Normally, when PJM presents information about load forecasting, it’s discussed in terms of seasonal or yearly.
Just as important, however, is the daily load forecast, which brings the broader picture into focus and is critical for the daily operation of the grid.
PJM posts forecasts for seven days, including the current day, for 22 zones in addition to a few aggregated areas over the entire footprint. There are a number of facets to forecasting the daily load, according to PJM meteorologist Elizabeth Anastasio, Dispatch.
“We look at the calendar, as well as weather – both measurements and forecasts – and historical load,” she said. “For temporal factors, it could be the time of the year or day of the week. Holidays play a big part, too. Thanksgiving Day, for example, is like no other day of the year. For weather, we look at factors like cloud cover, wind and humidity, in addition to the temperatures.”
PJM operators have eight load forecast models, as well as a number of tools that help them visualize weather data, identify trends and highlight anomalies. Many of these models, including the basic one used on average days, are neural network models.
Neural networks are statistical learning models, inspired by biological neural networks (such as the central nervous system and the brain). These networks have systems of interconnected “neurons,” which send messages to each other.
“In certain situations, some of our non-traditional models may perform better,” said Anastasio. “There are specific models for specific situations – for drastic changes in weather or extreme temperatures, significant effects from humidity or wind. For holidays, our own tool, the PJM Similar Day Lookup, often performs the best.”
The Similar Day Lookup algorithm suggests a number of historical days from which the PJM operator can choose. The operator then plots the models and makes an informed decision to either raise or lower loads. The dispatchers’ experience and knowledge adds value to the process and makes for a better forecast, she said.
PJM dispatchers can update forecasts for any day, but will typically focus on the intraday and day-ahead forecasts. In addition, PJM dispatchers can update forecasts at any time. The forecasts are captured and posted online twice an hour – quarter-of and quarter-after the hour.
Refining the load forecast models and monitoring weather forecast performance is part of PJM’s continuous improvement goal. This includes participating in meetings with industry load forecasters hosted by the Northeast Power Coordinating Council. PJM has also implemented a solar forecast and has set up a load forecast analysis team.
“We’re working on improving the process by combining three weather vendors’ forecasts into one ‘smart’ mix to decrease the chances of weather-forecast-related errors,” said Anastasio. “We also have new neural net models that use as input the temperature forecast plus a couple of degrees and minus a couple of degrees, which provides a confidence interval.”