Modeling and optimization using machine learning techniques in buildings with renewable production, storage and electrical consumption

Abstract

Neutralizing the impact of greenhouse gasses requires major changes in the following years. The negative effect caused by the emissions related to the building sector is important due to the diverse need for the use of energy in all types of activities realized in this sector. Furthermore, trends towards electrification and increased use of interconnected devices indicate that these energy needs are growing. Thus, it is of vital importance to develop techniques to improve energy efficiency in this sector and thus respond to mitigation measures for sustainable development. The improvement of energy efficiency in the building sector requires the study and implementation of techniques that understand the patterns of electricity consumption and renewable energy. By recognizing the energy needs at the user level and understanding the production of the generation sources in situ, it is possible to carry out efficient energy management. With the implementation of a battery, an effective management system is capable of evaluating the most appropriate actions, reducing losses and increasing the coefficient of autarky. The main objective of this thesis is the development of a predictive methodology for efficient energy management in buildings. The process is divided in two phases. In the first one, the most robust techniques for modeling and predict the electricity consumption and photovoltaic production are recognized. In the second, an optimization of the resource is carried out from the implementation of a battery. By means of the electrical tariffs, the forecast of the electrical consumption and the photovoltaic production, the energy efficiency of the building is increased while obtaining an economic benefit. The techniques used to achieve the main objective are validated from the publication of four research articles published in peer-reviewed scientific journals. In the main body of this thesis, the research articles related to the prediction of photovoltaic production, publication one, and electricity consumption, publication two, are firstly presented. Because of the need to work efficiently with large amounts of data that is constantly increasing, the effectiveness of different methods of machine learning and deep learning is studied. These techniques also allow continuous learning and quick adaptation to new circumstances. After exposing the content of the first phase, the research articles related to the energy management in buildings are exhibited. In the first publication of this series, shown as publication three, a predictive optimization method of electricity consumption is carried out. From the application of a battery and using mixed integer linear programming, a load shifting is sought to obtain a demand flattening. In the latest publication, expressed as publication four, a management system is presented. The forecast of electricity consumption and photovoltaic production are considered. By optimizing the storage system using deep reinforcement learning, the aim is to maximize the autarky coefficient and minimize the environmental impact of building consumption. The studies presented in the first phase faithfully follow photovoltaic production and electricity consumption. The reliability of these models guarantees the possibility of predictive management. The energy trade-off carried out through the application of a storage system shows that the models used meet the initial objective. Thus, studies show an increase in energy efficiency and a reduction in greenhouse gas emissions in the buildings analyzed.