On the Clustering of Demand and Weather Data for Electricity Generation Expansion Planning

Abstract

This research investigates the performance of clustering algorithms for Electricity Generation Expansion Planning (GEP) with a focus on electricity demand and weather data. The primary objective is to identify the most effective algorithm for selecting representative periods that accurately reflect the variability in energy demand and renewable energy supply, influenced by weather conditions. The study examines the impact of different clustering algorithms, including Agglomerative Hierarchical, K-means, and K-medoids, across various settings by adjusting the number of representative periods and shifting the slice of the data. Through rigorous a priori and a posteriori analyses, the research evaluates the algorithms' performance in replicating the statistical characteristics and GEP decisions of the full dataset. It also assesses the impact of data slicing methods on the clustering outcomes. The results indicate that K-medoids algorithm stands out for its consistent accuracy in replicating the full dataset measures, making it the best performing algorithm in both a priori and a posteriori evaluation. This algorithm excels at capturing the essential statistical characteristics of the dataset, such as variance and correlations between demand and renewable energy outputs. However, it is notably susceptible to shifts in data slicing, which can significantly influence its performance. Shifts in the data slice often lead to variations in the algorithm's output, which highlights the delicate balance between data representation and the accuracy of clustering outcomes. The findings also highlight the dependency between initial data spread, which is a data characteristic, and the shifting effect on clustering. It is shown that a data possessing higher spread indicator will experience high shifting effect on both a priori and a posteriori measures. This study not only contributes to the theoretical understanding of clustering in GEP but also offers practical insights for energy policy and system design, emphasizing the critical role of accurate data representation in optimizing energy planning and operations. Furthermore, this study sheds light on a topic often neglected in common practices, showing that slice shifting might have considerable effect on outputs if the historical data possess a high spread indicator.