Taming the Curse of Dimensionality in the Generation Expansion Planning Problem

Abstract

Integrating a significant amount of generating capacity from intermittent renewable energy sources (IRES) requires a change in the long-term generation investment plans. The variability and stochasticity of these sources mean that it is essential to consider not only the cost of meeting the annual demand for electrical energy but also the hourly or even sub-hourly changes in operating conditions in the generation expansion plan (GEP). However, introducing these operational constraints dramatically increases the dimension of the GEP problem and the computational burden. Selecting a small set of the most representative profiles makes it possible to consider the operational constraints in GEP models within a reasonable computing time. In this work, the application of feature engineering and machine learning in this area of research has been revisited to design a rigorous algorithm for systematically selecting representative profiles from a one-year horizon net load series. A new metric to evaluate the representative profiles has also been proposed. Further, the long-term impact of basing GEP on the selected sample data using this new metric and other metrics in the literature is investigated. In addition, sensitivity tests regarding the size of the sample data and different penetration levels of IRES are carried out.