Building energy consumption prediction using deep learning

Razak Olu-Ajayi
Big Data Technologies and Innovation Laboratory, University of Hertfordshire, Hatfield, AL10
9AB, UK
[email protected]

Hafiz Alaka
Big Data Technologies and Innovation Laboratory, University of Hertfordshire, Hatfield, AL10
9AB, UK
[email protected]

Abstract: The consumption of energy in buildings has elicited the occurrence of many environmental
problems such as air pollution. Building energy consumption prediction is fundamental for improved
decision-making towards regulating or decreasing energy usage. There have been several applications of
Machine Learning (ML) algorithms for predicting the energy consumption of operational buildings
without much exploration into forecasting the potential building energy consumption at the early design
stage. On the topic of reducing energy inefficient buildings, it is essential to address the root of the
problem, the essentiality of predicting energy use before construction to alleviate futuristic problems of
constructing new buildings that are harmful to the environment. At the early design stage, the customary
model utilised for predicting energy consumption is the forward model, based on building energy
modelling tools, which is stated to be mundane and time consuming. In contrast, the Machine Learning
(ML) model is recognized as the most contemporary and best technique for prediction. To address this
gap, this paper (1) presents the utilization of deep learning for predicting annual energy consumption of
buildings, and (2) conduct a comparative analysis of the prediction performance of the models. The
originality of this paper is to build a model trained by a dataset of multiple buildings that enables building
designers to input key features of a building design and forecast the annual average energy consumption
at the early stages of development. The ANN method outperforms SVM and DT for predicting annual
energy consumption.

Keywords: Building energy consumption prediction, Deep learning, Data-driven model,

Machine Learning.

1. INTRODUCTION
In recent years, the forecasting of building energy consumption has gained more attention by
researchers and specialists (Amasyali and El-Gohary, 2017). This emanates from the report that
the amount of energy consumed in buildings deploys a severe impact on the existence of
mankind through the cause of major environmental problems such as climate change, air
pollution, among others (Dandotiya, 2020). Energy inefficient buildings are recognised as the
main contributors to global energy consumption and Green House Gas (GHG) emission (Pham
et al., 2020; United Nations Environment Programme, 2017). According to the United Kingdom
(UK) Building Energy Efficiency Survey (BEES), the energy consumed in buildings is
accountable for 70 per cent of the total consumption (Building Energy Efficiency Survey, 2016).
Therefore, the prediction of building energy use is essential to provide building owners and
facility managers with the capacity to make informed decisions towards reducing energy
consumption. However, accurate building energy forecasting remains a complex task due to
certain factors that cannot be easily determined or obtained, such as occupant energy-use
behaviour, physical properties of the building, among others (Hamed and Nada, 2019).

Building energy simulation tools, such as TRNSYS, DOE-2, and EnergyPlus are extensively
utilized for forecasting energy consumption of operational buildings and buildings at the design
stage. However, these tools are very detailed and elaborate, often requiring a large number of
input parameters about the building and its environment (e.g., HVAC (Heating, Ventilation and
Air Conditioning) system, Physical properties, internal occupancy loads, solar information and
so on) which are not generally available to users (Runge and Zmeureanu, 2019). In many cases,
the inability to provide the required input parameters leads to poor estimation performance
(Amasyali and El-Gohary, 2017). In contrast, data driven models use Machine Learning (ML)
algorithms such as Support Vector Machine (SVM), Decision Tree (DT), among others for
energy use prediction. These models do not require a significant number of input parameters
and make predictions based on historical data obtained from Building management systems and
smart meters (Tardioli et al., 2015). However, in machine learning, the achievement of good
prediction performance is predicated on these three factors: model selected, quantity and quality
of the data (Runge and Zmeureanu, 2019).

