Die Elektromobilität mit teils hohen Ladeleistungen ist für den sicheren Betrieb der elektrischen Verteilnetze zukünftig eine Herausforderung. Zur Reduzierung von Überlastungen in der Niederspannung werden daher Steueralgorithmen benötigt, um die Ladeleistung der Fahrzeuge zu steuern. Hierbei ergibt sich allerdings das Problem, dass die Niederspannungsnetze in der Regel messtechnisch nicht überwacht werden und so Eingangsdaten für Steueralgorithmen fehlen. In der Arbeit wird die Kombination von zwei maschinellen Lernalgorithmen untersucht. Die Steuerung der Ladeleistung von Elektrofahrzeugen ist als Markow-Entscheidungsprozess definiert, der mittels dem bestärkten Lernen gelöst wird. Für die Bereitstellung der Eingangsdaten wird ein künstliches neuronales Netz verwendet, das den Zustand eines Niederspannungsnetzes abschätzt. Durch das Zusammenspiel beider Algorithmen können die durch die Ladung von Elektrofahrzeugen ausgelösten Netzüberlastungen reduziert werden.
The expansion of renewable energies and the charging of electric vehicles, in some cases with high power, pose a new challenge for the safe operation of electrical distribution grids in the future and, in some cases, today. Therefore, algorithms for controlling the charging power of electric vehicles are needed to prevent and reduce low-voltage equipment overloads. However, when implementing a control algorithm for the charging power of electric vehicles, the fact that low-voltage grids are not typically monitored by measurement technology, due to historical reasons, presents a problem. Therefore, it is necessary to be able to precisely estimate the condition of the low-voltage grid. This thesis presents an autonomous control of the charging power of distributed private electric vehicles to prevent and reduce equipment overloads in the low-voltage grid while maintaining a short charging time. This control utilizes a Markov decision process. Machine learning algorithms were used to solve the Markov decision process and generate its state. An artificial neural network was used to estimate the node voltage in real time and as a state for the Markov decision process. To solve the Markov decision process, reinforcement learning in a decentralized approach as a multiagent system was used as a machine learning algorithm. Each charging point was assigned a so-called agent that, using a defined reward function and action vector for controlling the charging power of the respective charging point, attempted to achieve the optimal balance between reducing equipment overload in the low-voltage grid and minimizing the charging time of electric vehicles. The estimation of node voltage and the autonomous decentralized control of electric vehicles using machine learning algorithms were validated and analyzed in three different scenarios of a grid model. This thesis investigates how the estimation of the node voltage as a state of the Markov decision process, which is subject to inaccuracies, affects the efficacy of the autonomous control. It also explores whether increasing the percentage of electric vehicles without expanding primary resources using the system of linear learning algorithms described is possible.
Nutzung und Vervielfältigung:
Alle Rechte vorbehalten