Efficient management of energy systems including storage systems

Abstract

In this master thesis, we are going to work towards a data based control algorithm, that given a complex system and a goal, it can learn how to manipulate the control actions of the system towards achieve that goal. The objective of developing such a control algorithm is to ensure the correct operation of a small smart microgrid with several components that balances the operational costs under optimal conditions and the economical performance of the system. Particularly, in this thesis, we will describe the components that conforms a standard microgrid including Energy Storage Systems (ESS) considering the efficiency and operational limits. Energy Generation Systems that relies on climatic condition cannot be on demand. Consumption hubs, like a regular household, whose consumption of electrical power depends on the daily habits can change from time to time. The composition of a microgrid of a "prosumer " - generally consumers of energy from the elec- trical grid with production and storage capabilities - usually relies in several types of ESS with complementary characteristics like a battery with higher storage capacity and a super capacitor with higher power density. The production must have at least one energy source, that is usually a renewable source, but the combination of several sources could increase the reliability of the system. At this level of management, the system can be considered as a set of ESS whose State of Charge (SOC) can be considered the working space and can be leveled with the control actions. This type of management is traditionally done with an economical criterion, whose optimization requires the definition of an accurate model of the system and the correct parametrization of the goals. We propose a different method that can be used without the previous knowledge of the system dynamics and can simultaneously generate control signals and learn the optimal policy given a parametrized cost function and the sensing of the system states. The proposed set of methods are called Structured Online Learning methods, and relies in two distinctive parts: the System Identification module, that will learn the dynamics of the system given the collected data as a linear combination of non-linear basis functions of the state; and a Value Function learning that will be updated given the learned model and can generate the control signal that is showed optimal in a long term window. As those systems require the definition of a differentiable and convex cost over the control effort, we introduced some assumptions over the operational cost that can adapt the economical problem to an equivalent of a Quadratic Regulator. This will also bring us the opportunity to compare our method with some standard solutions like the Ricatti equation for a Linear Quadratic Regulator. Comparing the standard solutions for the adapted problem that leverages the knowledge of the real model and several versions of the Structured Online Learning algorithm, we can confidently state that the proposed method generates comparable results in a reference tracking problem. But, it can even enhance the economical performance if the generation and consumption profiles has a certain degree of predictability