The investigation of the applicability of a considerable number of machine learning algorithms

for predicting energy consumption have been conducted in several studies (Aversa et al., 2016;
Dong et al., 2005; Q. Li et al., 2009a; Pham et al., 2020; Tardioli et al., 2015). Among them,
the most utilized for energy prediction are Decision Tree (DT), Artificial Neural Networks
(ANN) and Support Vector Machine (SVM). These ML algorithms have been applied for
predicting the energy consumption of operational buildings without much exploration into
forecasting the potential energy consumption at the early design stage. Energy use prediction at
the design stage is often implemented using building energy modelling tools, which is
quantified to be resource demanding and time consuming (Zhu, 2006). Contrary to the
utilization of ML models, which is recognized as the most contemporary and best technique for
prediction (Canales, 2016; Vorobeychik and Wallrabenstein, 2013). To significantly decrease
the construction of energy inefficient buildings, it is essential to optimize and identify the best
model for energy consumption prediction during the design period. The development of a model
with good prediction performance would enable building designers to check the energy
consumption level at the early design stage, to detect if there is a need for design adjustment. In
this study, the major ML algorithms namely DT, multi-layer perceptron ANN and SVM are
applied to forecast the annual energy consumption of residential buildings. This paper focuses
on developing a model that enables building designers to input parameters that are readily
accessible at the early design stage to forecast the annual average energy consumption of the
building before construction.

The remainder of the paper is structured as follows: Section 2 delivers a brief background on
the most utilized supervised machine learning techniques for forecasting building energy
consumption. Section 3 presents the research methodology, which consists of the description of
the data collected, data pre-processing, model development and performance measures. Section
4 provides and examines the result, while Section 5 tenders the conclusions and future work.

2. LITERATURE REVIEW
Energy use predicting plays a vital role in energy conservation, financial cost reduction and
enabling facility managers to make informed decisions towards reducing the energy consumed.
In the application of ML algorithm for energy prediction, majority of the research aims to obtain
the highest level of accuracy (García-Martín et al., 2019). A large variety of supervised machine
learning algorithms have been utilized for energy prediction, although ANN, SVM and DT are
recognised as the most utilized (Amasyali and El-Gohary, 2017). These algorithms comprise of
both advantages and disadvantages in various situations. For instance, ANN and SVM often
produce more accurate results than DT. However, DT techniques are elementary and easier to
implement (Tso and Yau, 2007).

Artificial Neural Networks (ANN) is a non-linear computational algorithm that emulates the
functional concepts of the human brain (Alaka et al., 2018; Amasyali and El-Gohary, 2018). A
basic form of ANN consists of three consecutive layers: input, hidden and output layer. ANN
is identified as a recurrent model and has gained more attention due to its good prediction
performance. It is known to be dominant with big datasets, which enable the neural network
sufficient data to train the model (Bourhnane et al., 2020). ANN algorithm is also known to
produce good accuracy in energy load prediction (K. Li et al., 2018). Multi-layer Perceptron
(MLP) is a function of a deep neural network that utilizes a feed forward propagation process
with one hidden layer where latent and abstract features are learned (Donoghue and Roantree,
2015). In research by Khantach et al., Multi-layer Perceptron ANN produced the most accurate
result between support vector machine (SVM), Gaussian process and radial basis function
(RBF) with a Mean Absolute Percentage Error (MAPE) of 0.96 (Khantach et al., 2019). In 2019,
Runge and Zmeureanu concluded that ANN produces good results when applied to single and
multi-step ahead forecasting (Runge and Zmeureanu, 2019). Neto et al conducted a comparative
analysis of the prediction performance of ANN and an energy simulation tool called EnergyPlus
using a single dataset. The results revealed that data driven methods (ANN) are better suited for
predicting building energy load (Aversa et al., 2016). The study by Bagnasco et al concludes
that ANN performs better in the winter season when forecasting electrical consumption based
on meteorological data and time/day variation (Bagnasco et al., 2015).

