By Robert Mullin
LAS VEGAS — Increased adoption of behind-the-meter generation is complicating short-term load forecasting across the Western Energy Imbalance Market (EIM), especially in the Arizona Public Service area.
The challenge is caused by the unpredictability of cloud cover, which can cause sharp and sudden drops in solar production.
“In the past, cloud cover was always a variable that came in for load forecasting, but it was really interrelated to temperatures,” Amber Motley, CAISO manager of short-term load forecasting, said during a March 1 meeting of the EIM Governing Body at The Palazzo hotel.
The conventional understanding: Clouds would move over an area, causing temperatures to fall, which would in turn reduce system load.
“Now, when you get high penetration levels of rooftop solar, there is a point in time when clouds come over and your [net] load is going to increase instead of decrease” because of reduced output from rooftop solar, Motley said.
A Caveat
Motley offered one caveat to that assessment: When daily temperatures average about 80 degrees Fahrenheit, temperature is still the main driver of the load forecast.
Under those conditions, air-conditioning load still drives enough electricity consumption that a cloud system causing a 10-degree drop in temperatures is going to reduce load.
Further complicating matters is humidity, which causes air conditioners to work harder and support load even under cloud cover. The situation is especially problematic in summer when monsoon moisture is thrown into the mix.
“You really have a question to ask yourself: Is my load going to increase because I am losing the rooftop solar, or is it going to decrease because I have a 10-degree temperature drop?” Motely said. “And we’ve seen both situations happen.”
Motley called APS the “most challenging load-forecasting region” within the EIM.
“It has a combination of a significant amount of rooftop solar, which is a driving factor, combined with some of those strong monsoon days in the summertime,” she said.
APS began transacting in the EIM last October, after the summer solar and monsoon peaks. But CAISO began running EIM load forecasting models ahead of the go-live date, giving operations staff an indication of what to expect this summer.
High Error Rates
So far, even outside the summer months, short-term load forecasts for the APS area are recording relatively high error rates compared with other EIM balancing areas (see chart). In November, the region’s hour-ahead forecast error rates reached nearly 2%, falling to 1.5% the following month. NV Energy has had similarly high error rates in the summer because of the prevalence of dust storms — a phenomenon that affects Arizona as well. The error calculations represent the average deviation between hour-ahead forecasted load and actual load.
The ISO’s goal is to keep error rates below 1%, Motley said, adding that such accuracy is not always attainable in some regions.
“If you have more rooftop solar, your accuracy is going to be worse because you now have another characteristic behind the scene that is influencing it,” Motley said.
She pointed out that short-term load forecasting is an important component for market optimization and reliability. It also is used as a key input for dispatch operation functions such as unit commitment, economic dispatch, fuel scheduling and generation and transmission maintenance.
EIM Governing Body member John Prescott wondered if there was a “nexus” between load forecasting errors and the high number of flexible ramping test failures observed in the EIM late last year — particularly in APS. (See EIM Sees Sharp Increase in Flexible Ramping Test Failures.)
“There are several factors that play into that and we have to isolate each one to see what’s driving it,” said Justin Thompson, director of resource operations and trading at APS. “But load forecast is one piece of it. Also, how well have [we] forecasted wind? … How well [have] we forecast the solar output?”
Phoenix Baseline?
Alyssa Koslow, a regulatory analyst at Salt River Project, said she had heard CAISO was using Phoenix as the baseline for forecasting for Arizona, despite the fact that APS’s territory extends into high-elevation areas.
Motley clarified that the ISO’s approach to forecasting is more comprehensive than that.
“We have multiple temperature stations within Arizona, and [the load forecast] is always driven by the temperature station that’s closest to where your load pattern moves the most,” Motley said. “So we work with all of the EIM entities on which station in which area moves the most for your load and then we incorporate that into the design.”
“One of the problems with models is ‘garbage in, garbage out,’” said Clay MacArthur of Deseret Power. “There’s a lot of behind-the-meter generation going on. How do you aggregate” the capacity?
Motley responded that the ISO takes a bottom-up approach that starts with the zip code and capacity for every interconnection on the distribution system. That information lays the foundations for system load forecasts for individual areas.
“And then we forecast the irradiance — which is essentially the amount of sunlight that’s going to come from the atmosphere to the roof for that resource — and we put that into the forecast as its own variable,” Motley said.
Neural Net
That last point is important for CAISO’s “neural net” forecasting method, which relies on the dynamic interplay between “highly interconnected processing elements” — the data fed into the model. As Motley explained, the neural net is modeled on the human brain and can synthesize copious amounts of information and “learn” to weight the importance of certain factors over others in their predictive processing.
“Storing the information by technology type is very important so that the neural net can have the correct connections,” Motely said. If that information gets “blended in with the rest of the model,” then the neural net has a difficult time distinguishing whether it was a change in temperature or solar output that caused load to move up or down.
CAISO continues to seek ways to improve its load-forecasting model, Motely said. Future improvements could include having EIM participants share their own load forecasts to provide comparisons, as well as having them provide balancing area information about demand response, hydroelectric behavior, rooftop solar and irrigation patterns.
“Can we fix everything? No, it’s forecasting — it’s good job security,” Motley joked. “But are there some things that we can fix? Yes, there are some things.”
