Sponsored Advertising Feature
There’s a lot riding on the complex load forecasts electric utilities use to schedule the dispatch of their power plants. A mismatch between predicted and actual load can cost thousands or even hundreds of thousands of dollars.
Associated Electric Cooperative Inc., which provides power for its six generation and transmission members in Missouri, Iowa and Oklahoma, has a roughly 5,000 MW system. If a load forecast turns out to be even one percentage point off from actual load, it could translate into the equivalent of running a mid-sized coal plant.
About six years ago, AECI changed how it handles its load forecasts, and it has saved the cooperative “quite a bit of money,” Kirk Clark, supervisor for power marketing at AECI, said. “It makes us much more reliable too. We are hitting our peak loads more closely.”
Prior to that, AECI was predicting its load internally, using similar day forecasts. Back then, its mean average percentage error between forecast load and actual load was about 8.5%, Clark says. Since AECI began using external forecasting services from Pattern Recognition Technologies (PRT), their error rate has dropped below 4 percent.
PRT provides load, price, wind and solar forecasts based on machine learning algorithms.
Clark says the change to PRT was well worth it. One of the factors that prompted the change, he says, is that more wind power was showing up on AECI’s system. And compensating for the variability of wind power can magnify errors in forecasting. “You have to dispatch more power than you want,” Clark says.
AECI’s business has also changed in other ways. Seven or eight years ago, most of its power purchases were done through bilateral deals. “Now 90 percent of our business is done with RTOs,” Clark said. AECI now engages in power transactions in the PJM Interconnection, the Midcontinent ISO and the Southwest Power Pool.
In addition to local load forecasts, AECI also buys forecasts for the loads and indicative pricing in those RTOs from PRT. “It helps us make better market decisions,” Clark says. “It helps us marry up our load decisions and our dispatch decisions.”
Like many other aspects of the electric power industry, load forecasting has changed over the past decade or so, as has the need for more accurate forecasts. When utilities used to rely on the steady output of central station fossil fuel plants, predicting load was a more straightforward exercise. Now, there is greater and growing penetration of intermittent generation sources such as wind and solar power, which increases the volatility of load and prices. In the past, utilities used to primarily trade power bilaterally. Now, they are taking part in competitive markets where prices can be influenced by an even wider array of factors, from cheap natural gas in PJM to abundant wind power in MISO. Those changes can bring opportunities and present challenges. They also call for new tools.
As they become more deeply immersed in this new environment, many utilities have moved away from the days when load forecasts were done manually and have turned to software that not only can automate forecasts but can learn and update forecasts on the fly. Algorithms and machine learning are replacing the time-consuming process of manually updating forecasts throughout the day as conditions change.
A case study from PRT details how Missouri public power utility Independence Power & Light could lower its forecast error rate. By switching to a PRT forecasting product, the utility cut its error rate down from 8 to 6 percent in two weeks. Within a year of making the change, the error rate was down to within 4 percent, resulting in annual savings of over $100,000 for the city.
It is increasingly important for utilities to be able to look at load in a more nuanced way. Residential solar power, for instance, not only adds generation to a system, it can also reduce load. That requires a new approach to load forecasting.
“The problem with solar is that it takes out the load from behind the meter” and that is difficult to see, Jeff House, short term load planning program manager at the Tennessee Valley Authority, says. The best way to approach that problem, he says, is to do separate calculations for irradiance and cloud cover. If it is cloudy, there is going to be more load and less behind the meter generation, House says.
However, that calculation is slightly different for winter and summer. A cloudy summer day can mean less behind the meter generation but less load because it would not be as hot. In winter, a cloudy day would mean more load and less behind the meter generation.
“It is a dance,” House says. “You don’t want to be short of generation, but you also don’t want to generate too much. You don’t want to be swimming in megawatts.”
Ninety percent of the time, he says, “computers nail it.”
Accurate forecasts are a key part of that dance. For more than 12 years, TVA has been using PRT’s forecasting services and has seen its error rate shrink. “We are between 2 and 2.25 percent. We want to avoid getting to 3 percent. We haven’t seen that in a decade,” House says.
“If we had a perfect weather forecast, we could get to 1 or 1.5 percent,” House estimates, but “in real life, even with a good forecast, it is hard to get under 2 percent.”
In another case study, a public power utility in the Electric Reliability Council of Texas area was able to use PRT’s forecasting products to identify days when ERCOT would be likely to overestimate demand. The utility saved its customers money when PRT was within 600 MW of actual demand on a day when ERCOT was over 2,000 MW higher.
In the quickly changing and ever more complex electric power sector, if a utility does not have an accurate load forecast, “you are behind the eight ball,” Clark at AECI says. “You need those services to be an effective utility.”
For more information about PRT and its products, click here.