Analyzing controller conflicts in multimodal smart grids - framework design

Research focus

External events such as a short circuit in power line or change in wind speed force autonomous power system controllers to react, tune up against the events and move the system back to a stable state. These controllers will be modeled as self-organizing agents, thus reflecting the autonomous characteristics of the controllers they represent. They can modify their behavior during runtime for a more efficient performance as a wanted effect. Mutually and from undesired effects of emergence, the interaction of agents could cause an undesirable system behavior in form some unstable states.

Framework design

In this section, the methodology of work will be introduced. Using the technological basis of a multi agent system, an exemplary system will be modeled as adaptive multi-controller system combined with a smart energy system simulation framework. With the modeling approach of a self-organized multi-agent system, the system offers an abstraction layer regarding single controller types. This allows for an efficient automation of the scenario adaptation needed for an intense simulative scenario exploration, thus identifying the needed set of unstable system states.

The proposed project follows an artefact-oriented research method to ensure replicability and stepwise evaluation possibilities as has been proposed in (Nieße et al., 2014) and established research recommendations (Hevner & Ram, n.d.). For this reason, each research question is either connected to an artefact implemented to setup the simulation and software system needed to answer the research question, or results from analyses yielded from using this simulation and software system. These artefacts are shown in Fig. 1. In the following, the three layers of the planned research and simulation framework will be outlined briefly.

Layer of physical component simulation

As a basis for all later investigations, the physical system has to be defined and modeled. The general goal of this modeling is to set up the multimodal smart grid simulation as the object of observation. Implementing the idea of the smart grid architectural model (SGAM) 2nd IEEE PES International Conference and Exhibition on Innovative Smart Grid Technologies (ISGT Europe), 2011: 5–7 Dec. 2011, Manchester, United Kingdom (2011), the resulting simulation system can be understood as an instance of the component layer. In addition to the electrical system, different heating systems and gas networks have to be modeled as well as heat sinks and coupling points. Co-simulation tools will be used to integrate various simulation models from different domains.

Layer of observer/controller multi-agent system

With the Observer/Controller multi-agent system (see second lower layer in Fig. 1) an agent abstraction layer is added to the layer of physical component simulation to represent autonomous controllers and other energy system components as defined in layer 1. The resulting agent abstraction layer is needed for several purposes: First, this layer allows abstraction from the precise and specific view on the modeled energy systems. The abstraction method using agents thus allows to integrate the system model in the intense and automated simulation studies needed to identify unstable system states that result from controller conflicts (Ringler et al., 2016). Second, the system represents an agent-based control system as depicted in Fig. 1. With this purpose in mind, later mitigation actions for controller conflicts can be integrated in the system without changes in the overall design.

With the Observer/Controller approach, a generic architecture has been proposed for the development of self-organizing systems based on multi-agent systems. The main purpose of this architecture is to embed a self-organizing system in a control environment and thus allow for both keeping track of the actual performance of the System under Observation and Control and passing triggers to the system components (the agents) to control this performance. Different observation models can be activated during runtime if needed for the task to be solved by the System under Observation and Control. The general approach has been evaluated in different use cases (McCluskey et al., 2016; Müller-Schloer et al., 2011).

Simulation study layer

The simulation study layer (see the third layer in Fig. 1) is proposed for generating controller conflict scenarios and developing a metrics. In this layer two blocks work parallel and develop each other. The simulation study driver has two main functionalities, scenario generation and scenario modification. Unstable system states in multimodal energy systems are dynamic by definition. To analyze these states though, an approach is needed to deterministically reproduce these system states. The problem is twofold: First, a scenario exploration strategy has to be applied to identify severe (i.e. performance degraded) system states. Second, based on unstable states recognized in the first step, further examinations are necessary to develop metrics that can later be applied to help identify unstable states (RQ3).

Metrics driver

It has been already stated that a bad system performance does not necessarily hint to a controller conflict – the system might show bad performance independent of the current controller setup. Whereas in the first case an adaptation of the controller parameters might allow to avoid unstable systems states or lead the system to a stable operation again, the second one cannot be influenced. Parallel to the simulation study layer, the goal is to develop metrics as foundation for the identification of those system states that are classified as unstable caused by controller conflicts. One of the adjacent questions is if instabilities can be derived as global rules or depend on the current and preceding system states.

In (McCluskey et al., 2016), two different approaches are identified to distinguish a bad system performance resulting from undesired behavior and bad system performance that does not result from a “performance hole”. If the system layout allows separating the control system from the simulation system, different control concepts can be applied to a scenario that lead to an unstable system state. If this leads to a better performance, we can assume an undesired system adaptation.

Analysis of controller conflicts

With the presented concept, the multi-layer experimental framework will be set up and examination objects – i.e. deterministically generated scenarios prone to unstable system states – will have been identified. Metrics will have been developed to differentiate unstable system states that are a result of controller conflicts (undesired adaptation) from general bad system performance.

One method to analyze controller conflicts is to model the development of the system state from the simulated scenario over time. At some point in time, the system state proceeds from stable to unstable, so this transition can be regarded as a phase transition. With the generated metrics developed in the last section, a distinction function is given to discriminate if the set of system states belongs to the set of controller conflict system states or not. For this kind of systems, classifiers based on support vector decoders have been successfully applied (Tax & Robert, 1999). With the given complexity of interconnected energy systems though, choosing the appropriate set of system attributes to train the support vector is a challenging task. As has been shown for distributed energy scheduling, the quality of the qualifier substantially depends on choosing the appropriate training data (Bremer & Sonnenschein, 2013). Therefore, a structured approach has to be developed to evaluate different excerpts of simulation attributes for support vector training. The set of attributes will have to consist of both – attributes characterizing the scenario and attributes characterizing the current simulation state, i.e. the system state over time (Blank et al., 2011).