# Load forecasting: predicting demand from weather, behavior, and DERs
A heat wave is forecast to hit your service territory next Tuesday. Your day-ahead load forecast says peak demand will hit 8,200 megawatts (MW) at 5 PM. You procure power to match. But you were off by 3 percent. That 250 MW gap just cost your utility real money in the real-time market, where prices during a heat wave can spike to ten times the day-ahead price.
That is the entire game of load forecasting: get the number right, or pay for being wrong.
A load forecast is a prediction of how much electricity customers will consume over a future period. Utilities and grid operators use it to decide how much generation to schedule, how much power to buy in advance, and whether the grid can handle demand.
The key term is day-ahead procurement: buying most of your expected power one day in advance in the wholesale market, where prices are calmer and more predictable. Whatever you fail to procure, you must buy in the real-time market (also called the balancing market), which settles in five-minute or fifteen-minute intervals and is far more volatile.
Forecast well, and you buy cheap power in advance. Forecast poorly, and you scramble in real time at whatever price the market demands.
Forecast error is usually measured as MAPE (Mean Absolute Percentage Error): the average percentage difference between predicted and actual load. Utility day-ahead MAPE typically lands in the 1 to 3 percent range for a stable region.
For an 8,000 MW system, a 2 percent error is 160 MW, roughly the output of a mid-sized power plant. Over-procure, and you sell the excess back at a loss. Under-procure, and you buy the shortfall at a premium. Both directions cost money, but under-procurement during a scarcity event is the expensive one.
Old-school forecasting was mostly about weather and the calendar. Today you have to model three distinct forces.
Temperature is still the single strongest driver. The relationship is not linear: it is a U-shape. Demand is lowest around 18°C (65°F) and rises in both directions as heating and cooling kick in.
The most useful engineered variable here is cooling degree days (CDD) and heating degree days (HDD), which measure how far the temperature sits above or below a comfort baseline. A 35°C afternoon drives air conditioning load hard; a cold snap drives electric heating.
Humidity, cloud cover, and wind matter too. A humid 32°C feels hotter, so air conditioners run longer.
People are creatures of habit. Load has strong daily and weekly cycles: a morning ramp, an evening peak, low overnight demand, and lighter weekends. Holidays break the pattern (a Monday holiday looks like a Sunday).
Behavioral shifts matter more than they used to. Remote work flattened the sharp morning commute ramp in many regions after 2020. Time-of-use pricing, where electricity costs more during peak hours, nudges customers to shift laundry and dishwashing to off-peak windows.
This is what makes 2026 forecasting genuinely hard. DERs are small energy resources located at the customer site: rooftop solar panels, home batteries, and electric vehicle (EV) chargers.
Here is the problem. The utility only meters net load: the power flowing across the customer's meter. It does not directly see how much a rooftop solar array is generating behind that meter.
Rooftop solar creates the famous duck curve: a demand shape that sags in the middle of the day as solar floods the system, then ramps up steeply in the early evening as the sun sets and everyone comes home. The California grid operator popularized this term, and you can see the shape in the CAISO duck curve explainer.
EV charging adds a new evening spike. If thousands of drivers plug in at 6 PM, that stacks right on top of the existing evening peak.
Let us build the logic for that heat wave Tuesday.
Step 1: Start with the baseline. Pull historical load for similar days: same season, same day of week, similar temperature. This is your reference shape.
Step 2: Overlay the weather forecast. Feed in tomorrow's hourly temperature, humidity, and cloud cover. Your model maps CDD to expected cooling load. A 35°C peak pushes the afternoon curve up sharply.
Step 3: Subtract expected solar injection. Estimate how much your rooftop solar fleet will generate. On a clear hot day, midday solar is high, which suppresses net load. But clouds change everything, and solar forecast error is a major source of overall error.
Step 4: Add expected EV charging. Layer in the evening charging load, which is growing every year.
Step 5: Add the ramps. The dangerous moment is the evening. Solar drops to zero as the sun sets, air conditioning is still running, and EVs plug in. Net load can climb steeply in a couple of hours. Grid operators care intensely about this ramp rate because generators need time to spin up.
Here is a simplified sketch of the net-load logic in code:
# Simplified day-ahead net load estimate (per hour)
def net_load(base_demand, temp, solar_output, ev_charging):
# Cooling load rises as temperature exceeds 18C comfort point
cooling = max(0, temp - 18) * COOLING_SENSITIVITY
gross_demand = base_demand + cooling + ev_charging
net = gross_demand - solar_output # behind-the-meter solar
return net
# Evening ramp = the gap between the 6 PM peak and the midday solar trough
evening_ramp = net_load_6pm - net_load_1pmThe real models are far more sophisticated (gradient-boosted trees, neural networks, or weather-ensemble methods), but the accounting logic is exactly this: gross demand, minus solar, plus EVs, tracked hour by hour.
Knowledge check
1. Why does a small percentage forecast error (like 2 percent) matter so much for a large utility?
2. What is the primary strategic reason utilities procure most of their power in the day-ahead market rather than waiting for real time?
3. During a scarcity event like a heat wave, why is under-procurement generally more costly than over-procurement?
4. Select ALL correct answers about how load forecasts are used by utilities and grid operators.
Select all the correct answers.
5. Select ALL correct answers about forecast error and MAPE.
Select all the correct answers.
Now connect the forecast to the money.
Suppose your day-ahead forecast for the 5 PM peak is 8,200 MW, and you procure exactly that. Actual load comes in at 8,450 MW. You are short 250 MW.
You must buy that 250 MW in the real-time market. During a heat wave, real-time prices can be many multiples of the day-ahead price. If day-ahead was around 60 dollars per MWh and real-time spikes to 600 dollars per MWh during scarcity, that shortfall gets expensive fast. (These are illustrative figures; actual prices vary widely by region and event.)
The reverse also costs you. If you over-procure by 250 MW, you are stuck selling surplus power back into a market that does not need it, often at a loss.
The trickiest error is behind-the-meter solar. Imagine you assumed clear skies and high solar output, so you procured less power. Then clouds roll in, solar underperforms, and net load jumps unexpectedly. You are now short during peak hours, buying at scarcity prices.
This is why utilities invest heavily in solar irradiance forecasting and satellite cloud tracking. A better solar forecast directly reduces procurement risk.
Modern practice is shifting from a single number to a probabilistic forecast: a range with confidence bands. Instead of "8,200 MW," the forecast says "8,200 MW expected, with a 90 percent chance between 7,900 and 8,600 MW."
That range lets a utility make smarter bets. If the upper tail is dangerous and expensive, you might procure a little extra as insurance, or line up flexible demand response (paying large customers to cut usage on command) as a cheap backup. The U.S. Energy Information Administration publishes accessible background on how electricity demand and forecasting work.