Accuracy of different machine learning algorithms and added-value of predicting aggregated-level energy performance of commercial buildings

Research output: Contribution to journalArticleAcademicpeer-review

11 Downloads (Pure)

Abstract

As with many other sectors, to improve the energy performance and energy neutrality requirements of individual buildings and groups of buildings, built environment is also making use of machine learning for improved energy demand predictions. The goal of achieving energy neutrality through maximized use of on-site produced renewable energy and attaining optimal level of energy performance at building-cluster level requires reliable short term (resolution shorter than one day) energy demand predictions. However, the prediction and analysis of the energy performance of buildings is still focused on the individual building level and not on small neighborhood scale or building clusters. In a smart grid context, to better understand electricity consumption at different spatial levels, prediction should be at both individual as well as at building-cluster levels, especially for neighborhoods with definite boundaries (such as universities, hospitals). Therefore, in this paper, using data from 47 commercial buildings, a number of machine learning algorithms were evaluated to predict the electricity demand at individual building level and aggregated level in hourly intervals. Predicting at hourly granularity is important to understand short-term dynamics, yet most of the neighborhood scale studies are limited to yearly, monthly, weekly, or daily data resolutions. Two years of data were used in training the model and the prediction was performed using another year of untrained data. Learning algorithms such as; boosted-tree, random forest, SVM-linear, quadratic, cubic, fine-Gaussian as well as ANN were all analysed and tested for predicting the electricity demand of individual and groups of buildings. The results showed that boosted-tree, random forest, and ANN provided the best outcomes for prediction at hourly granularity when metrics such as computational time and error accuracy are compared.

Original languageEnglish
Article number109705
Number of pages14
JournalEnergy and Buildings
Volume209
DOIs
Publication statusPublished - 15 Feb 2020

Fingerprint

Energy Levels
Electron energy levels
Learning algorithms
Learning systems
Learning Algorithm
Machine Learning
Energy
Prediction
Electricity
Neutrality
Random Forest
Granularity
Smart Grid
Renewable Energy
Buildings
Sector
Metric
Predict
Interval
Demand

Keywords

  • ANN
  • Groups of buildings
  • Hourly demand prediction
  • Regression-trees

Cite this

@article{9d7102151bf345d5a06c9d2c728212f7,
title = "Accuracy of different machine learning algorithms and added-value of predicting aggregated-level energy performance of commercial buildings",
abstract = "As with many other sectors, to improve the energy performance and energy neutrality requirements of individual buildings and groups of buildings, built environment is also making use of machine learning for improved energy demand predictions. The goal of achieving energy neutrality through maximized use of on-site produced renewable energy and attaining optimal level of energy performance at building-cluster level requires reliable short term (resolution shorter than one day) energy demand predictions. However, the prediction and analysis of the energy performance of buildings is still focused on the individual building level and not on small neighborhood scale or building clusters. In a smart grid context, to better understand electricity consumption at different spatial levels, prediction should be at both individual as well as at building-cluster levels, especially for neighborhoods with definite boundaries (such as universities, hospitals). Therefore, in this paper, using data from 47 commercial buildings, a number of machine learning algorithms were evaluated to predict the electricity demand at individual building level and aggregated level in hourly intervals. Predicting at hourly granularity is important to understand short-term dynamics, yet most of the neighborhood scale studies are limited to yearly, monthly, weekly, or daily data resolutions. Two years of data were used in training the model and the prediction was performed using another year of untrained data. Learning algorithms such as; boosted-tree, random forest, SVM-linear, quadratic, cubic, fine-Gaussian as well as ANN were all analysed and tested for predicting the electricity demand of individual and groups of buildings. The results showed that boosted-tree, random forest, and ANN provided the best outcomes for prediction at hourly granularity when metrics such as computational time and error accuracy are compared.",
keywords = "ANN, Groups of buildings, Hourly demand prediction, Regression-trees",
author = "Shalika Walker and Waqas Khan and Katarina Katic and Wim Maassen and Wim Zeiler",
year = "2020",
month = "2",
day = "15",
doi = "10.1016/j.enbuild.2019.109705",
language = "English",
volume = "209",
journal = "Energy and Buildings",
issn = "0378-7788",
publisher = "Elsevier",

