A novel artificial intelligence based energy management system for microgrids

Abstract

In many countries, including our own, large amounts of electrical power are generated where the energy source is located, while it is consumed in areas with large industries and populations. This distance between energy generation and consumption leads to the transmission of energy, which results in the waste of energy as heat and increases energy costs. Microgrids have emerged as a solution to energy use by applying the principle of energy generation and consumption at the same place. Microgrids are small-scale electrical grids that can use distributed energy resources in conjunction with conventional grids. They can combine solar panels or plants, wind turbines, energy storage systems, generators, and the utility grid. This reduces energy loss during transmission, improves energy efficiency, and allows energy to be used efficiently. In addition, microgrids that operate in small settlements such as university campuses, military facilities, towns, or neighborhoods can work in "island mode" without a connection to the utility grid when needed. Many microgrids are currently operated using classical control methods and operate in certain size that has only been determined using optimization methods. This limits the efficiency that can be achieved during the operation of the microgrid and makes it difficult to follow new trends in energy storage technologies. The crux and significance of this thesis revolves around the notion that contemporary energy storage technologies can be utilized efficiently within the system, and that the existing artificial intelligence technology can serve as the foundation of the microgrid energy management system. The energy management system designed in this structure reduces energy waste, lowers costs, improves efficiency, and improves grid stability, while also producing effective solutions for energy demand by controlling the use of various sources together. Moreover, this energy management system contributes to reducing carbon emissions while allowing for the easy adaptation of new technologies. In light of all these advantages, this thesis presents an artificial intelligence-based energy management system design for microgrids. To further explain the concept of artificial intelligence, it encompasses machine learning algorithms as a subset, while machine learning includes deep learning algorithms and concepts. In this thesis, microgrid applications of various sizes and properties are examined, and a microgrid simulation model was created at commonly used sizes. This simulation model assumed a microgrid applied to a university campus, with a solar power plant and wind turbines serving as renewable energy sources. The energy management system being designed predicts the power that these sources will generate, using the up-to-date prediction algorithms within artificial intelligence. When designing, the focus is initially on predicting the power that solar and wind turbines will generate, using five years of meteorological data collected at five-minute intervals. The meteorological dataset, consisting of nine different data types, has undergone a series of data pre-processing. Missing data is filled in accordance with the characteristics of the dataset, and outliers are removed. The characteristics of this dataset were analyzed with different graphs and their suitability for training was examined. The labeled data consisting of the generation values at the same region and at the same time/minute intervals were added to the meteorological data set that was deemed suitable for training. Seven prediction models were developed using four prevalent machine learning methods and three novel algorithms based on the gradient boosting machine to predict the power generated by the solar power plant. These prediction models were trained separately using the training dataset made suitable for training. The results obtained from these seven prediction models were presented in both graphical and tabular formats. In addition to comparing which algorithm gave how successful results for this study, the computation costs were also compared. The designed energy management system must also predict the power generated from wind turbines. In this regard, prediction models were created using three different machine learning algorithms, and the results were obtained. These prediction models were compared using various performance metrics. This study conducted within this thesis, which achieved successful results, offers new approaches and unique results to the literature on the prediction of the power generation of renewable energy sources. An artificial intelligence-based energy management system should provide not only energy efficiency but also low energy costs and profitability for the user. The widespread use of dynamic electricity pricing should also be considered, which is determined based on the relationship between countrywide generation and consumption level. In this thesis, it is assumed that the microgrid simulation model developed is located in a country where dynamic pricing is applied. A five-year dataset was created from actual dynamic pricing data obtained from open-source platforms and analyzed. The dataset was examined, preprocessed, and made ready for the training of prediction models. Four deep learning algorithms with memory cell structures were selected for this study. Using these algorithms and the training dataset, price prediction models were developed, and the results were obtained. The learning performances, error values, and accuracies of the models were presented comparatively. These innovative prediction models were integrated into the designed energy management system. Knowing the power demand from a microgrid makes operational decisions more appropriate and robust. The load demand at which time of the day is an important parameter. Knowing the load demand in advance affects decisions regarding resource utilization. Considering this fact, the energy management system designed should also be able to predict load demand. To this end, load demand prediction models were developed using four deep learning methods with memory cell structures similar to price prediction. Actual load values obtained from open sources were scaled according to the simulation model of the microgrid created. Deep learning models were trained using the five-year load dataset, and the results were obtained. The results were presented comparatively using many performance metrics. As a result of this study, successful prediction models were developed and integrated into the designed energy management system. An artificial intelligence-based energy management system uses many prediction models described above. The theoretical and mathematical foundations of all machine learning and deep learning methods used are provided in the second chapter of this thesis. The energy management system described requires an additional controller to manage the microgrid in addition to human management. In this context, this thesis proposes another artificial intelligence-based controller. Data-driven control methods that have replaced classical control methods are popular topics nowadays. This thesis focuses on machine learning-based control methods of this type. In this context, reinforcement learning, which is one of the three main branches of machine learning, is investigated and its foundations are given. Reinforcement learning is the general name for methods based on the principle of controlling the system without the need for a mathematical model of the system. It is possible to separate this concept into methods based on table creation and methods using deep neural networks. In this thesis, controller agents using both types of methods are created. The agent, which will learn to control the system in reinforcement learning, needs to optimize itself. This optimization process is done through trial and error. For the agent to be able to take the best version through these trials, the system it will control, which is a microgrid environment model in this thesis, must have specific characteristics. Five different control agents were designed specifically for the energy management system, three of which were temporal-difference-based and two were deep reinforcement learning-based. Three environment models designed specifically for the microgrid are proposed in this thesis to enable these agents to train themselves. These environment models with unique reward strategies present a new approach to the literature. These environment models that use renewable energy sources, load demand, and dynamic prices for the training of agents have shown quite successful results in terms of energy management. The trained reinforcement learning agents have learned to manage the microgrid and offer considerable profitability to the user. The energy management system whose design steps are explained in this thesis uses many different artificial intelligence algorithms. These artificial intelligence models created, trained, and successful results achieved have been consolidated under a single graphical interface in this thesis. A unique graphical interface has been designed, and all prediction models and control agents have been integrated into this design. This interface design, which consists of seven pages in total, offers many variables and control actions related to the microgrid to the user. The user can see the powers that will be generated for the future, load demand, and the price. In addition, the user can apply many control actions to the microgrid through this interface. The user, who can also see many real-time parameters, can analyze the performance of prediction models and control agents through relevant pages. In conclusion, this thesis proposes an artificial intelligence-based energy management system that contains many current and innovative algorithms for microgrids and uses them uniquely. Artificial intelligence-based prediction models determine the decisions that an artificial intelligence-based control agent will make. This agent, which learns to select the correct control actions for the microgrid, presents the determined control action to the user through the designed interface. Additionally, the originally designed energy management system interface allows the user to see many parameters related to the microgrid in advance. This thesis proposes an energy management system that contributes to the literature with its original approach and can be used in real-world applications.