and solar, not load alone. Size the evening ramp, set reserves from forecast uncertainty and commit the right units before the sun goes down.
- Why net load, not load, decides which units run
- Where net-load forecast error comes from
- How probability bands set ramp reserves
Load Versus Net Load
Thermal plants do not serve load. They serve what is left after wind and solar.
A classic load forecast predicts total demand. But the units a plant must commit depend on net load: the demand left over once wind and solar output is taken away. CAISO describes net load as the difference between forecasted load and expected output from variable resources. As renewable capacity grows, net load swings more than load ever did, and it is net load that drives starts, ramps and reserves. Our forecasting team can show what your net load really looks like.
Deep midday dip
Solar pushes net load down, so units back down to minimum or shut down.
Steep evening ramp
Sunset removes solar while demand stays high, so units must climb fast.
Wider uncertainty
Weather errors in wind and solar add to load error, so bands widen.
Solar on homes and businesses usually is not metered by the grid. It shows up as lower load on sunny days. A forecast that does not model it will miss midday load and then miss the evening ramp when that hidden solar fades. As rooftop systems keep growing, this error grows with them unless the model tracks installed capacity.
Where Net-Load Forecast Error Comes From
Three forecasts, three sets of errors, one combined miss.
A net-load forecast combines a load forecast, a wind forecast and a solar forecast. Each has its own error, and the errors do not always cancel out. On a cloudy afternoon, solar can come in far below forecast just as load rises, and the combined miss lands at the worst time. Plan your error breakdown in an accuracy review.
Temperature and calendar
Usually the smallest error. Driven by temperature forecasts, holidays and unusual events.
Speed at hub height
Errors grow with forecast horizon and are largest when wind is near cut-in or cut-out speeds.
Clouds and haze
Clear days are easy. Passing clouds, fog and dust cause sudden misses, often in the afternoon.
A daily average error can look fine while the evening ramp hours are badly wrong. Track error for each hour of the day and each season, because the hours that matter most for unit decisions are often the hardest to forecast.
From One Number to a Range: Probabilistic Forecasts
A single forecast hides the risk. A range shows it.
A point forecast says what is most likely. A probabilistic forecast also says how far reality could stray. The common form is a set of quantiles: P10, P50 and P90, where actual net load should fall below P90 nine times in ten. The gap between P50 and P90 is a direct guide to how much fast reserve to hold for the evening ramp. Questions on forecast types go to our support desk.
Low case
Net load should be below this only one time in ten. Useful for over-generation and curtailment risk.
Middle case
The most likely value. Used for base unit commitment and dispatch plans.
High case
Net load should exceed this only one time in ten. Used to size ramp reserves.
A good P90 is exceeded about 10% of the time, no more and no less. Track how often actual net load falls outside each band. If P90 is beaten 25% of the time, the forecast is overconfident and reserves are too thin.
What Better Net-Load Forecasts Are Worth
Each percent of forecast error costs real money in extra starts, idle reserve and expensive short-notice power. Net-load error is usually larger than load error, so the gains are often larger too.
Building a Net-Load Forecasting Model
Good inputs matter more than a clever model.
Machine learning models such as gradient-boosted trees and neural networks handle the many interacting drivers of net load well. But they can only learn from the data they are given, so input quality comes first. The inputs below cover most of what drives net load at a plant or balancing-area level. Map your data in a data review.
Renewable fleets grow every month. A model trained on last year's capacity will under-forecast solar and wind, and over-forecast net load, more each month. Feed new capacity into the model as soon as it connects.
From Net-Load Forecast to Unit Decisions
The forecast is only useful when it changes the plan.
A net-load forecast should feed straight into commitment and dispatch. That means turning bands into ramp windows, matching those windows to units that can start and ramp fast enough, and re-checking as weather forecasts sharpen through the day. Questions about the workflow go to our planning team.
Forecast with bands
P10, P50 and P90 net load for each hour of the next days.
Find ramp windows
Hours where net load climbs or falls fastest.
Match unit capability
Units with the start time and ramp rate each window needs.
Size reserves
Fast reserve set from the P50 to P90 gap in ramp hours.
Re-forecast intraday
Bands narrowed each hour and commitment adjusted.
Every unit has a latest time by which a start decision must be made. Waiting until then lets newer weather forecasts narrow the bands, which often avoids a start that would not have been needed. Show each unit's decision deadline next to the forecast so planners know how long they can wait.
Net-Load Forecasting Mistakes to Avoid
Most misses come from data gaps, not from the model.
