Low-dimensional representation of monthly electricity demand profiles. (March 2023)

• Record Type: Journal Article
• Authors: Luque, Joaquin; Personal, Enrique; Perez, Francisco; Romero-Ternero, MCarmen; Leon, Carlos
• Is Part Of: Engineering applications of artificial intelligence. Volume 119(2023)
• Publication Date: 2023-03
• Subjects: Electricity demand profile -- Dimensionality reduction -- Graphic representation -- Profile clustering -- Profile labeling; Engineering -- Data processing -- Periodicals; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Ingénierie -- Informatique -- Périodiques; Intelligence artificielle -- Périodiques; Systèmes experts (Informatique) -- Périodiques; Artificial intelligence; Engineering -- Data processing; Expert systems (Computer science); Periodicals; 620.00285
• DOI: 10.1016/j.engappai.2022.105728
• Languages: English
• ISSNs: 0952-1976

Abstract

Abstract: This paper addresses the problem of reducing the number of values required to characterize an electricity demand profile, which is usually known as its dimensionality. This reduction may have a significant impact on the computational efforts and storage capacities required to analyze and process high volumes of electricity load curves. Also, the reduction to 2 or even 1 component enables its graphic representation. Specifically, this work is mainly focused on profiles defined by their monthly demand values, and where the clients are aggregated by locations and/or economic activities. This approach is of great interest for marketing analysis and decision-making of electricity retailers. In this sense, the use of dimensionality reduction techniques based on knowledge (calendar and temperature) along with the application of data-driven procedures (Principal Component Analysis and autoencoders), are explored in the paper. The results of this research show that autoencoders clearly outperform the other techniques, yielding errors in the reduction process between 15% to 40% lower and preserving distances between profiles in the low-dimensional spaces, with a correlation of 0.93 with the distances in high dimensional space. Additionally, the bidimensional graphical representation of a profile can easily be interpreted in a polar way, where the angle denotes the shape of the profile, and the radius reveals its scale. To reach these results, a very large dataset has been employed, with about half a million aggregated profiles corresponding to the electricity consumption during 3 years of more than 27 million clients in Spain. Graphical abstract: Highlights: Autoencoders are the best option to reduce the dimensionality of demand profiles. Data-driven techniques outperform knowledge-based models by reducing components. Low-dimensional characterization preserves the relative distances between profiles. The graphic representation of bidimensional profiles admits a polar interpretation.