Improving sustainable office building operation by using historical data and linear models to predict energy usage

Sustainable Cities and Society 29 (2017) 107–117

Contents lists available at ScienceDirect

Sustainable Cities and Society journal homepage: www.elsevier.com/locate/scs

Improving sustainable office building operation by using historical data and linear models to predict energy usage Majeed Safa a,∗ , Mahdi Safa b , Jeremy Allen c , Arash Shahi d , Carl T. Haas e a

Department of Agricultural Management and Property Studies, Lincoln University, New Zealand College of Business, Lamar University, USA Energy Solution Providers Ltd, Auckland, New Zealand d Department of Civil Engineering and Mineral, University of Toronto, Canada e Department of Civil and Environmental Engineering, University of Waterloo, Canada b c

a r t i c l e

i n f o

Article history: Received 21 October 2016 Received in revised form 2 December 2016 Accepted 2 December 2016 Available online 6 December 2016 Keywords: Energy modelling Energy auditing Office buildings Energy saving Artificial neural network Linear regression model

a b s t r a c t Offices and retail outlets represent the most intensive energy consumers in the non-residential building sector and have been estimated to account for more than 50% of a building’s energy usage. Accurate predictions of office building energy usage can provide potential energy savings and significantly enhance the efficient energy management of office buildings. This paper proposes a method that applies multiple linear regression (MLR) and artificial neural network (ANN) models to predict energy consumption based on weather conditions and occupancy; thus, enabling a comparison of the use of these two types of modelling methods. In this study, four models of office sites at research institutions in different New Zealand regions were developed to investigate the ability of simple models to reduce margins of error in energy auditing projects. The models were developed based on the monthly average outside temperature and the number of full-time employees (FTEs). A comparison of the actual and predicted energy usage revealed that the models can predict energy usage within an acceptable error range. The results also demonstrated that each building should be investigated as an individual unit. © 2016 Published by Elsevier Ltd.

1. Introduction In light of today’s sustainable development (SD) goals, buildings need to be constructed and operated so they have a minimal environmental impact, a controlled consumption of natural resources, and limited use of the primary energy life cycle (Gustavsson and Joelsson, 2010). Significant increases in energy consumption and costs due to rising populations, the expanding economy, and improved qualities of life in recent years, have raised concerns about the environment and the possible exhaustion of energy resources (Azhar, Brown, & Sattineni, 2010; Pérez-Lombard, Ortiz, & Pout, 2008; Harish and Kumar, 2016a; Yu, Haghighat, Fung, & Yoshino, 2010). For example, in the US, buildings represent about 74% of electricity usage and 41.1% of primary energy consumption (Li and Wen, 2014). Optimal building energy consumption, along with its design and operation, has been discussed in a number of recent research papers (Li, Liu, & Fang, 2010; Lund, Marszal, &

∗ Corresponding author. E-mail address: [email protected] (M. Safa). http://dx.doi.org/10.1016/j.scs.2016.12.001 2210-6707/© 2016 Published by Elsevier Ltd.

Heiselberg, 2011; Santamouris, 2013; Thyholt and Hestnes, 2008; Thomsen, Schultz, & Poel, 2005; Xu and Niu, 2006; Xue and Chen, 2014; Yilmaz and Basak Kundakci, 2008; Zhang, Lin, Yang, Di, & Jiang, 2006). Offices and retail outlets represent the most intensive energy users in the non-residential buildings, and account for more than 50% of energy usage (Pérez-Lombard et al., 2008). In New Zealand, typical office buildings consume electricity more than other energy resources. A study by the Building Research Association of New Zealand (BRANZ) showed that only 3.5% of non-residential buildings used diesel and fuel oil and 11% of non-residential buildings used natural gas and that was mostly for heating (Isaacs et al., 2010). The majority of electrical usage in the buildings investigated was for plug loads, water heating, lighting, and air-conditioning (HVAC) (Harish and Kumar, 2016a; Isaacs et al., 2010; Pérez-Lombard et al., 2008; Yu, Haghighat, & Fung, 2016). HVAC systems have a significant influence on energy consumption and the occupant’s comfort in commercial building. Lighting systems probably account for the second highest electricity use in commercial buildings, which mostly depends on occupancy pattern, day light available and the control system used (Harish and Kumar, 2016a)