Dong et al 2005, first proposed the utilization of Support Vector Machine (SVM) for building
energy use prediction and used four buildings dataset for forecasting monthly electricity
consumption. Dong et al proclaimed that SVM outperforms other related research using neural
networks with a coefficient of Determination R2 higher than 0.99 (Dong, Cao and Lee, 2005).
Similarly, Li et al explored the utilization of SVM for forecasting hourly cooling load in
buildings and concluded that SVM produces a good result with a Root Mean Square Error
(RMSE) of 1.17% (Li et al., 2009). On the other hand, Decision Tree (DT) does not outperform
neural networks for non-linear data. However, its popularity can be attributed to its ease of use
and ability to produce predictive models with interpretable structures. In Hong Kong, Tso and
Yau conducted a comparative analysis of the prediction performance of decision tree, neural
network and regression method in predicting weekly electricity consumption. Tso and Yau
gathered that decision trees and neural networks perform slightly better than the regression
method with a root of average squared error (RASE) of 39.36 (Tso and Yau, 2007).

3. RESEARCH METHODS
In this research, a prediction approach is explored based on supervised machine learning
regression algorithms to forecast the annual energy consumption of residential buildings. The
dataset is collected for this research is contains residential buildings data from various cities in
the United Kingdom (UK) from January 2020 to December 2020. The prediction method will
employ three machine learning algorithms which are multi-layer perceptron ANN, SVM and
DT. The model development will be implemented in jupyter notebook utilizing the python
programming language. Prior to training and testing the model, the raw data will first be
analysed and cleaned or pre-processed to avoid any possible complexity, such as missing data
during the training stage. Lastly, the performance of each model will be evaluated using
performance metrics. Furthermore, the energy use prediction framework will consist of four
sections: Data collection, Data pre-processing, Model development and Model evaluation.

3.1 Data collection

The building database was gathered from an online repository called Ministry of Housing
Communities and Local Government (MHCLG) Repository. The project recorded the metadata
of 300 real residential buildings around the UK. Each budling data encompasses the annual
average energy consumption data. Additionally, the metadata includes several characteristics
that are accessible at the design stage, such as floor area, wall type, roof type and so on. There
are two types of data: meteorological data and building data. The meteorological data was
obtained from an online repository called Meteostat repository, and it was recorded each
instance which consist of temperature, wind speed and humidity. The monthly meteorological
dataset was obtained from January 2020 to December 2020.
All variables utilized in this study are listed in Table 1. The building metadata is labelled as
internal variables, while the meteorological data is labelled as external variables.

Table 1: List of variables selected

ID Variable Internal/External Type Unit Label

o
1 Temperature c

2 Wind speed km/h

External Continuous
3 Pressure Hg

4 Total Floor Area m2

5 Floor Description 0, 1, 2, 3 Independent

6 Walls Description Categorical 0, 1, 2, 3

7 Windows Description 0, 1, 2, 3
Internal
8 Number of Heated 0, 1, 2, 3
Rooms
Discrete 0, 1, 2, 3
9 Number of Habitable
Rooms
10 Energy Consumption Continuous kWh/m2 Dependent

3.2 DATA PRE-PROCESSING

The preliminary procedure in machine learning involves data pre-processing for the preparation
of data, and it often expends a significant amount of time and computational power (Shapi et
al., 2021). This process is requisite as the dataset sometimes contains missing values and a few
outliers. In this study, the data was pre-processed by employing the method of the imputation
of missing data and normalization. The meteorological data had few missing values that were
managed by utilizing the mean value imputation as presented in by Newgard and Lewis, 2015.
In the raw data, the metadata consisted of specific variables that were descriptive, such as wall
description, windows description and so on. The affected variables in the metadata were
assigned values and filled using python library. The metadata was merged with meteorological
annual average data using each location postcode as the primary key.

At this point, the data was normalized using the friedman of the sklearn python package. Data
normalization is a very common technique of data pre-processing as various features have
dissimilar dimensions, it is requisite to normalize different data to eradicate the influence of the
dimension (Liu et al., 2020).

3.3 Model development (training)

This research utilized a supervised machine learning approach to forecasting the annual energy
consumption of buildings. Subsequent to data preparation, it was fed into the learning
algorithms. The data was then partitioned into two sets namely train set and test set at a ratio of
7:3. The predictive modelling approach utilized the regression model to forecast energy
consumption. The three machine learning algorithms developed in this research are multi-layer
perceptron ANN, SVM, and DT. The partitioned training set of 70% was used to train each
machine learning algorithm to develop a predictive model that would estimate energy
consumption based on the specified input. Consequently, the other 30% was utilized to test the
trained prediction model.

