I completed my PhD in June 2013. Here are some explanations of my research:
Note: This explanation was inspired by XKCD's Up-goer Five and the Science in ten hundred words project, where research is described using only the "ten-hundred" most common words in the English language.
Deciding better using many guesses for wind
Wind has become a big part of how we get our power. But it is hard to guess what wind is going to do tomorrow. And the other places that make power need to know ahead of time if they should make power or not. And we can't store power very well. So when we guess wrong about wind, we get stuck and have to pay more for power.
My work is on how to use many guesses for what might happen with wind tomorrow to decide better. We can guess better by looking at the past. Using many guesses takes much longer, but works better for some cases. It works best in the summer, when people in Texas use the most power.
Scheduling the power system with uncertain wind power
Wind and other renewables have become a significant source of our electricity. But wind can be difficult to forecast accurately and the existing power system was not designed for uncertain generation. In the power system today there is little or no energy storage and large generators must be scheduled well ahead of time. When there is an error in the wind forecast, this scheduling becomes less efficient. These small differences in efficiency can add up to millions of dollars annually for a large system with significant wind.
Instead of relying on a single, imperfect forecast, one solution is to consider many possible wind scenarios and choose the schedule that does the best overall. In power systems, this idea is called stochastic unit commitment (UC). My research focused on how these wind scenarios are created and the savings that resulted from the stochastic UC.
I used two methods to create scenarios: building statistical models of forecast error and mining the past for similar events. A process called moment matching synthesizes a set of wind forecast errors with statistics (moments) which are similar to a statistical model. With the similar events or analogs method, we search through a historical archive for forecasts that look like tomorrow's. Then we use the observed wind power on those analogous days as scenarios for tomorrow. In this work, simple analogs were tested. Simple analogs are based only on the forecasts for total wind power (rather than comparing forecasts based on many weather variables).
To evaluate how effective these scenario creation algorithms are, I created a model of the Texas power system, known as ERCOT (see interactive map). Wind power was scaled up in this model to simulate a future system with higher renewable energy penetration. Results were based on a nine month cost comparison between stochastic UC and current practice (UC based on a single forecast). Stochastic UC is computationally intensive, a 10 scenario run took about 40 hours to complete. This research benefited from time on the UW's Hyak cluster computer.
Results suggested that for the model of the Texas power system stochastic savings would be modest -- about $6M saved over nine months -- at 25% wind energy penetration (wind accounted for 8.6% of the energy produced in ERCOT in 2012). Savings from stochastic UC were highest during high demand periods -- hot afternoons in the summer when electricity prices are highest in Texas. However, 25% penetration appears to be a high point for savings. At 30% penetration, higher amounts of wind result in reduced energy prices during the highest demand periods, reducing the savings from running stochastic UC.
The choice of scenario creation algorithm and number of scenarios was important, significantly affecting both the savings and the run time. The moment matching method produced higher savings than the simple analogs method. Further improvements in both methods hold the promise of faster, lower cost scheduling.
Wind scenarios for stochastic energy scheduling
Large amounts of wind generation have been added to the power system in recent years. However, wind breaks many of the core assumptions in the process used to schedule energy and is particularly difficult to forecast accurately. Rather than scheduling based on a single forecast, stochastic unit commitment (UC) minimizes the expected cost over several wind scenarios for the next day. Stochastic UC is often held up as a solution to help alleviate the high costs related to uncertain renewables. Yet there is no widely accepted method for creating high quality stochastic scenarios. In this dissertation, we examine two wind power scenario creation methods – moment matching and analogs. Moment matching is a general technique where scenarios are synthesized to match a set of statistics or moments. We propose a method for estimating these desired moments based on historical wind data. The analogs method looks back in time to find similar forecasts and uses the matching observations from those analogous dates directly as scenarios. This work proposes and tests a simple analogs method based solely on aggregate wind power forecasts.
The performance of these methods is tested on a realistic model of the Electric Reliability Council of Texas (ERCOT) power system based on actual data from 2012. UC and dispatch simulations showed modest stochastic savings for the relatively flexible ERCOT model at 25% wind energy penetration. The scenario creation method and number of scenarios had a significant impact on these stochastic savings. Contrary to our hypothesis and the increase in perfect forecast savings, stochastic savings decreased as wind penetration increased to 30%. Stochastic savings are often largely due to a few high cost events during peak load periods; stochastic costs may be higher than deterministic for extended periods – generally when demand and marginal prices are low. Together these results paint a more nuanced picture of stochastic UC and provide a roadmap for future scenario creation research.