108

M. Safa et al. / Sustainable Cities and Society 29 (2017) 107–117

The optimal operation of office buildings is a critical component in the reduction of overall energy usage in buildings. Such optimization requires robust systems to monitor and audit energy use (Juan, Gao, & Wang, 2010). Modelling is the most significant part when designing building energy controls and optimisation strategies (Harish and Kumar, 2016b). Several studies have shown that improving the energy efficiency of an existing building has a considerable, and positive, influence on the value of the building as well as on its environmental impact (Daly, Cooper, & Ma, 2014; Harish and Kumar, 2016a; Li and Wen, 2014). Yu et al. (Yu, Fung, & Haghighat, 2013) categorised the energy modelling studies based on analysing building-related data into typical indicator method, statistical analysis method, and building simulation method (Yu, Fung et al., 2013). The primary objective of this study was, therefore, to develop and compare two forecasting models for predicting office building energy usage to enable the measurement of energy savings during energy auditing projects. In this research, the ability of simple models to decrease margins of error in energy management and sustainability projects was investigated by comparing two energy prediction models in four research company offices in different regions of New Zealand. Determining and managing a practical energy savings plan for existing buildings is a complex optimization problem that has a number of technical limitations (Daly et al., 2014; Diakaki, Grigoroudis, & Kolokotsa, 2008). Building a performance simulation is a useful tool widely employed in the building design industry for monitoring and predicting the process of optimizing energy savings (Daly et al., 2014; Ma, Cooper, Daly, & Ledo, 2012). However, significant ‘performance gaps’ between operational energy consumption and the predictions produced by building performance simulations have been observed (Daly et al., 2014; Menezes, Cripps, Bouchlaghem, & Buswell, 2012). A number of factors, such as inaccurate data input variables or inappropriate simulation methods, could have resulted in these gaps (Menezes et al., 2012). Establishing an appropriate simulation method and accurate input variables is, therefore, crucial for the development of a practical prediction model. With respect to the development of energy consumption models for the construction industry, finding and accessing detailed data about the construction and energy performance of specific buildings is always a major challenge. This step is, therefore, critical for predicting the energy consumption of buildings. In many energy auditing projects, energy savings are predicted, based on a comparison of the current energy consumption with data from previous years (Abrahamse, Steg, Vlek, & Rothengatter, 2005; Brandon and Lewis, 1999). Because of environmental conditions, changing occupancy and other causes, energy consumption could be significantly different from year to year. Modelling energy consumption based on previous energy usage has the potential to deliver precise estimates of the energy savings. It should be noted that some energy information and variables are of more consequence than others, implying that the selection of variables is an important consideration during the modelling process (Ramesh, Prakash, & Shukla, 2010). A number of studies have analysed energy consumption in a variety of buildings. However, because of the lack of the advanced metering technologies available today, monthly electricity bills were used for analysing energy usage (Kavousian, Rajagopal, & Fischer, 2013) in most early energy modelling studies. If the data mining as the primary tool to analyse building-related data develop in future, huge amount of useful data will be available for similar studies (Yu, Haghighat et al., 2016; Yu, Fung et al., 2013). Also, many of the energy forecasting models look complicated mathematical equations especially for common users without advanced math-

