Using clustering algorithms to identify representative EV mobility profiles for complex energy system models

Abstract

The interaction between electric vehicles (EV) and the future energy system is subject of current research in the field of energy system analysis. EVs represent an additional electrical load on the one hand and a potential flexibility provider through smart charging on the other. Feedback effects on the energy system and potential benefits of intelligently charged EVs depend on a variety of technical parameters as well as the individual driving behavior of vehicle owners. Since no sufficient data on EV users driving behavior is currently available, synthetic profiles have to be used. In this paper we propose a methodological approach that combines the mobility data of the two main household travel surveys in Germany - the Mobility in Germany 2017 and the German Mobility panel - to synthesize annual mobility profiles that represent the German mobility behavior. To guarantee statistical soundness, the methodology requires a large number of individual profiles used for further evaluations. Computational power however limits the maximum number of usable profiles. In the context of this paper, we assess and compare potential revenues of a price optimized unidirectional and bidirectional charging strategy. Those evaluations are carried out for 10,000 profiles with the linear optimization model eFLAME. Resulting revenues and vehicle-specific indicators such as equivalent full cycles (EFC) and charging/discharging hours serve as a reference for further evaluations with a reduced number of profiles. To reduce that number, we compare two distinct methodological approaches. The first approach is based on randomly drawing an increasing number of profiles, while the second is based on applying various clustering algorithms to specifically identify representative profiles. In the context of clustering algorithms, we test and compare distinct feature definitions, preanalysis methods and include a principal component analysis (PCA) to identify the best cluster of representative profiles. To assess the validity of each approach, we use the deviation of 16 key indicators from the reference simulation run with 10,000 profiles. When considering randomly drawn profiles, we identified a minimum number of 1,000 profiles to adequately represent the German mobility behavior and keep deviations for all 16 key indicators low. The use of cluster algorithms can reduce this number even further. Even with a minimum of 10 identified representative profiles, deviations for most key indicators are comparatively low. Others on the other hand remain high.