Improving smart charging for electric vehicle fleets by integrating battery and prediction models

Frendo, Oliver

PDF
OliverFrendoDissertation.pdf - Veröffentlichte Version
Download (11MB)

URL:	https://madoc.bib.uni-mannheim.de/58770
URN:	urn:nbn:de:bsz:180-madoc-587704
Dokumenttyp:	Dissertation
Erscheinungsjahr:	2021
Ort der Veröffentlichung:	Mannheim
Hochschule:	Universität Mannheim
Gutachter:	Stuckenschmidt, Heiner
Datum der mündl. Prüfung:	25 Januar 2021
Sprache der Veröffentlichung:	Englisch
Einrichtung:	Fakultät für Wirtschaftsinformatik und Wirtschaftsmathematik > Practical Computer Science II: Artificial Intelligence (Stuckenschmidt 2009-)
Fachgebiet:	004 Informatik
Freie Schlagwörter (Englisch):	electric vehicles , smart charging , optimization , regression , real time systems
Abstract:	With increasing electrification of vehicle fleets there is a rising demand for the effective use of charging infrastructure. Existing charging infrastructures are limited by undersized connection lines and a lack of charging stations. Upgrades require significant financial investment, time and effort. Smart charging represents an approach to making the most of existing charging infrastructure while satisfying charging needs. Smart charging involves scheduling for electric vehicles (EVs). In other words, smart charging approaches decide which EV may charge at which charging station and at which current during which time periods. Planning flexibility is determined by the length of stay and the available electrical supply. First, we present an approach for smart charging combining day-ahead planning with real-time planning. For day-ahead planning, we use a mixed integer programming model to compute optimal schedules while making use of information available ahead of time. We then describe a schedule guided heuristic which adapts precomputed schedules in real-time. Second, we address uncertainty in smart charging. For example, EV departure times are an important component in prioritization but are uncertain ahead of time. We use a regression model trained on historical data to predict EV departure times. We integrate predictions directly in the smart charging heuristic used in the first approach. Experimental results show a more accurate EV departure time leads to a more accurate EV prioritization and a higher amount of delivered energy. Third, we present two approaches which allow the smart charging heuristic to take EV charging behavior into account. In practice, EVs charge using nonlinear charge profiles where power declines towards the end of each charging process. There is thus a gap between the scheduled power and the actual charging power if nonlinear charge profiles are not taken into account. The first approach uses a traditional equivalent circuit model (ECM) to model EV charging behavior but in practice is limited by the availability of battery parameters. The second approach relies on a regression model trained on historical data to directly predict EV charging profiles. In each of the two approaches, the model of the EV's charging profile is directly integrated into the smart charging heuristic which allows the heuristic to produce more accurate charge plans. Experimental results show EVs charge significantly more energy because the charging infrastructure is used more effectively. Finally, we present an open source package containing the smart charging heuristic and describe results from applying the heuristic in a one-year field test. Experimental results from the field test show EVs at six charging stations can be scheduled for charging when the grid connection only allows two EVs to charge concurrently. Runtime measurements demonstrate the heuristic is applicable in real time and scales to large fleet sizes.