As power systems around the world have integrated wind and solar power, a crucial first step is to develop or procure a short-term forecasting system to support operations. Right now, anyone with significant penetration of variable energy resources (VER) that we’re aware of has some kind of a forecasting system for these resources. There is a great deal of ongoing research to improve VER forecasts, whether through the foundational weather models, the application of machine learning or data-mining algorithms, or the increased use of new and expanded sensor networks.
That said, even with the best forecasts in the world, there will always be variability and uncertainty associated with VER. This increases the system’s need for flexibility, which must come from traditional technologies (e.g., gas turbines, hydro), emerging technologies (e.g., storage and demand side), new transmission, or from the wind or solar itself. It could also come from advanced scheduling techniques. Operators currently procure reserves to manage short-term variability and uncertainty; for example, frequency regulation is used within a dispatch interval and newer flexible ramping products are used between dispatch intervals (e.g., in the Midcontinent Independent System Operator (MISO) and California Independent System Operator (CAISO)). While still a small percentage of the total production costs, as wind and solar penetration increases, the need for such reserves may also increase, increasingly making these costs evident.
New Tool to Forecast Reserve Requirements
A method we’ve been working on at the Electric Power Research Institute uses all of the information that is available to system operators to predict reserve needs based on existing and anticipated conditions. That is, you can forecast the operating reserve requirement, just as you would forecast the load or renewable production levels. The main concept is to use readily available information (e.g., time of day, load/renewable forecasts), together with historical reserve needs and risk tolerances, to predict the reserves or additional flexibility that would be required in the next hours or days. Many existing reserve products, like contingency reserve, are typically static and remain constant regardless of system conditions, but the requirements we have examined are dynamic and change through time as new information becomes available. This idea was developed into a software tool called Dynamic Assessment and Determination of Operating Reserves, or DynADOR, that has been tested with some balancing areas across the U.S. and elsewhere. The output of the tool provides a forecast for the reserve requirements (MW quantities) that is tailored to a specific system.
Simultaneously improving costs and reliability. We have studied the use of dynamic reserve methods with several utilities and system operators and shown simultaneous benefits in terms of reduced operating costs (e.g., variable fuel costs, startup costs) and improved reliability (lower risk, reduced area control error and reserve shortages). This “win-win” may seem counter-intuitive, since one expects a trade-off between costs and reliability. However, we have observed that it is possible to improve both. For large portions of the year, reserve requirements can be reduced compared to the traditional procurement because the traditional route is based on rules of thumb that try to cover the worst-case event in all hours and then, for those periods identified to be higher risk, the reserve requirements are often increased with respect to the rule of thumb to avoid load balance violations. And another counter-intuitive condition has been observed: we have seen that increasing your reserve requirements in critical risky periods allows for flexible capacity to be procured in advance, rather than leaving it up to expensive last-minute corrections. So — spread the word, increasing your reserve requirements may lead to lower costs, if done effectively.
Ongoing Improvements and Testing: Machine Learning, Simulations, Probabilistic Forecasting
There are a few more steps to be taken before these types of methods are more widely used, including more testing and improvement in the forecasting methods. For example, last year we examined the benefits of machine learning approaches, which are attracting increasing interest. Several utilities and system operators now use similar dynamic reserve methods in planning studies. By demonstrating the concept in simulations, operators can become more familiar with the method, which will help bring the relevant stakeholders on board in order to make the enhancements. Some system operators, particularly those in systems with high renewable penetrations, are starting to use dynamic reserves in operations, providing useful first lessons. Additional improvements to the forecasting methods for reserve, just like those for renewable forecasting, will lead to greater adoption and greater confidence from the operators that may end up using these methods.
The research community is already working on the next steps and more advanced measures of utilizing renewable forecast information. One example is the utilization of probabilistic forecasts. Instead of just utilizing deterministic “point” forecasts of wind/solar/load, reserve requirements can be predicted based on the uncertainty spreads observed in a probabilistic forecast. Some U.S. system operators and European transmission system operators are already experimenting with this.
The final frontier in this area is to integrate probabilistic information more completely, not just in setting or predicting reserve requirements, but in supporting commitment and dispatch decisions and/or other operational decisions (interchange schedules, gas purchases, outage planning, etc.). Here, advanced programming methods such as stochastic unit commitment are promising and have theoretical advantages. A number of practical barriers currently hinder their adoption and warrant further research, including a need to reformulate existing tools, determine how to handle large quantities of information, grapple with complicated economic outcomes, and address computational limitations. But moving toward these methods will enable system operators to fully consider the uncertainty inherent in wind and solar resources and the load when making decisions, allowing for a smoother integration of these resources.
Discussion to Continue in ESIG’s Meteorology and Market Design Workshop in June
In sum, many years of academic research are now starting to pay off in the real world. We’re seeing applications of all the discussion on probabilistic forecasting that we’ve heard about for many years at the ESIG Forecasting Workshop, and we on the power systems side are (finally) starting to apply the forecasting methods we’ve heard so much about to our domain. We’ll continue to discuss these and related topics at the newly named ESIG Meteorology and Market Design workshop in Denver in June — we hope to see you there to continue to figure out how to improve system operations.
I wish to thank the additional EPRI contributors to this blog, including Robin Hytowitz, Nikita Singhal, Miguel Ortega-Vazquez, and Erik Ela.
Aidan Tuohy, PhD
Principal Project Manager
Grid Operations and Planning, Electric Power Research Institute