3.4 Model evaluation (test)

To assess estimation performance of each model, the following performance measure were
utilized: R-Squared (R2), Mean Absolute Error (MAE), Mean Squared Error (MSE) and Root
Mean Squared Error (RMSE), Among all stated measures, the most frequently employed are
MSE and RMSE (Dong et al., 2005; Q. Li et al., 2009a; Pham et al., 2020)

1. Mean Absolute Error (MAE) is a method of computing the variation between the predicted
values and true or actual values. MAE scores close to zero indicates better performance while
a score greater than zero signifies a bad performance.
1 𝑛
𝑀𝐴𝐸 = ∑ |𝐴𝐸𝑖 − 𝑃𝐸𝑖 | (1)
𝑛 𝑖=1

2. Mean Squared Error (MSE) is the evaluation of the squared variance between the estimated
values and the actual values. MSE is the calculation of the quality of a prediction model. MSE
score closer to zero have good performance.
1
𝑀𝑆𝐸 = ∑𝑛𝑖=1(𝐴𝐸𝑖 − 𝑃𝐸𝑖 )2 (2)
𝑛

3. Root Mean Squared Error (RMSE) is also a measure of calculating the variances between the
estimated value and the true value. It is calculated by the square root of the Mean Square Error
(MSE).

1
𝑅𝑀𝑆𝐸 = √ ∑𝑛𝑖=1(𝐴𝐸𝑖 − 𝑃𝐸𝑖 )2 (3)
𝑛

4. R-Squared (R2) is an evaluation method that determines the proportion of the variance in the
target variable that can be justified by the independent variables. It exhibits the degree to which
the data fits the model. R2 can generate a negative result, but the best performance of R2 is 1.0.

∑𝑛
𝑖=1(𝑦𝑝𝑟𝑒𝑑,𝑖 −𝑦𝑑𝑎𝑡𝑎,𝑖 )
2
𝑅2 = 1 − ∑𝑛 2 (4)
𝑖=1(𝑦𝑑𝑎𝑡𝑎,𝑖 −𝑦𝑑𝑎𝑡𝑎 )

The range of experimental task conducted from data pre-processing to model evaluation was
implemented using python. The computation was performed on a MacBook BigSur Os version
11.3 with M1 chip and 16gb RAM.

4. RESULT AND DISCUSSION

The prediction performance of the ANN and SVM models are comparable. However, DT
evidently has the lower performance. The four performance metrics utilized to evaluate the
performance were computed for the three models as presented in table 2 below. Based on the
accuracy and computational efficiency, The result reveals which model best predicts annual
energy consumption. Each model produced satisfactory results except the decision tree model.
The ANN based model slightly outperforms the SVM based model with better MAE, RMSE,
MSE and R2, However, during training, the SVM based model consumed a shorter time than
ANN. Contrarily, the DT model performed worse than both ANN and SVM model but trained
faster. The best models are presented in bold font in Table 2 below. This affirms ANN based
model as a potentially good model to consider for forecasting annual energy consumption.

Table 2: Performance result for all model

Model Training Time R-squared MAE RMSE MSE
Artificial Neural Network 1.2s 0.66 2.20 2.80 7.85
Support Vector Machines 1.0s 0.59 2.40 3.07 9.41
Decision Tree 877ms 0.27 3.17 4.10 16.83

Figures 1a-1c displays a visual representation of actual and predicted annual energy
consumption for each test instance. This presents the difference between the actual and
predicted values which corroborates ANN’s performance result as the better model. The
performance of ANN corroborates the proclaimed theory in several studies that ANN is the
better model for building energy prediction (Khantach et al., 2019; K. Li et al., 2018; Runge
and Zmeureanu, 2019).

