Versen, Dennis Salvador

In deutschen Stromnetzen ist die Vorhersage zur Vermeidung von Überlast eine anspruchsvolle Aufgabe. Eine wenig digitalisierte Infrastruktur sowie eine strenge Datenschutzvorgabe erschweren die Vorhersage von Lasten in Netz und damit das präzise Reagieren auf Überlast. In diesem Beitrag wird ein intelligentes Last- und Lademanagement für die Integration der Elektromobilität vorgestellt, welches auf neuronalen Netzen basiert, und Lastspitzen in einem deutschen Niederspannungsnetz reduzieren kann. Nach einem prädiktiven Regel-Ansatz wird eine neuronale Netzarchitektur als Systemidentifikator dienen und dazu genutzt werden den richtigen Leistungspegel für das Laden von Elektrofahrzeugen zu bestimmen und die Last somit intelligent im Niederspannungsnetz zu verteilen.

In recent years, electric vehicles (EVs) are considered to be a promising way to reduce greenhouse gas emissions from the transportation sector. However, the increasing penetration of EVs into the distribution network (DN) raises serious concerns about the network’s safe and reliable operation. The uncontrolled EV charging with random behavior will lead to volatile load peaks on the distribution transformer. In order to obtain more transformer loading capacity available for integration of further EVs, distributed energy resources (DERs) and related devices, such as heat pumps, the transformer loading must be limited to a certain range. For this reason, an intelligent charging management based on model-free Reinforcement Learning (RL) is proposed in this work. The RL management is able to control the charging power of all EVs connected to the network without previous knowledge about the arriving- and leaving time. The needed information for the RL-agent to perceive the current state of the system is formed with cumulated values such as the total energy requirement and the total charging power demand of all EVs. In this paper, the RL algorithm is trained on real-world energy consumption data for a month and on a reference network, created with selected characteristics of a substation network area in the northeast of Hamburg. Comparing with uncontrolled charging, the simulation results show that the RL-based charging management avoids 99 % of threshold violations regarding transformer loading and results in 1% of EV energy requirement is not satisfied. Through sensitivity analysis regarding the state space representation in the employed RL process, the necessity of providing the state of charge (SOC) or the energy requirements of EV users are proven to improve the charging control performance.