ematical knowledge. Which make it more difficult to understand and use for energy conservation (Yu, Fung et al., 2013). The first important step in analysing energy usage in buildings is to obtain a clear understanding of most significant parameters with respect to energy usage in different types of buildings. Kavousian et al. (Kavousian et al., 2013) categorised four main effective parameters in residential buildings: appliances and electronic stock, weather and location, physical properties of the building, and occupancy. Nevertheless, in the survey of commercial buildings, Issacs et al. (Isaacs et al., 2010) classified buildings based on staff numbers, client numbers, business sector and activities, and operating periods. For sustainability plans and energy auditing projects, predicting energy consumption is a significant challenge. Forecasting energy usage is challenging because of the complexity of office buildings, which differ substantially with respect to design, construction, occupancy and activity. Such wide discrepancies make it is too challenging to classify the small numbers of buildings that could serve as representative samples in the majority of office buildings (Ahmad et al., 2014; Korolija, Marjanovic-Halburd, Zhang, & Hanby, 2013). An examination of numerous studies revealed that environmental parameters have been the primary factors used in most research directed at estimating energy consumption in buildings, with humidity, temperature and lux level constituting the most important environmental factors that can directly impact energy usage (Considine, 2000; Gugliermetti, Passerini, & Bisegna, 2004; Harish and Kumar, 2016a; Isaacs et al., 2010; Johnsen, 2001; Li and Wen, 2014; Moral-Carcedo and Vicéns-Otero, 2005; Pardo, Meneu, & Valor, 2002; Yu et al., 2010). Additional factors, such as electrical devices, geographical location and the timing of building use, can also indirectly influence energy usage predictions (Ahmad et al., 2014; Erkoreka, Garcia, Martin, Teres-Zubiaga, & Del Portillo, 2016). While few optimization methods have been developed for estimating energy consumption; however, a variety of modelling techniques have been applied over recent years. Energy forecasting models are made up of input variables, output variables and the structure of the model (Ahmad et al., 2014; Harish and Kumar, 2016a; Mathews, Botha, Arndt, & Malan, 2001; Moral-Carcedo and Vicéns-Otero, 2005; Mukherjee et al., 2010; Ma, Qin, Salsbury, & Xu, 2012; Magnier and Haghighat, 2010; Pandharipande and Caicedo, 2011; Roche and Milne, 2005; Rubinstein, Neils, & Colak, 2001; Safa and Allen, 2014; Üc¸tu˘g and Yükseltan, 2012; Vakiloroaya, Su, & Ha, 2011; Wang, Zmeureanu, & Rivard, 2005). Lack of validation data, the small size of input variables, complications in their use, and focusing only on construction elements and environmental factors, would be the main limitations of many of the energy usage prediction models developed. In many studies, such as Erkoreka, A. et al’s. (Erkoreka et al., 2016), the energy prediction models in buildings have been developed mostly based on details of the building structure, such as HVAC systems, insulation, design, material and orientation. The models look practical but the model development process would be very complex and time consuming, and with several factors needing to be considered. If the appropriate historical data are available, developing models to estimate energy usage in buildings is much easier and it is possible to develop models with a small number of input variables. However, the model of each building would be completely different and all changes during the investigation process should be considered carefully (Safa and Allen, 2014). Linear regression (Catalina, Virgone, & Blanco, 2008; Cheung and Braun, 2016; Ghiaus, 2006) and artificial neural networks (ANNs) (Magnier and Haghighat, 2010; Zhang and Haghighat, 2010; Deb, Eang, Yang, & Santamouris, 2016; Mba, Meukam, & Kemajou, 2016; Sholahudin and Han, 2016) have been used more frequently than other modelling methods for forecasting energy usage in build-

M. Safa et al. / Sustainable Cities and Society 29 (2017) 107–117

109

ings. While, previously, linear regression analysis was the most common modelling method for energy management projects; artificial neural networks (ANN) have, nevertheless, been become more common in recent years (Ahmad et al., 2014; González and ˜ 2005; Karatasou, Santamouris, & Geros, 2006; Sözen, Zamarreno, 2009; Safa and Samarasinghe, 2011a; Yang, Rivard, & Zmeureanu, 2005; Yu, Haghighat et al., 2016). Because of ANNs’ ability to shape nonlinear networks in a flexible and adaptive model, they are being developed more and more for investigating a variety of scientific and engineering problems (Jebaraj and Iniyan, 2006; Yu, Haghighat et al., 2016). Their primary advantage is their ability to use existing information (historical data relating to the underlying processes) to develop a precise illustration of the process of interest (Safa and Samarasinghe, 2011b). However, since NN architecture differs from the architecture of a microprocessor, ANNs must be produced as specific replicas, which makes developing a large neural network very time consuming (Ahmad et al., 2014). When the input is processed by the ANN model, each neuron in the first layer runs the weighted input variables through a transfer function in order to produce an output. Transfer functions, which may be either linear or nonlinear, include hyperbolic tangent, logistic, Gaussian and sine versions. The output of the first layer, which is dependent on the applied transfer function, is then sent to the neurons in the following layer through weighted corrections. These neurons then complete their output by processing the sum of the weighted input values through their transfer functions. When this layer becomes the last layer (output layer), the output of the neurons is the predicted output model (Heinzow and Tol, 2003; Hornik, Stinchocombe, & White, 1989; Jebaraj and Iniyan, 2006; Safa and Samarasinghe, 2011a). In both MLR and ANN models when the model training ends; and with the final weights thus obtained, another independent validation dataset is used to test the model predictions (Safa and Samarasinghe, 2011a; Samarasinghe, 2007). Fig. 1. New Zealand site locations: Dunedin (D), Christchurch (C), Palmerston North (P), and Hamilton (H).

