JultikaUniversity of Oulu repository

Abstract

Indoor air quality is important for our health and well-being. It has been proven that the air quality affects the performance of the workers. One way to achieve better air quality, would be to adjust the air conditioning and heating. By predicting the room conditions one can react faster to changes and ensure that the room conditions stay favourable.

The existing machine learning (ML) models that are used for CO2 prediction are rather basic. This study aims to improve upon the performance of the models compared to earlier studies. The focus of this thesis is to study what type of model would give the best results, what type of training data should be used, and how long history should be fed into the model.

One of the goals of this thesis is to examine whether a deep neural network is better than a wider one. The used data consists of indoor air measurements from nine variables: CO2, temperature, pressure, illuminance, volatile organic compound (VOC), movement detection, humidity, door state. The data was gathered in VTT’s Oulu office (in Finland). Different combinations of input variables are experimented on, to find out, which inputs should be fed into the network. The performance is compared to other models commonly used in prior studies. These models include previous value forward (PPV), line fit, and a Multilayer perceptron with one hidden layer (MLP1). Several hyperparameters are tested to find out which combination of parameters has the lowest error.

Compared to earlier studies, the developed deep Multilayer Perceptron (MLP) model improved root mean squared error (RMSE) by 0,997 ppm. This indicates that deep models perform better for CO2 prediction tasks. The total root mean squared error was 6.07 ppm. This improvement makes it possible to give more accurate readings for the air conditioning control system, which in turn makes it easier to keep CO2 levels low. A history length of seven minutes is used as the input, and the model predicts ten minutes ahead.