Energy Management System for Smart Buildings and Microgrids Using Sampling-Based Model Predictive Control (SBMPC) and Machine Learning

As the cost of renewable energy resources decreases and environmental concerns, such as global warming, arise, new ways of generating, storing, and using clean energy are being encouraged by governments and organizations around the world. Due to the growing energy demand, the modernization of the power grid has become an immediate priority that is leading to the decentralization of power systems, the creation of transactive energy markets, and the integration of distributed energy resources (DER). Traditionally, bulk power generation is based on a unidirectional structure, where electricity is generated on monolithic fossil fuel power plants and is then delivered to customers through different stages of transmission and distribution systems. This classical framework has many disadvantages, including, but not limited to, high inefficiencies and power losses in the system over lengthy transmission and distribution lines, a central point of failure of the system, and the emission of greenhouse gases that contribute to polluting our planet. That is why, progressively, governments and companies are starting to heavily invest in the development of new ways of integrating renewable power generation and DER systems into the modern power grid. The integration and control of these systems are considered essential problems that need to be tackled in order to build the future smart grid and advance universal electrification efforts. However, these systems are still not ready for a harmonious integration to the main electrical grid due to problems related to their intermittent nature and lack of proper control. Additionally, most of these systems still have a high price tag when compared to traditional power generation, and that is why their integration is not seen as a good investment for regular consumers and even large utility companies. Many researchers around the world are currently working on developing methods and frameworks that could allow the harmonious integration of renewable energy systems and other distributed energy resources with the main electrical grid. Nonetheless, since this problem does not have a trivial solution, sophisticated control methods and techniques need to be designed and developed to control DER systems and advance the decentralization of power. This dissertation is focused on the development of novel methods and algorithms, with particular emphasis on energy management applications that can facilitate the integration of DER systems to the modern power grid. This dissertation begins with an in-depth literature review of state-of-the-art energy management controllers designed to control available DERs in microgrids and buildings connected to modern distribution systems. Here, we aim to identify some of the key knowledge gaps present in this area and propose solutions designed to tackle these problems. The second chapter presents an overall review of microgrid systems and all the major components it is composed of. Additionally, it focuses on reviewing the state-of-the-art technologies currently being used and researched in the field of energy management of buildings and microgrids, with a detailed explanation of all the types of control and features these systems could include, like energy management and renewable generation forecasting. In this chapter, we explore some of the current solutions available for energy management systems and present the reason why our approach is being proposed as a cost-effective solution to the energy management problem. After this, the next chapters concentrate on explaining the fundamental theory behind the individual modules required by the proposed control method while proving its applicability in a real-world scenario. The modules developed can be described as follows: 1. A load and generation forecasting module based on a novel short-term PV power forecasting model and a neural network load forecasting model. 2. An energy management controller module based on Sampling-Based Model Predictive Control (SBMPC). The performance of this module is improved using models based on deep reinforcement learning. 3. A controller hardware-in-the-loop (CHIL) testbed designed to test the proposed model in a real-time environment. The development of the forecasting module is the main topic presented in the third chapter of this dissertation. This chapter focuses on exploring the theory behind the development of a novel short-term PV power generation forecasting architecture that significantly improves the performance of PV power forecasting by using a combination of the stationary wavelet transform (SWT), long-short term memory (LSTM) networks, and deep neural networks. The proposed forecasting model exhibits significant performance improvements in metrics such as forecast skill scores and the mean absolute percentage error (MAPE) when compared to state-of-the-art models used in short-term PV power forecasting applications. The fourth chapter of this dissertation is concerned with describing the theoretical background and concepts behind the proposed energy management solution. Here, we present an alternative graph-search formulation of the energy management problem and adapt Sampling-Based Model Predictive Control (SBMPC) to solve the formulated optimization problem. Moreover, chapter 5 considers the use of a deep reinforcement learning agent based on an asynchronous advantage actor-critic (A3C) model to enhance the performance of the formulated solution via learning a dynamic sampling process. Finally, chapter 6 focuses on the development of a real-time controller hardware-in-the-loop (CHIL) testbed designed to evaluate and test the proposed solution in a real-time simulated environment. In this chapter, interfacing modules between the external physical controller and the real-time system are developed using standard communication protocols such as TCP/IP and DNP3. In summary, the proposed solutions presented in this manuscript are aimed to improve methods related to the optimal control of DER systems under varying price schemes by leveraging the use of novel control and machine learning techniques. The main objective of this dissertation is to provide a detailed view of the theory behind the development of a proposed energy management solution that aims to minimize the operating cost for consumers and facilitates the integration of distributed energy resources (DER) into our modern power grid.