Figure 1a: Visualization of ANN based model performance

Figure 1b: Visualization of ANN based model performance

 Figure 1c: Visualization of actual and predicted values of DT

The predictive performance of ANN states it as a better model than SVM in an application on
the same data and same situation. This presents a motivation for utilization of ANN for energy
prediction before construction which would reduce the development of energy in efficient
buildings. In related works by Dong et al 2005, SVM performed better than ANN using a dataset
of 4 buildings, as opposed the utilization of 300 buildings in this research. The performance of
ANN can be attributed to its dominance in big datasets to enable the neural network sufficient
data to train the model (Bourhnane et al., 2020).

5. CONCLUSIONS AND RECOMMENDATIONS

The course of the research focused on the annual energy prediction approach of residential
buildings using the most utilised data driven based algorithms. The framework of this study
affirms the possibility of a model that enable building designers to make informed decisions at
the predesign stage of development. In this study, the models utilised, namely ANN, SVM and
DT were trained, tested, and evaluated based on computational efficiency and accuracy. The
performance result conveys multi-layer perceptron ANN has the best predictive model with an
R2 value of 0.66 and MAE value of 2.20. Similarly, SVM achieved good performance with an
R2 value of 0.59 and MAE value of 2.40. On the other hand, the DT model had the fastest
training time of 877ms but produced the worst result for annual energy prediction. The

The performance result of Artificial Neural Network (ANN) shows promising potential for the
estimation of annual energy consumption, especially on a larger dataset. Future research should
focus on the application of ANN on a larger dataset and exploration into the suitability of other
data driven regression models for annual energy use prediction.

6. REFERENCES

Alaka, H. A., Oyedele, L. O., Owolabi, H. A., Kumar, V., Ajayi, S. O., Akinade, O. O., and

Bilal, M. (2018). Systematic review of bankruptcy prediction models: Towards a

 framework for tool selection. Expert Systems with Applications, 94, 164–184.

https://doi.org/10.1016/j.eswa.2017.10.040

Amasyali, K., and El-Gohary, N. (2017, May 31). Deep Learning for Building Energy

Consumption Prediction.

Amasyali, K., and El-Gohary, N. M. (2018). A review of data-driven building energy

consumption prediction studies. Renewable and Sustainable Energy Reviews, 81,

1192–1205. https://doi.org/10.1016/j.rser.2017.04.095

Aversa, P., Donatelli, A., Piccoli, G., and Luprano, V. A. M. (2016). Improved Thermal

Transmittance Measurement with HFM Technique on Building Envelopes in the

Mediterranean Area. Selected Scientific Papers - Journal of Civil Engineering, 11.

https://doi.org/10.1515/sspjce-2016-0017

Bagnasco, A., Fresi, F., Saviozzi, M., Silvestro, F., and Vinci, A. (2015). Electrical

consumption forecasting in hospital facilities: An application case. Energy and

Buildings, 103, 261–270. https://doi.org/10.1016/j.enbuild.2015.05.056

Bourhnane, S., Abid, M. R., Lghoul, R., Zine-Dine, K., Elkamoun, N., and Benhaddou, D.

(2020). Machine learning for energy consumption prediction and scheduling in smart

buildings. SN Applied Sciences, 2(2), 297. https://doi.org/10.1007/s42452-020-2024-9

Building Energy Efficiency Survey. (2016). Building Energy Efficiency Survey (BEES).

GOV.UK. https://www.gov.uk/government/collections/non-domestic-buildings-

energy-use-project

Canales, V. R. (2016). Using a Supervised Learning Model: Two-Class Boosted Decision

Tree Algorithm for Income Prediction. 7.

Dandotiya, B. (2020). Climate-Change-and-Its-Impact-on-Terrestrial-Ecosystems.

https://doi.org/10.4018/978-1-7998-3343-7.ch007

Dong, B., Cao, C., and Lee, S. E. (2005). Applying support vector machines to predict

building energy consumption in tropical region. Energy and Buildings, 37(5), 545–

553. https://doi.org/10.1016/j.enbuild.2004.09.009
Donoghue, J. O., and Roantree, M. (2015). A Framework for Selecting Deep Learning Hyper-

parameters. In S. Maneth (Ed.), Data Science (Vol. 9147, pp. 120–132). Springer

International Publishing. https://doi.org/10.1007/978-3-319-20424-6_12

García-Martín, E., Rodrigues, C. F., Riley, G., and Grahn, H. (2019). Estimation of energy

consumption in machine learning. Journal of Parallel and Distributed Computing,

134, 75–88. https://doi.org/10.1016/j.jpdc.2019.07.007

Hamed, M., and Nada, S. (2019). Statistical Analysis for Economics of the Energy

Development in North Zone of Cairo. International Journal of Finance and

Economics, 5, 140–160.