2. Methods This study involved an investigation of four very similar research office buildings, in Hamilton (37◦ 47 S,175◦ 17 E), Palmerston North (40◦ 21.3 S,175◦ 36.7 E), Christchurch (43◦ 31 48 S,172◦ 37 13 E), and Dunedin (45◦ 52 S,170◦ 30 E) belonging to the same research company and located in different areas of New Zealand (Fig. 1). All the buildings examined were built in 1970. Each building was investigated independently, based on the existing monthly energy usage from previous years. Multiple linear regression (MLR) models and artificial neural networks (ANNs) were developed for predicting energy consumption. For the purposes of this work, the models were created with a minimum number of input variables. After a number of the available variables were examined, occupancy (full-time employees) and average air temperature were applied as the input variables in the model development. Average monthly temperatures were collected from the National Institute of Water and Atmospheric Research (NIWA) database, and the number of employees was collected through property managers and HR staff of the company being studied. Based on NIWA reports, North Island of New Zealand has warmer climate than Sought Island, Christchurch and Dunedin climate is temperate and Hamilton and Palmerston North climate is oceanic. Peltarian Synapse software was used to develop the neural network models based on minimum number of input variables. Peltarian Synapse has the attractive feature of genetic algorithm (GA) to optimize the network parameters such as number of hidden layers, number of neurons in each layer and training parameters. Also, SPSS software was used for MLR model development.

Several error estimation methods were considered. The mean square error (MSE) (Eq. (1)) was an error indicator that has frequently been used in modelling studies. MSE proved very valuable for comparing the models developed because it indicated the ability of a network to predict the most accurate output. The mean square error can be written as 1  (ti − zi )2 2N N

MSE =

(1)

i

where ti and zi are the actual and predicted outputs for the ith training pattern, and N is the sample size (Catalina, Iordache, & Caracaleanu, 2013; Samarasinghe, 2007). An additional error estimation employed then was the root mean square error (RMSE) method, which showed the unit error in the actual and predicted data. 3. Results and discussion 3.1. Multiple linear regression models Multiple linear regression models (MLR) have previously been used extensively in energy-modelling projects (Athienitis, Kalogirou, & Candanedo, 2012; Catalina et al., 2008; Catalina et al., 2013; Gugliermetti et al., 2004; Korolija, Zhang, MarjanovicHalburd, & Hanby, 2013; Lam, Wan, Liu, & Tsang, 2010; Neto and Fiorelli, 2008; Suganthi and Samuel, 2012). Compared with nonlinear models, MLR models provided an understandable, and more practical, means of solving a variety of issues (Catalina et al., 2013).

110

M. Safa et al. / Sustainable Cities and Society 29 (2017) 107–117

Fig. 2. Actual and predicted energy usage using the MLR models.

Fig. 3. Correlation between the actual and predicted energy usage using the MLR models (Training data).

M. Safa et al. / Sustainable Cities and Society 29 (2017) 107–117

111

For the model in this study, it was essential to choose independent variables without any selective bias (Safa and Samarasinghe, 2011a). The final model was then designed, based on the highest r2 . In the final model the terms should be linked significantly at p = 0.05 (Alvarez, 2009). In the first step, the correlation between energy usage and each independent variable was investigated with the use of a basic MLR technique, using r2 as the decision standard. A MLR model (Eq. (2)) was then applied to predict energy usage, as follows: Y = a0 + a1 V1 + a2 V2 + ... + an Vn

(2)

