dc.description.abstract | Reduction of energy consumption in buildings is a desired goal for many people. This goal may be
achieved if sufficient planing of building construction in early design phase is done. In this project
is the accuracy of 8 different machine learning models tested.
The purpose of this study was to analyze which machine learning model predict most accurate results of heating demand and electricity consumption in two case buildings. Each model was analyzed
based on input variables, hyper-parameters settings, and how the accuracy of predictions was compared to previous studies. Two case buildings in eastern part of Norway is used for data analysis and
testing of machine learning models. For building 1 it is provided data about hourly heating demand
and for building 2 it is provided data about hourly electricity consumption. Data mining are done
on both dataset before they are used to predict energy consumption of the two buildings. Machine
learning models that are tested in this thesis is a total of 8 machine learning models. 4 of these models are ANN models, namely Levenberg-Marquardt back propagation, Bayesian regularization back
propagation, Scaled conjugate gradient back propagation and Quasi-Newton back propagation. The
rest of the machine learning models are XGBoost, Adaboost, Linear regression and Support vector
regression model.
ANN models are simulated in a programming software called MATLAB 2021a and the other 4 models
are simulated in a programming software called Python. Accuracy of each models are analyzed based
on 4 different quality measures, namely RMSE, MAE, MAPE and R2.
Results show that the models struggle to adapt to data on heating demand. Predictions on electricity
consumption is however more accurate. A challenge when it comes to prediction with machine
learning models is that some of the models predict values below 0 Watt-hour. Tuning of hyperparameters show that the accuracy vary.
This study conclude that the QNBP model gave the most accurate results for prediction on electricity
consumption. SVR model gave the most accurate results on prediction of heating demand. Comparison between results in other studies show that predictions in this thesis perform badly. Weather data,
weekday and working hour index does not describe the heating demand and electricity consumption
in the buildings sufficient. It is no clear indication that hyper-parameters in this project produces
more or less accurate prediction on dataset 1 and 2. However, changes on different hyper-parameter
settings had an impact on prediction accuracy . | |