Simulation-Based Evolutionary Dynamic and Stochastic Optimization for Smart Electric Power Systems

Stephan Hutterer

Abstract

The electric power systems research society early identied the necessity of optimization both for planning and operation tasks, where formulations such as the optimal power ow (OPF) problem shape this research domain ever since. At the same time, technological changes to electric power grids challenge new methods, requiring optimization in both dynamic as well as uncertain systems. Additionally, optimization tasks for the control of numerous distributed devices fundamentally require scalability aspects.

In this context, heuristic optimization methods have evolved as being capable of managing many of those upcoming needs. Simulation optimization with metaheuristics provides a promising fundament for optimization under uncertainty, and oers the basic approach for handling manifold challenging optimization issues within this work. While various aspects of future power grid optimization tasks are being analyzed, simulation-based methods both for static but mainly for dynamic problems are developed. Further on, simulation-based evolutionary policy function approximation is being discussed for dynamic power ow control problems, which is presented both in a generic manner as well as tailored to the electric engineering domain.

In a rst experimental part, the developed methods are applied both to a static probabilistic planning problem and to a dynamic OPF control task within benchmark systems. Showing that approximate optimal policy-based control yields competitive results compared to reference solutions, it additionally is able to make quick and robust control actions within dynamic and stochastic environments, being scalable to numerous devices.

The second experimental part treats optimal load control issues in the smart electric grids context. Electric vehicle (EV) charging control is defined as a generic problem for optimal load control over time. Existing works in the literature are being analyzed while major lacks can be identified that need to be tackled. Here, once more simulation-based evolutionary policy function approximation comes into play, which is shown to evolve control policies for a holistic electric vehicle charging problem that unifies all important requirements for smart EV charging. Finally, policies are evolved that enable intelligent charging decisions to multiple EVs within an experimental system. These decisions are even able to satisfy system-wide goals while being exible to dynamic and uncertain conditions. Comparisons to deterministically optimized charging decisions in a test simulation show its validity.

Finally, a scalable technology has been presented, which enables approximate optimal control in volatile as well as uncertain systems. Representing a technology for managing dynamic stochastic optimal power ow problems makes it highly promising for future smart electric grid control.

Kurzfassung

Optimierung stellt einen wichtigen Bestandteil in Betrieb und Planung moderner Stromnetze dar. Zeitgleich trägt der aktuelle Wandel der Versorgungsparadigmen elektrischer Energie - hin zu so genannten Smart Grids - neue Anforderungen an Optimierungsanwendungen mit sich: Optimale Entscheidungen müssen zunehmend in dynamischen und unsicheren Systemen getroffen werden. Gleichzeitig verlangt die Steuerung einer Vielzahl verteilter Endgeräte nach Skalierbarkeit der eingesetzten Methoden.

In diesem Kontext gewinnen heuristische Optimierungsverfahren an Bedeutung. Speziell mit simulationsbasierter Optimierung können Probleme in komplexen und unsicheren Szenarien bearbeitet werden, wodurch das methodische Fundament dieser Arbeit begründet wird. Mittels Analyse aktueller und zuküunftiger technologischer Herausforderungen in Stromnetzen werden sowohl statische als auch dynamische Optimierungsprobleme formuliert, behandelt und analysiert. Darauf aufbauend wird ein Ansatz der simulationsbasierten evolutionären Approximation von so genannten policies entwickelt, welcher sowohl generisch beschrieben, aber speziell für dynamische und stochastische Last ussprobleme entwickelt und validiert wird. Speziell diese Verquickung von evolutionärer Simulationsoptimierung und der policybasierten dynamischen Optimierung für reellwertige Entscheidungsprobleme bietet einen neuartigen methodischen Ansatz.

PhD thesis, Johannes Kepler University Linz, November 2013