Browsing by Author "Falk, Sebastian"

The idea underlying this thesis is to use data gathered by building management systems to build machine learning models in order to improve these systems. Our goal is to create models which can use data from multiple different sensors as its input and output some predictions about that data. We will then use these predictions when implementing new applications. At our disposal we have data gathered by both motion sensors as well as carbon dioxide (\ce{CO2}) sensors. This data is gathered at regular intervals, and will be in the form of time-series, after some transformations, which is the first topic we cover. We want to improve the systems to which these sensors are connected. For a concrete example we can consider the ventilation systems which control the air-conditioning. They usually have \ce{CO2} sensors connected to them. By keeping an eye on the \ce{CO2} value the system is able to adjust the air flow when the value becomes too high. The problem with this is that when that value is reached it takes some time before it is again lowered to a normal level. If we were able to predict when this value will begin to rise the system could increase the airflow beforehand, meaning that it can avoid reaching the threshold level. This improves the effectiveness of the system, making the air quality constantly stay at a comfortable level. Another example is the lighting control systems which commonly have some motion detection sensors which control the lights. A motion detection event occurs when one of these sensors sees some movement. Sensors are connected to one or multiple luminaires, turning the luminaires on when an event happens. The luminaires also turn off automatically after a set amount of time. Being able to predict when these events happen would make it possible to turn on the lights before a person actually walks into the room in question. The system would also be able to turn off the lights if it knows that no one will be in the room, which means that the lights will not be on unnecessarily. For creating these models we will be using multiple different prediction methods. In the thesis we will discuss some time-series forecasting models such as the autoregressive integrated moving average model as well as supervised learning algorithms. The supervised learning models we will cover are decision tree models, random forest models, feedforward neural network models as well as a recurrent neural network model called long short-term memory. We will explain how all of these models are created as well as how they can be used for time-series prediction on the data which we have at our disposal.