Teams often blame the algorithm when a forecast misses, but the cause is usually a missing input, an out-of-date capacity figure or an error measure that hides the worst hours. A short rollout plan session helps avoid these traps.
What works
- Forecast net load directly, with its parts shown
- Model rooftop solar inside load
- Use bands to size reserves by hour
- Re-forecast as each weather run arrives
What fails
- Forecasting load and ignoring renewables
- Out-of-date installed capacity
- One daily accuracy number for every hour
- Fixed reserve rules regardless of weather
When a net-load forecast misses, planners need to know whether load, wind or solar caused it. Showing each part next to the total makes misses easier to explain and faster to fix, and builds trust in the forecast over time.
How iFactory Forecasts Net Load
Load, wind, solar and uncertainty in one on-prem model.
iFactory brings metered load, weather forecasts, renewable output and unit data onto an on-prem AI server inside your network. It forecasts each part of net load, combines them with probability bands and links the result to unit commitment and ramp plans. Fit questions go to our solutions team.
Net-load forecast
Load, wind and solar forecast together and shown as parts of one total.
Probability bands
P10, P50 and P90 for every hour, checked against actuals for honesty.
Rooftop solar estimate
Hidden solar inside metered load estimated from weather and capacity.
Ramp alerts
Large or uncertain ramps flagged early, with the hours and size.
Commitment links
Forecast bands matched to unit start times, ramp rates and reserves.
Intraday re-forecast
New weather runs fed in automatically, with changes to the plan highlighted.
Results depend on your data, weather inputs and renewable mix. We measure forecast error by hour, band reliability and reserve use on your own data during the pilot, rather than promising a general figure.
Turnkey AI: Delivered, Connected and Live in 6–12 Weeks
You do not build this. It arrives ready.
iFactory ships as a pre-configured NVIDIA AI server with the software pre-loaded, racked and ready. Rack it, plug in power and Ethernet, and the AI is live on your network. Our team handles cabling, network setup, DCS, SCADA, PLC, metering and weather data integration, planner training and 24×7 remote monitoring. Data stays on your network. For a scope matched to your system, request a turnkey quote.
Ship, network and data
Server installed. Load, weather, renewable output and unit data connected.
Model training and pilot
Net-load and band models trained on your history. Pilot forecasts run beside your current ones.
Go-live and training
Forecasts and ramp alerts live. Planners trained. 24×7 monitoring begins.
Frequently Asked Questions
What is net load?
Net load is demand minus the output of variable renewables such as wind and solar. It is the part of demand that controllable units, such as thermal plants, hydro and storage, must serve.
Why forecast net load instead of load?
Because unit starts, ramps and reserves depend on net load. With a lot of wind and solar, net load can swing far more than load, especially around sunrise and sunset.
What is the duck curve?
It is the shape of daily net load in solar-heavy systems: a deep midday dip as solar peaks, then a steep evening rise as solar fades. CAISO has described needing about 13,000 MW of extra supply within roughly three hours at spring sunsets.
What is a probabilistic forecast?
It gives a range instead of one number, usually as P10, P50 and P90 values. The range shows how uncertain each hour is, which helps size reserves and decide when to commit extra units.
How is rooftop solar handled?
It usually is not metered by the grid, so it appears as lower load. The model estimates it from weather and installed rooftop capacity, so midday load and the evening ramp are forecast correctly.
How is this different from a standard load forecast?
A load forecast predicts total demand. A net-load forecast also predicts wind and solar output, combines the three with their uncertainty and links the result to unit decisions.
Does battery storage change net-load forecasting?
Yes. Batteries charging at midday and discharging in the evening reshape the net load that thermal units see. Reports on CAISO show batteries cutting the evening ramp sharply on some days, so storage schedules belong in the forecast alongside wind and solar.
How far ahead can net load be forecast well?
Accuracy is best for the next few hours and the next day, when weather forecasts are sharpest. Beyond a few days, wind and solar errors grow quickly, so longer horizons are more useful for planning outages and fuel than for committing units.
How do we start?
Start with one balancing area or plant portfolio and a year of load, weather and renewable data. A 6-week pilot runs net-load forecasts beside your current ones so you can compare. To plan it, contact our team.
See Tomorrow's Ramp Before It Arrives
In thirty minutes we look at your load, renewables and how you forecast today, and show where net-load forecasting could change your unit decisions. You keep the notes whether or not you go further with iFactory.
- 1A year of hourly load data
- 2Wind and solar output in your area
- 3Installed renewable capacity, including rooftop
- 4Your current forecast and its errors
- 5A day when the evening ramp caught you short
