Data-driven decision-making under uncertainty in power systems

In this dissertation, we construct decision-making frameworks and decision-support tools that seek to aid power system operators (SOs) in delivering an economical and reliable power system operation under uncertainty. The dissertation comprises nine publications, each of which lays out an optimization framework under uncertainty or a decision-support tool, on which SOs can capitalize in ensuring a reliable power system operation around the clock. The deepening penetration of renewables greatly exacerbates the uncertainty and variability of net load, that is, system load minus renewable generation. In scheduling the short-term operation of power systems, SOs must explicitly take into account the uncertainty and variability in net load. To that end, SOs solve two key problems that are at the cornerstone of the short- term scheduling of power systems: unit commitment (UC) and economic dispatch. These problems do not only play a central role in vertically integrated utilities, but also form the basis of market optimization problems employed to clear the day-ahead and real-time markets across the globe. Given that in the U.S. alone SOs solve these problems to clear over a $400 billion electricity market annually [1], improvements in UC decision quality could unlock billions of dollars’ worth of savings every year. To take UC decisions under uncertain net load, most existing studies in the literature employ a two-stage stochastic optimization framework, which requires committing to a single, pre-specified distribution to represent the probability distribution of net load. However, since the true, underlying distribution of net load cannot be known in practice, such studies are typically plagued by a poor out-of-sample performance. Further, these studies fail to draw on information on covariates, which refer to contextual information (such as the day of the week, temperature, or lagged observations of net load) that may directly influence the materialized net load levels. Although net load, or other uncertain parameters typically encountered in short-term optimization problems such as demand, market-clearing prices, or renewable generation, are typically influenced by various covariates, classical two-stage stochastic optimization frameworks largely use a one-size-fits-all representation of uncertainty. In this dissertation, we address these shortcomings by working out alternative optimization frameworks under uncertainty. Two of the publications presented herein discuss a distributionally robust optimization (DRO) framework for the UC problem. A major advantage of DRO frameworks is that it immunizes the UC decisions against using an inaccurate probability distribution for net load. The framework we propose relies on forming the nominal distribution using directly real-world data in conjunction with the k-means clustering algorithm and different distance metrics. Subsequently, an ambiguity set of probability distributions is constructed around the nominal distribution using the Kullback- Leibler divergence. In two publications, we demonstrate the application of this framework for different problems (one for the UC problem under uncertain net load, the other for a price-taker microgrid under the uncertainty in both net load and market-clearing prices). The results provide valuable insights into the influence of degree of conservatism and the dataset size on the in-sample and out-of-sample performances. Motivated by the actionable insights provided by covariates, we further discuss in three publications of this dissertation contextual stochastic optimization (CSO) approaches to the UC problem. In these publications, we leverage directly data in conjunction with different machine learning algorithms to prescribe UC decisions by taking explicit account of covariate observations. The UC frameworks proposed in these publications present a significant contribution to the state of the art. These publications are among the first to take a CSO approach to UC, and they present novel methods to construct prescriptions using three ML algorithms, which have to date not been used in any CSO framework. These studies further propose important innovations that help improve decision quality, which are empirically borne out by several numerical studies conducted using real-world data. In addition to the above risk-neutral decision-making approaches, we further propose a risk-averse self-scheduling framework for storage resources bidding in decentralized electricity markets under uncertain market-clearing prices. Among the many contributions of this study is a novel explanation for the seemingly nonintuitive storage dispatch decisions encountered in practice, wherein storage resources fail to discharge at markedly low prices or discharge at significantly high prices. We lay out how risk can serve as an underlying factor for such decisions, a concomitant of which is that such decisions could well be rational and persist—signifying that storage resources may indeed forgo price spikes and not dispatch during periods of scarcity, during which the system is in dire need. In another study, we approach the short-term scheduling problem from a sustainability perspective, and we discuss a UC approach tailored to microgrids, which takes explicit account of greenhouse gas emission amounts and its monetary implications. The results bring home how different carbon tax rate values can influence the scheduling decisions of a microgrid. In the last two publications of this dissertation, we present decision-support tools that seek to aid SOs in coping with the several challenges brought by the deepening penetration of renewables. One such tool focuses on forecasting the start time and magnitude of net load ramps, so that SOs can capitalize on the forecasts to ensure that the system has sufficient rampable capacity to meet steep ramping needs that may eventuate. In the last publication presented in this dissertation, we discuss a prediction model that seeks to exploit the underlying factors that influence the frequency response characteristic of a system so as to forecast the power system frequency in the next minute. This decision-support tool is geared toward imparting extra time to SOs in the event of a large frequency excursion, during which SOs typically race against the time to bring the frequency toward its nominal value. In all nine publications laid out in this cumulative dissertation, we use real-world data to demonstrate the application of the proposed decision-making frameworks and decision-support tools, and we compare our frameworks and tools with benchmark methods proposed in the literature or used in practice based on several evaluation metrics. Extensive simulation results and comparative assessments set out in each publication serve to make evident the effectiveness of each proposed framework and tool.

Author dc:creator

• Yurdakul, Ogün

Advisor dc:contributor.advisor

• Albayrak, Sahin

Licence dc:rights.uri

• http://rightsstatements.org/vocab/InC/1.0/

Language dc:language.iso

en

Identifier URI

https://doi.org/10.14279/depositonce-19543

OAI identifier oai:identifier

oai:depositonce.tu-berlin.de:11303/20745