where a0 -an are the regression coefficients, and V0 -Vn are the independent variables. As shown in Figs. 2–4, the final MLR models of the Christchurch, Dunedin, and Palmerston North sites were applied based on 30 months (2.5 years) of monthly data; due to lack of data the MLR model of the Hamilton site was based on 1.5 years of monthly data. The Canterbury earthquake on 4 September 2010 significantly decreased energy consumption in Christchurch for two months; therefore, the data from the affected months (September and October 2010) have not been used in the model. In addition, because the earlier historical energy usage information for the Hamilton site was unavailable, the Hamilton model began with data for July 2009, rather than for July 2008, as with the other models. Also, one year data was applied as validation data to investigate the models performance. It was notable that when the MLR models for all sites were based only on the average outside temperature, they looked accurate, and had a high correlation coefficient between their actual and predicted energy consumptions. However, when the MLR models with one input variable (temperature) were subsequently compared with models based on two input variables (FTEs and temperature), the results revealed that including FTEs can improve model accuracy through better correlation coefficients with respect to the actual and predicted energy usages. As it was expected the r and r2 of training data (Fig. 3) is higher than validation data (Fig. 4) in the same sites. However, the high correlation between predicted and actual data in validation data confirm the capability of the MLR models to predict energy usage in investigated office buildings. 3.2. Artificial neural network models In this study, after a number of learning algorithms, transfer functions and network structures were tested, different multilayer perceptron (MLP) networks with two hidden layers were also developed. Due to its better performance than other gradient descent methods, the Quickprop learning method was employed in the final model. A variety of functions were investigated: hyperbolic tangent (tanh), linear, logistic sigmoid, Gaussian, and sine. A combination of a logistic (Eq. (3)), a hyperbolic tangent (Eq. (4)) and a linear functions, were variously applied in the final ANN models. These functions are expressed as follows: tanh(u) =

1 + e−u 1 − e−u

L(u) = [1 + e−u ]

Fig. 4. Correlation between the actual and predicted energy usage using the MLR models (Validation data).

−1

(3) (4)

where tanh(u) is the hyperbolic tangent function, L(u) is the logistic function; and u is the weighted sum of the values input into a neuron (Samarasinghe, 2007). To determine the optimum model with a minimum margin of error, different transfer functions were investigated relative to the structure, as described above (Table 1). For example, for the Palmerston North site, a hyperbolic tangent (tanh) function was applied to the first hidden layer and a logistic function was selected for the

112

M. Safa et al. / Sustainable Cities and Society 29 (2017) 107–117

Fig. 5. Structure of the two MLP layers for the Palmerston North site.

Table 1 Artificial neural network components. First Hidden Layer

Palmerston North Christchurch Dunedin Hamilton

Second Hidden Layer

Iteration

Number of neurons

Function

Number of neurons

Function

14 20 19 16

Hyperbolic tangent Linear Linear Linear

13 18 18 7

Logistic Hyperbolic tangent Logistic Hyperbolic tangent

second hidden layer (Fig. 4). The number of neurons in each layer was optimized using a genetic algorithm optimizer that produced a different number for each site (Fig. 5). Figs. 6 and 7 reveal a high degree of correlation between the predicted and actual data using the ANN models. As can be seen, the predicted and actual energy usage patterns fitted almost exactly. The main gap was for September and October 2010 in the Christchurch results, which represented the data removed from the models. The ANN models were, thus, capable of predicting energy usage extremely well. After developing the prediction models the accuracy of the each model had been examined through using 12 months data as the validation data. Comparing the validation and training data (Figs. 6 and 7) show the ANN model are the capable tools to predict energy usage in buildings under different conditions. Results of the both MLR and ANN models indicated that monthly energy consumption in the buildings investigated can be estimated

402 923 2213 102

from the occupancy and temperature data. Energy consumption depended, primarily, on the weather conditions. The low usage in summer was mostly attributable to the average air temperatures and to the Christmas and New Year holidays, which occur during the summer in the Southern Hemisphere. It was clearly evident; however, that energy consumption in these buildings could also be influenced by other parameters (Fig. 8). Figs. 2 and 6 show that, with both models, the predicted and actual data were closely coordinated for most months; however, the different numbers of days and holidays per month, as well as other factures, can influence the predictions produced by the models. The final RMSEs for the MLR results of training data were calculated as 2752, 9508, 1905, and 1412 kW&903;h for the Christchurch, Palmerston North, Hamilton, and Dunedin research office buildings, respectively, (Table 2), which were 36%, 5 3%, 50%, and 56%, respectively, higher than the RMSEs for the ANN model results for

Table 2 Model accuracy, with MSE and RMSE (kW&903;h) for training (T) and validation (V) data. R2

MSE

MLR

Palmerston North Christchurch Dunedin Hamilton

ANN

RMSE

MLR

ANN

MLR

ANN

V

T

V

T

V

T

V

T

V

T

V

T

0.88 0.78 0.84 0.61

0.89 0.81 0.79 0.76

0.90 0.81 0.84 0.78

0.96 0.94 0.89 0.95

13857862 411498798 1887930 5251713

7571201 90405451 3627612 1993101

13037957 38426668 4597969 3981211

3046003 19216085 917914 379401

3722 20285 1374 2292

2752 9508 1905 1412

3611 6199 2144 1995

1745 4383 958 616

M. Safa et al. / Sustainable Cities and Society 29 (2017) 107–117

Fig. 6. Comparison of the actual and predicted energy usage using the ANN models.