}

TY - JOUR

T1 - Accuracy of different machine learning algorithms and added-value of predicting aggregated-level energy performance of commercial buildings

AU - Walker, Shalika

AU - Khan, Waqas

AU - Katic, Katarina

AU - Maassen, Wim

AU - Zeiler, Wim

PY - 2020/2/15

Y1 - 2020/2/15

N2 - As with many other sectors, to improve the energy performance and energy neutrality requirements of individual buildings and groups of buildings, built environment is also making use of machine learning for improved energy demand predictions. The goal of achieving energy neutrality through maximized use of on-site produced renewable energy and attaining optimal level of energy performance at building-cluster level requires reliable short term (resolution shorter than one day) energy demand predictions. However, the prediction and analysis of the energy performance of buildings is still focused on the individual building level and not on small neighborhood scale or building clusters. In a smart grid context, to better understand electricity consumption at different spatial levels, prediction should be at both individual as well as at building-cluster levels, especially for neighborhoods with definite boundaries (such as universities, hospitals). Therefore, in this paper, using data from 47 commercial buildings, a number of machine learning algorithms were evaluated to predict the electricity demand at individual building level and aggregated level in hourly intervals. Predicting at hourly granularity is important to understand short-term dynamics, yet most of the neighborhood scale studies are limited to yearly, monthly, weekly, or daily data resolutions. Two years of data were used in training the model and the prediction was performed using another year of untrained data. Learning algorithms such as; boosted-tree, random forest, SVM-linear, quadratic, cubic, fine-Gaussian as well as ANN were all analysed and tested for predicting the electricity demand of individual and groups of buildings. The results showed that boosted-tree, random forest, and ANN provided the best outcomes for prediction at hourly granularity when metrics such as computational time and error accuracy are compared.

AB - As with many other sectors, to improve the energy performance and energy neutrality requirements of individual buildings and groups of buildings, built environment is also making use of machine learning for improved energy demand predictions. The goal of achieving energy neutrality through maximized use of on-site produced renewable energy and attaining optimal level of energy performance at building-cluster level requires reliable short term (resolution shorter than one day) energy demand predictions. However, the prediction and analysis of the energy performance of buildings is still focused on the individual building level and not on small neighborhood scale or building clusters. In a smart grid context, to better understand electricity consumption at different spatial levels, prediction should be at both individual as well as at building-cluster levels, especially for neighborhoods with definite boundaries (such as universities, hospitals). Therefore, in this paper, using data from 47 commercial buildings, a number of machine learning algorithms were evaluated to predict the electricity demand at individual building level and aggregated level in hourly intervals. Predicting at hourly granularity is important to understand short-term dynamics, yet most of the neighborhood scale studies are limited to yearly, monthly, weekly, or daily data resolutions. Two years of data were used in training the model and the prediction was performed using another year of untrained data. Learning algorithms such as; boosted-tree, random forest, SVM-linear, quadratic, cubic, fine-Gaussian as well as ANN were all analysed and tested for predicting the electricity demand of individual and groups of buildings. The results showed that boosted-tree, random forest, and ANN provided the best outcomes for prediction at hourly granularity when metrics such as computational time and error accuracy are compared.

KW - ANN

KW - Groups of buildings

KW - Hourly demand prediction

KW - Regression-trees

UR - http://www.scopus.com/inward/record.url?scp=85076861563&partnerID=8YFLogxK

U2 - 10.1016/j.enbuild.2019.109705

DO - 10.1016/j.enbuild.2019.109705

M3 - Article

AN - SCOPUS:85076861563

VL - 209

JO - Energy and Buildings

JF - Energy and Buildings

SN - 0378-7788

M1 - 109705

ER -