Khantach, A. E., Hamlich, M., Belbounaguia, N. eddine, Khantach, A. E., Hamlich, M., and

Belbounaguia, N. eddine. (2019). Short-term load forecasting using machine learning

and periodicity decomposition. AIMS Energy, 7(3), 382–394.

https://doi.org/10.3934/energy.2019.3.382

Li, K., Xie, X., Xue, W., Dai, X., Chen, X., and Yang, X. (2018). A hybrid teaching-learning

artificial neural network for building electrical energy consumption prediction.

Energy and Buildings, 174, 323–334. https://doi.org/10.1016/j.enbuild.2018.06.017

Li, Q., Meng, Q., Cai, J., Yoshino, H., and Mochida, A. (2009a). Predicting hourly cooling

load in the building: A comparison of support vector machine and different artificial

neural networks. Energy Conversion and Management, 50(1), 90–96.

https://doi.org/10.1016/j.enconman.2008.08.033

Li, Q., Meng, Q., Cai, J., Yoshino, H., and Mochida, A. (2009b). Applying support vector

machine to predict hourly cooling load in the building. Applied Energy, 86(10), 2249–

2256. https://doi.org/10.1016/j.apenergy.2008.11.035

Liu, Y., Chen, H., Zhang, L., Wu, X., and Wang, X. (2020). Energy consumption prediction

and diagnosis of public buildings based on support vector machine learning: A case

study in China. Journal of Cleaner Production, 272, 122542.

https://doi.org/10.1016/j.jclepro.2020.122542
Newgard, C. D., and Lewis, R. J. (2015). Missing Data: How to Best Account for What Is Not

Known. JAMA, 314(9), 940–941. https://doi.org/10.1001/jama.2015.10516

Pham, A.-D., Ngo, N.-T., Ha Truong, T. T., Huynh, N.-T., and Truong, N.-S. (2020).

Predicting energy consumption in multiple buildings using machine learning for

improving energy efficiency and sustainability. Journal of Cleaner Production, 260,

121082. https://doi.org/10.1016/j.jclepro.2020.121082

Runge, J., and Zmeureanu, R. (2019). Forecasting Energy Use in Buildings Using Artificial

Neural Networks: A Review. Energies, 12(17), 3254.

https://doi.org/10.3390/en12173254

Shapi, M. K. M., Ramli, N. A., and Awalin, L. J. (2021). Energy consumption prediction by

using machine learning for smart building: Case study in Malaysia. Developments in

the Built Environment, 5, 100037. https://doi.org/10.1016/j.dibe.2020.100037

Tardioli, G., Kerrigan, R., Oates, M., O‘Donnell, J., and Finn, D. (2015). Data Driven

Approaches for Prediction of Building Energy Consumption at Urban Level. Energy

Procedia, 78, 3378–3383. https://doi.org/10.1016/j.egypro.2015.11.754

Tso, G. K. F., and Yau, K. K. W. (2007). Predicting electricity energy consumption: A

comparison of regression analysis, decision tree and neural networks. Energy, 32(9),

1761–1768. https://doi.org/10.1016/j.energy.2006.11.010

United Nations Environment Programme, U. N. (2017, September 26). Sustainable buildings.

UNEP - UN Environment Programme. http://www.unep.org/explore-topics/resource-

efficiency/what-we-do/cities/sustainable-buildings

Vorobeychik, Y., and Wallrabenstein, J. R. (2013). Using Machine Learning for Operational

Decisions in Adversarial Environments. 9.

Zhu, Y. (2006). Applying computer-based simulation to energy auditing: A case study.

Energy and Buildings, 38(5), 421–428. https://doi.org/10.1016/j.enbuild.2005.07.007