Application of artificial intelligence techniques to the monitoring and study of energy in buildings

  1. Martínez Comesaña, Miguel
Dirixida por:
  1. Javier Martínez Torres Director
  2. Pablo Eguía Oller Director

Universidade de defensa: Universidade de Vigo

Fecha de defensa: 13 de xullo de 2023

Tribunal:
  1. Elena Arce Fariña Presidente/a
  2. Jorge Morán González Secretario
  3. Ofélia Maria Serralha Vogal
Departamento:
  1. Enxeñaría mecánica, máquinas e motores térmicos e fluídos

Tipo: Tese

Resumo

In this work different tools, based on artificial intelligence (AI) and optimisation algorithms, are implemented and compared with the aim of analysing and improving the energy efficiency of buildings. The first step has been the use of different machine (ML) and deep learning (DL) models (also known as black box models) in order to replicate the thermal behaviour of buildings showing some improvements over existing methods such as simulations or continuous monitoring. In this way, AI models demonstrate their usefulness in predicting the thermal demands and indoor temperatures of a building reducing the monitoring requirements. These techniques are faster and require less previous knowledge about the specific case study to obtain accurate results. Furthermore, it has been shown that the thermal inertia inside the buildings is significant and an important input for model training. In the first published article, the influence of the thermal inertia has been analysed introducing time lags in the model training and observing that considering a certain number of time-lags for the indoor variables its performance improves. The estimations generated by AI models that replicate the thermal behaviour of buildings allows the subsequent estimation of coefficients such as HLC (Heat Loss Coefficient). Specifically, HLC characterises the building envelope and give a numerical value to the thermal efficiency of a building. In the second published paper, it is demonstrated that this coefficient can be accurately estimated using the estimates provided by an AI model. With the appearance of the COVID 19 pandemic, and with it the consequent state of alarm and confinement, the interest in controlling the indoor environmental conditions within in-use buildings arose. Thus, the direction of this research, maintaining its main objective of analysing and contributing to improve the energy efficiency of buildings, focused on this area. In this way, based on a low-cost monitoring installation, in the third published article a DL model is built and trained to interpolate data directly relate with the indoor environmental quality (IEQ) throughout a large space. The monitoring installation is composed of eight wall-mounted devices and a mobile device that are employed as independent variable for DL model training. The variables considered to summarise IEQ were indoor temperature, relative humidity and CO2 levels. Moreover, the implementation of AI model allows that, once the model is trained, the mobile device is removed and no human intervention is needed. Therefore, the DL model uses the collected values by the wall-mounted devices and accurately interpolates the mentioned variables to any place of the analysed space. Taking into account the performance of AI models interpolating IEQ throughout large spaces, in the fourth published article an IoT (Internet of Things) system is created and combined with a trained AI model. Therefore, all the monitoring devices connected between them and with a trained AI model that generates estimations continuously, allows to see, measure and control the indoor environmental conditions inside a building in real time. Moreover, based on the installed IoT platform, an optimisation study was developed in the fifth published article. Using a multi-objective genetic algorithm, the minimum number of monitoring devices required to train a black box model that accurately estimates the indoor environmental conditions was efficiently analysed. Since the number of wall-mounted devices used in training affects the performance of the AI model, the optimal values of their internal hyperparameters vary with the number of devices. In order to use the best possible architecture for the DL models, the optimisation of the number of monitoring devices was performed simultaneously with the optimisation of the model hyperparameters. As a result of this analysis, it was demonstrated that only considering three specific wall-mounted devices (reducing them into five), the AI model is able of perform accurate interpolations. This fact shows the contribution of this latest study with a tool to optimise monitoring installations and reduce the monitoring cost maintaining an accurate model that is able to generate interpolations in real time.