Smart Grids are complex systems which involve multiple domains such as electrical grids and the information and communication technology infrastructure. Especially the integration of DER into the distribution grid has led to an increased complexity of grid control with a multitude of intricate optimization problems to solve. In this thesis the motivating use case are regional ancillary services. Formerly ancillary services were provided by large generation units. The shift of generation into the distribution grids leads to the necessity to provide these services at least partly with DER. As part of ancillary services, congestion management in distribution grids will be the use case with the main focus. For congestion management, a set of DER and controllable loads have to coordinate their schedules to achieve a necessary adjustment in power generation or consumption. The distribution of this deviation from the planned power output among the various units is a reactive scheduling task. The results obtained by the investigation of congestion management may be applicable to other regional ancillary services that can be performed by reactive scheduling, like voltage control.
The provision of such services with a large number of distributed energy resources and controllable loads leads to an increased complexity for the control mechanisms. This rise in complexity can be handled through the use of distributed coordination approaches. A natural way to implement these distributed algorithms are Multi-Agent Systems (MAS). MAS are a form of distributed artificial intelligence. They consist of a number of autonomous agents. An intelligent agent is a computer system which can sense its environment and take autonomous actions to accomplish its individual goals. Furthermore it is capable of social interaction with other agents or even humans. Agents can react to changes in their environment as well as to information received from social interaction. But they are also able to act in a pro-active manner. Global system behavior emerges through the interaction of agents who aim to achieve potentially contradictory individual goals (Wooldridge 2009).
In literature a vast number of possible applications for MAS in smart grids can be found: For example microgrid operation in (Pipattanasomporn et al. 2009), load management in (Amini et al. 2013) or demand side management in (Ramchurn et al. 2011).
Figure 1 shows how a MAS could be realized in a smart grid setting. The way the agents are situated in the grid is only one possibility. It shows the case where each agent is located directly on-premise, close to the DER or controllable loads they belong to. Here the agents communicate with each other via the internet. There are many degrees of possible centralization and decentralization of the agents. The alternative design concept that differs the most from the one shown in Fig. 1 would be complete centralization of the agents, for example in a data center.
The location of the agents in a central facility revokes some of the major benefits of the distributed algorithm: The robustness of the system is reduced through the reintroduction of a single point of failure. Furthermore the scalability is decreased.
Another concern are privacy issues. Especially when small DER and loads that belong to private households are included. It is likely possible to derive private data like behavioral patterns from the sets of feasible schedules agents need for the optimization process. When the agents are hosted centrally this data has to be transferred entirely. When the agents are hosted on-premise only a small fraction of this information is communicated to the outside.
This thesis focuses on cases where the agents are either completely or at least to a great extent distributed. The reasons for this limitation are the advantages this approach offers that are especially relevant for the considered use case, namely robustness, scalability and privacy.
In either case the communication of the agents is characterized by certain properties.
Talbi (Talbi 2009) identifies four design questions:
-
Exchange content (Which information is exchanged?)
-
Exchange criterion (When is the information exchanged?)
-
Exchange topology (Between which entities is the information exchanged?)
-
Integration policy (How is the new information handled after the exchange?)
In this work, the primary focus lies on the exchange topology. The exchange topology determines which agents communicate directly. It is often modeled as an overlay network. Only the entities that are direct neighbors in the overlay network are exchanging information during the execution of the algorithm. The exchange topology is therefore part of the algorithm itself and effects the (runtime) behavior of the algorithm profoundly. Talbi refers to this type of algorithms as algorithmic-level parallel heuristics (ALP-heuristics).
This is the reason why the exchange topology must be designed with care, as it has a great impact on both, the output of optimization and runtime characteristics, in particular the speed of convergence, robustness to failures and communication efficiency (Baras et al. 2009).
The construction of the communication topology should lead to an appropriate degree of connectedness. Fully connected graphs are not scalable enough, whereas very sparse graphs will lead to a slow convergence speed. Furthermore certain graph topologies can result in undesirable behavior. For example, individual agents can exert greater influence than others (Baras et al. 2009). In (Nieße et al. 2017) the authors stated the communication topology as a possible cause for the convergence to local optima. This is another example for the affects the communication topology can have on the results of the algorithms.
The creation of optimal overlay networks presents a challenge that can be solved with a variety of approaches. In (Elias 2009), the author gives a broad overview of works from different contexts, which address the problem with methods from completely different areas, which range from game theory to using heuristics like simulated annealing. Furthermore, graph theory plays an important role in analyzing and modeling of overlay networks. One example is the theoretical framework for the analysis of consensus algorithms in (Olfati-Saber et al. 2007).