113

114

M. Safa et al. / Sustainable Cities and Society 29 (2017) 107–117

Fig. 7. Relationship between the actual and predicted energy consumptions using ANN models (Training).

M. Safa et al. / Sustainable Cities and Society 29 (2017) 107–117

115

the same sites. Correspondingly, the RMSE’s of validation data of ANN models are less than MLR models for the same sites. Fig. 3 and Table 2 show that the MLR models could be fitted to the energy usage for Palmerston North, Christchurch, Dunedin and Hamilton, accounting for about 89%, 81%, 79%, and 76% of the variance, respectively; however, the ANN models could predict energy consumption of about 96%, 94%, 89%, and 95%, respectively, of the variance for the same sites (Fig. 7). Comparison of the modelling methods showed that the data predicted by the ANN models were more accurate than those produced by the MLR models. The ANN models were powerful tools for predicting the correct levels of building energy usage with a minimal margin of error, especially for sites where the data were insufficient. The differences between the models demonstrated the difficulty of developing a reliable model estimate, even for similar constructions. Nevertheless, the development of the models was simple and produced an accurate comparison of the actual and predicted energy usage as well as useful estimates for energy auditing in buildings. Compared with complex nonlinear modelling methods, the final MLR models were very simple (two-variable regression models) and even worked well in the MS Excel program. The MLR modelling method can be used by people with basic technical knowledge, such as, property managers, building owners, and consultants, to audit energy consumption in the buildings being investigated. The hypothesis that formed the basis of this study can also be applied to other similar energy management projects. Based on the mentioned MLR models in this study, several additional MLR models have since been developed for forecasting energy usage in other types of commercial buildings, such as swimming pools and libraries, where the data included outside temperatures and the number of visitors.

4. Conclusion

Fig. 8. Correlation between the actual and predicted energy usage using ANN models (Validation).

This paper presented a modelling method for predicting energy consumption to estimate energy savings in energy management projects, based on separate investigations using MLR and ANN models. Both modelling methods were suitable tools to estimate energy usage and energy savings during energy auditing projects, especially for online energy monitoring projects. This study showed it was possible to develop accurate energy prediction models for office buildings without using construction elements and with few input variables. The results revealed that a MLR model with simple input variables can estimate energy consumption within an acceptable error range. The MLR models were very simple and understandable for people who lacked a technical background. However, it appeared from the data that the ANN models were more effective and accurate at predicting energy usage with only a minimal margin of error. The primary challenge with this method was finding accurate data and the right input variables over an acceptable time period. The results of the study indicated that each building should be considered as an individual project, as many independent factors, such as environmental factors, climate, architectural design, could change the prediction models. Even similarly designed and age buildings managed by the same company could have different prediction models. Outside temperature was a critical input variable; however, finding other input variables will improve the accuracy of energy usage predictions. Based on the results of this study, a precise investigation using historical data was recommended. For example, the Christchurch 2010 earthquake resulted in significantly reduced energy consumption during September 2010, which would have affected the final results of the models. Energy usage should also be considered and analysed for each building individually. When the information from the Hamilton site was compared with that from the other

