Wind is one of the cleanest and most sustainable ways to generate electricity, but the variable nature of wind makes it difficult to incorporate into the energy grid.
Imprecise wind forecasting systems keep wind power costs unnecessarily high and limit its growth as an energy source. But a Siebel Energy Institute-supported research project, Novel Physically-Driven Approaches for Multi-Scale Wind-Energy Forecasting, has made significant progress toward developing more reliable wind forecasting models.
Led by Elie Bou-Zeid, Ph.D., associate professor in the Department of Civil and Environmental Engineering at Princeton University, and postdoctoral researchers Seyed Hossein Hezaveh and Mostafa Momen, the project produced a hybrid of forecasting models that reduced large prediction errors by 40 percent. The team presented its findings in January 2017 at the annual meeting of the American Meteorological Society.
While the approach requires further testing and refinement, the results offer promise for the future of wind power forecasting—and the growth of this renewable energy source.
Bou-Zeid spoke with the Siebel Energy Institute about the project.
What benefits would better wind forecasting models provide?
Because of wind’s variability, wind power companies need backup generation capacity—usually from gas-powered turbines in the United States—to compensate for times when demand exceeds supply due to errors in forecasting of the latter. More accurate prediction models would enable wind suppliers to better match supply with demand, drive down costs, and enable the industry to grow. Studies show that, in a scenario where 20% of the grid’s electricity comes from wind, perfect forecasting could yield $1.6 to $4.1 billion in savings per year in the United States.
Lay people who are unfamiliar with the industry might assume that excess backup energy can be stored in batteries. Why doesn’t this happen?
The main barrier is cost. Batteries are very expensive. The technology really isn’t there yet to store those amounts of energy cost-effectively. In places where hydroelectric power can be used as backup—places like Oregon, Norway, or Sweden—you could use the excess energy to pump water up the damn and use it again. But we don’t have a lot of hydroelectric power in the United States.
Your project creates a new hybrid between two wind forecasting models—short-term data-driven models and mid-term numerical weather prediction models. What are these models?
Short-term forecasting extrapolates weather conditions in the near term—a couple of hours—based on conditions in the past few hours. It is based on a concept called persistence—that the wind in an hour will be close to what it is now. If you want to forecast the wind for tomorrow, you need to use another model—mid-term numerical weather prediction models.
The mid-term numerical weather prediction model is the one the National Weather Service uses. It takes into account atmospheric conditions in a given region, and what is happening at the boundaries of that region—factors like heat, airflow, and temperatures. For instance, what happens in New Jersey today could depend on what air masses are coming from Canada and the ocean. The mid-term model gives you the possibility of a long-term forecast, which the short-term model cannot provide.
Each model produces errors, however. In the short-term model, error greatly increases over time. In the mid-term model, the error you have now will be about the same as the error you have later.
When you plot these errors out on a graph, you see that, at about 2-4 hours, the errors in the short-term model cross the errors in the mid-term. Beyond that timeframe, errors in the short-term model exceed those of the mid-term model.
Most current approaches subjectively blend these models to produce a continuous forecast for up to 48 hours. What’s different about your project?
This field is still new. Thus far there hasn’t been an effort to do a more sophisticated blending of the two models. We wanted to create a dynamic hybrid in which each model is simultaneously used to inform the other. We’re trying to come up with a practical approach to blend the short- and mid-term forecasts to produce the best overall forecast.
With a dynamic approach, you look at how well a model performed over the last two hours and what errors it produced. Then you adjust the model based on what you’ve learned—to give you the best forecast in the future over the next 2-6 hours. The dynamic or learning model turns out to be quite useful.
How did you test these models?
We obtained observational data from 2012 from the CHLV2 station of the National Oceanic and Atmospheric Administration’s (NOAA) Buoy Center, which is located offshore from Virginia Beach, as well as from a long-established meteorological tower in the Cabauw, Netherlands. The data include wind speed and direction for every 10 minutes.
In one-month segments, we ran short-term and mid-term models to predict wind speeds in 2- and 6-hour increments and compared the models against the observational data. The mid-term model showed significant errors in predicting wind speed, while the short-term model gave more accurate predictions.
We wanted to see if we could further improve the short-term predictions, so we applied an ensemble of three models: a persistence model, a linear regression model, and a Mass Spring Damper model. Based on the performance of each method in a given time period, we selected the model with lower error rates for the next time period.
For the two-hour timeframe, this approach showed a 9% improvement in mean errors and an 18% improvement in large errors. For the six-hour timeframe, this approach produced a 17% improvement in mean errors and a 40% improvement in large errors.
What is equally interesting is that we are now also finding that the blending significantly improves the mid-term forecast for the day ahead (around 24 hours into the future), by about 10 to 20%.
What implications do these findings have for the wind power industry?
If you are a wind operator, you don’t care about small or average errors. You care about the large errors. This blended learning model reduced large errors significantly.
When wind comprises a small percentage of power in the grid, no one cares much about variability. When it grows, you care a lot. A colleague of mine at Princeton says that the current uncertainty in forecasting wind will limit wind power on the East Coast to just 30% of total power. By improving the forecast, we can increase the maximum percentage of wind power that can come from variable renewables.
What future directions will your research take?
We still need more observational data to additionally validate and improve the short-term models. We will run the mid-term model in real-time and compare the results to real-time observational data. Another key factor in running large wind farms is prediction of the wind direction. We will use the same techniques to provide short- to mid-term wind direction predictions.
We will use these findings to prepare a proposal for a full study to the U.S. Department of Energy. We hope to assess the challenges of applying this hybrid model in operational settings. In the current climate, we are uncertain about energy funding, but there’s still a lot we can do improve these models in realistic environments.
BY RAYMOND RIGOGLIOSO