2.4 Operational Probabilistic Tools for Solar Uncertainty

Monday, 7 January 2019: 9:45 AM
North 129A (Phoenix Convention Center - West and North Buildings)
Aidan Tuohy, Electric Power Research Institute, Palo Alto, CA; and D. B. Kirk-Davidoff

Handout (2.6 MB)

Electrical power systems must continuously satisfy demand for electricity that varies over a wide range each day. Decisions must be made in advance to allocate resources among different generation technologies, with ample lead time allowing more efficient operation. But decisions must be made under uncertainty about the demand to be met, and about the supply of electricity that will be available from each potential source. Meteorologists have long been able to supply probabilistic forecasts (where forecasts of the level of electrical demand or generation are given for several quantiles of the probability of exceeding a given value). However, uptake of these probabilistic forecasts by most end users has been slow.

This presentation will describe ongoing work for a DOE-funded research project, as part of the Solar Forecasting II program, to improve the usefulness of probabilistic forecasting in power system operations. The focus of the project is developing and demonstrating new methods to operate power systems with high penetrations of solar power. These methods will use probabilistic solar power forecasts, which capture the uncertainty inherent in solar power output, to support decision making in utility operations. Using data from three utility regions – Duke Energy (Carolinas region), Southern Company (entire system), and Hawaiian Electric (Oahu system), the outcome will be a platform that enables new power system operating strategies.

Probabilistic solar power forecasts will be developed for time scales from days to hours ahead for distributed and central solar PV. Improvements will focus on better capturing the uncertainty for the specific underlying weather conditions. This will be based on applying machine learning methods to ensembles of solar irradiance and power forecasts, to better clarify the reliability and usefulness of the full distribution associated with solar power uncertainty. Machine learning methods will be applied to improve the fit and narrow the width of the distribution, resulting in an improved skill scores. We will focus on skill scores like ignorance, that measure of the quality of the forecasting scheme in terms of the expected returns that would be obtained by placing bets proportional to the forecast probabilities. This is expected to be a useful metric to assess whether probabilistic forecast improvements can lead to improved outcomes.

Probabilistic forecasts will be used as input to advanced grid operations simulation tools, tailored to specific operations of each utility studied, to investigate the economic and/or reliability benefits of different uses of the data. Advanced simulation tools that replicate system operations in detail will allow for a more accurate and realistic understanding of the use and value of probabilistic forecasts. By comparing results from a base case using deterministic methods to improved operating practices using probabilistic forecasts, the value of using this information will be explored. Multiple use cases for probabilistic information, including stochastic optimization, interval programming, robust optimization and dynamic reserve determination, will be applied to each utility system. Some of these directly apply probabilistic information to decision making, while others will use the probabilistic information to inform a deterministic decision-making process. An example of the latter is the use of dynamic operating reserve requirements, where the reserves are dynamically sized to reflect uncertainty in the forecasts, as opposed to current practice of basing reserves on historical system conditions. As well as examining the value of using probabilistic information, the simulation framework will also be applied to show the value of the improvements made to the probabilistic forecasts in the forecasting workstream of the project.

Based on results from the simulation phase, a scheduling management platform (SMP) will be demonstrated for each of the three utilities, to show the feasibility of the proposed methods in real world operations. The integrated approach, tested across multiple real world systems with varying degrees of solar and different existing operational practices, will demonstrate the full value of using new operational methods based on probabilistic forecasts. By testing and demonstrating a range of different potential use cases, the most beneficial methods, balancing economic efficiency, reliability, and ease of use will be identified for each system. Performing these studies across multiple systems provides for a range of benefits to be shown and thus increases the likelihood of adoption beyond the utilities studied here.

This presentation provides an update on current progress from the first year of the project. Initial results and planned efforts will be described, focusing on the methods that are being used for improvement of probabilistic forecasting methods, initial analysis of the use cases to be studied and plans for future deployment of these forecasts in utility operations. Feedback from the participating utilities will be summarized in terms of how end users think about the use of probabilistic forecasts.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner