No abstract

Explainable Artificial Intelligence (XAI), i.e., the development of more transparent and interpretable AI models, has gained increased traction over the last few years. This is due to the fact that, in conjunction with their growth into powerful and ubiquitous tools, AI models exhibit one detrimential characteristic: a performance-transparency trade-off. This describes the fact that the more complex a model's inner workings, the less clear it is how its predictions or decisions were achieved. But, especially considering Machine Learning (ML) methods like Reinforcement Learning (RL) where the system learns autonomously, the necessity to understand the underlying reasoning for their decisions becomes apparent. Since, to the best of our knowledge, there exists no single work offering an overview of Explainable Reinforcement Learning (XRL) methods, this survey attempts to address this gap. We give a short summary of the problem, a definition of important terms, and offer a classification and assessment of current XRL methods. We found that a) the majority of XRL methods function by mimicking and simplifying a complex model instead of designing an inherently simple one, and b) XRL (and XAI) methods often neglect to consider the human side of the equation, not taking into account research from related fields like psychology or philosophy. Thus, an interdisciplinary effort is needed to adapt the generated explanations to a (non-expert) human user in order to effectively progress in the field of XRL and XAI in general.

Renewable energy sources provide an ever increasing amount of the global electric power generation. However, for many regions, even whole countries, the go-to primary renewable energy sources that are available in large quantities are wind and solar radiation, which are highly volatile since they depend on a factor that is not human-controllable: the weather. The traditional power grid featured centralized power generation and a hierarchical structure; but in addition to the volatility, renewable energy sources blur the distinction between generator and consumer: Through photovoltaic panels on rooftops, a consumer can alternate between becoming a generator and a consumer during any one day. Moreover, wind farms and larger photovoltaic power plants feed into the intermediate layer of the power grid, the distribution grid. Today, lower levels of the power grid feed back into the upper levels, voiding the traditional hierarchical structure of the power grid; power generation has also become distributed, just as the consumption of power. Due to technical restrictions that are inherent in the type of power plant, the traditional coal, oil, or nuclear power plants cannot accompany this volatility with arbitrarily changing their power generation.At the same time, the notion of the smart grid introduces a vast array of new data coming from sensors in the power grid, at wind farms, power plants, and consumers. The new wealth of information can help in managing the different actors in the power grid. This thesis proposes to view the outlined problem of power generation and distribution as a problem of information distribution and processing.To accommodate the new, decentralized architecture of the power grid, an equally decentralized approach to grid-wide information processing and distribution is sensible. Each power plant, substation, transformer, and large consumers, such as factories, become agents that exhibit proactive behavior i ii ABSTRACT and communicate to maintain the grid-wide power balance.Local forecasting is the basis for these entities. Every agent forecasts its future power balance or imbalance from historic data. The agents utilize individually trained Artificial Neural Networks to exhibit this forecast. The agent now seeks the help of other agents to solve this disequilibrium. The rules of this exchange are governed by a protocol designed in this thesis. The core principle of the rules that govern the information exchange is to arrive at a power equilibrium while being as scalable as possible without any agent having a priori knowledge of other agents. For this, the protocol remodels the power grid in the communication architecture to take advantage of the properties of the electric grid. Which agent contributes which part to the power equilibrium, however, remains a combinatorial problem. This thesis models the demand and supply of power in the Boolean domain. The power balance solver leverages Ternary Vector Lists and the XBOOLE system to master the emerging complexity. Thus, a distributed demand-suppl...

Multi-microgrids address the need for a resilient, sustainable, and cost-effective electricity supply by providing a coordinated operation of individual networks. Due to local generation, dynamic network topologies, and islanding capabilities of hosted microgrids or groups thereof, various new fault mitigation and optimization options emerge. However, with the great flexibility, new challenges such as complex failure modes that need to be considered for a resilient operation, appear. This work systematically reviews scheduling approaches which significantly influence the feasibility of mitigation options before a failure is encountered. An in-depth analysis of identified key contributions covers aspects such as the mathematical apparatus, failure models and validation to highlight the current methodical spectrum and to identify future perspectives. Despite the common optimization-based framework, a broad variety of scheduling approaches is revealed. However, none of the key contributions provides practical insights beyond lab validation and considerable effort is required until the approaches can show their full potential in practical implementations. It is expected that the great level of detail guides further research in improving and validating existing scheduling concepts as well as it, in the long run, aids engineers to choose the most suitable options regarding increasingly resilient power systems.

Unlocking and managing flexibility is an important contribution to the integration of renewable energy and an efficient and resilient operation of the power system. In this paper, we discuss how the potential of a fleet of battery-electric transportation vehicles can be used to provide frequency containment reserve. To this end, we first examine the use case in detail and then present the system designed to meet this challenge. We give an overview of the tasks and individual sub-components, consisting of (a) an artificial neural network to predict the availability of Automated Guided Vehicles (AGVs) day-ahead, (b) a heuristic approach to compute marketable flexibility, (c) a simulation to check the plausibility of flexibility schedules, (d) a multi-agent system to continuously monitor and control the AGVs and (e) the integration of fleet flexibility into a virtual power plant. We also present our approach to the economic analysis of this provision of a system-critical service in a logistical context characterised by high uncertainty and variability.