Representative day selection in capacity expansion modeling : an accelerated energy transition for Texas

Long-term electricity planning models with high temporal granularity are computationally expensive, but models that study high penetrations of variable renewable energy source(s) (VREs) in electric grids require the inclusion of a certain level of temporal detail. Studying the tradeoffs between computational intensity and model specificity is a growing area of research. In particular, a key question in electricity modeling is how to maintain reliability and accuracy in CEMs that study high VRE penetrations, while also abiding by reasonable computational limits. Modelers have used a variety of approaches to reduce computational burden in planning models, while also ensuring that model accuracy and reliability is preserved. One common simplification involves selecting representative days or "time slices” to represent a grid’s behavior for an entire year. This thesis uses a generation-level capacity expansion model built with R’s linear optimization programming library to compare the differences in different time slice selection methods’ projection of an accelerated energy transition. The Electric Reliability Council of Texas (ERCOT), which serves about 90% of Texas’s electric load, is used as a virtual testbed to test and demonstrate the efficacy of this method. The research compares how future capacity mixes differ when selecting days using (1) seasonal averaged days with a peak day, (2) functional boxplots, and (3) k-means clustering. The study finds that model results are particularly sensitive to assumed solar and wind availability, which differs depending on hour of day and day of year. Under a strict emissions constraint when higher demand gaps must be satisfied by VREs, the selection methods are more sensitive to input assumptions such as the way that wind and solar availability is accounted for, and to the weight that is assigned to each representative day. Tradeoffs exist in each of the three methods. The methods have similar run times when the number of time slices are held constant, but while methods that study seven 24-hour periods (7-day methods) take an average of 10 minutes each to run for a time horizon of 2018 to 2030, methods that study thirteen 24-hour periods (13-day methods) take an average of 47 minutes to run on the same time horizon. While the functional boxplot and k-means clustering methods are useful for their ability to highlight extreme events such as instances of high load-low availability, they may overestimate the impact of events such as this. While the seasonal average methods are not very sensitive to adding additional representative days, clustering and functional boxplots do yield significantly different results when more days are added, especially later in the time horizon when the model is forced to meet a higher demand gap under a lower emissions constraint, and when the system retires some capacity and makes more endogeneous investment choices. Clustering has a similar output to seasonal methods in the early part of the time horizon. The clustering methods most closely match ERCOT's reported load duration curve, but a drawback of clustering compared to the other two methods is that the peak day is not represented. The seasonal averaging methods most closely matches ERCOT's annual reported data in 2018, but this speaks more to this method's ability to replicate pure dispatch than to the precision of its capacity expansion results. More research is needed to verify which method is most precise compared to a capacity expansion model that runs for all 8760 hours.