116

M. Safa et al. / Sustainable Cities and Society 29 (2017) 107–117

sites, it was evident that a longer data set can provide superior estimates. It appeared that long term and accurate data can improve the prediction of energy usage. In the future, the models developed in this study have the potential to be widely used to monitor energy consumption, estimate energy savings and facilitate energy audit projects. As additional data become available, the method introduced in this study will also be more widely applicable for use in other sectors, such as commercial buildings and industrial facilities. References Abrahamse, W., Steg, L., Vlek, C., & Rothengatter, T. (2005). A review of intervention studies aimed at household energy conservation. Journal of Environmental Psychology, 25(3), 273–291. Ahmad, A. S., Hassan, M. Y., Abdullah, M. P., Rahman, H. A., Hussin, F., Abdullah, H., et al. (2014). A review on applications of ANN and SVM for building electrical energy consumption forecasting. Renewable and Sustainable Energy Reviews, 33(0), 102–109. Alvarez, R. (2009). Predicting average regional yield and production of wheat in the Argentine Pampas by an artificial neural network approach. European Journal of Agronomy, 30(2), 70–77. Athienitis, A., Kalogirou, S. A., & Candanedo, L. (2012). 3.11 – modeling and simulation of passive and active solar thermal systems. In A. Sayigh (Ed.), Comprehensive renewable energy (pp. 357–417). Oxford: Elsevier. Azhar, S., Brown, J., & Sattineni, A. A. (2010). Case study of Building performance analyses using building information modelling. In 27th international symposium on automation and robotics in construction (ISARC 2010) (pp. 213–222). Brandon, G., & Lewis, A. (1999). Reducing household energy consumption: a qualitative and quantitative field study. Journal of Environmental Psychology, 19(1), 75–85. Catalina, T., Virgone, J., & Blanco, E. (2008). Development and validation of regression models to predict monthly heating demand for residential buildings. Energy and Buildings, 40(10), 1825–1832. Catalina, T., Iordache, V., & Caracaleanu, B. (2013). Multiple regression model for fast prediction of the heating energy demand. Energy and Buildings, 57, 302–312. Cheung, H., & Braun, J. E. (2016). Empirical modeling of the impacts of faults on water-cooled chiller power consumption for use in building simulation programs. Applied Thermal Engineering, 99, 756–764. Considine, T. J. (2000). The impacts of weather variations on energy demand and carbon emissions. Resource and Energy Economics, 22(4), 295–314. Daly, D., Cooper, P., & Ma, Z. (2014). Understanding the risks and uncertainties introduced by common assumptions in energy simulations for Australian commercial buildings. Energy and Buildings, 75, 382–393. Deb, C., Eang, L. S., Yang, J., & Santamouris, M. (2016). Forecasting diurnal cooling energy load for institutional buildings using Artificial Neural Networks. Energy and Buildings. Diakaki, C., Grigoroudis, E., & Kolokotsa, D. (2008). Towards a multi-objective optimization approach for improving energy efficiency in buildings. Energy and Buildings, 40(9), 1747–1754. Erkoreka, A., Garcia, E., Martin, K., Teres-Zubiaga, J., & Del Portillo, L. (2016). In-use office building energy characterization through basic monitoring and modelling. Energy and Buildings, 119, 256–266. Ghiaus, C. (2006). Experimental estimation of building energy performance by robust regression. Energy and Buildings, 38(6), 582–587. ˜ J. M. (2005). Prediction of hourly energy González, P. A., & Zamarreno, consumption in buildings based on a feedback artificial neural network. Energy and Buildings, 37(6), 595–601. Gugliermetti, F., Passerini, G., & Bisegna, F. (2004). Climate models for the assessment of office buildings energy performance. Building and Environment, 39(1), 39–50. Gustavsson, L., & Joelsson, A. (2010). Life cycle primary energy analysis of residential buildings. Energy and Buildings, 42(2), 210–220. Harish, V. S. K. V., & Kumar, A. (2016a). A review on modeling and simulation of building energy systems. Renewable and Sustainable Energy Reviews, 56, 1272–1292. Harish, V. S. K. V., & Kumar, A. (2016b). Reduced order modeling and parameter identification of a building energy system model through an optimization routine. Applied Energy, 162, 1010–1023. Heinzow, T., & Tol, R. S. J. (2003). Prediction of crop yields across four climate zones in Germany: an artificial neural network approach. Hamburg: Centre for Marine and Climate Research, Hamburg University. Hornik, K., Stinchocombe, M., & White, H. (1989). (5). Multilayer feedforward networks are universal approximators (vol. 2) Oxford, UK: Elsevier Science Ltd. Isaacs, N., Saville-Smith, K., Babylon, M., Bishop, R., Camilleri, M., Donn, M., et al. (2010). Building energy end-Use study (BEES) year 3. BRANZ. Jebaraj, S., & Iniyan, S. (2006). A review of energy models. Renewable and Sustainable Energy Reviews, 10(4), 281–311. Johnsen, T. A. (2001). Demand, generation and price in the Norwegian market for electric power. Energy Economics, 23(3), 227–251. Juan, Y.-K., Gao, P., & Wang, J. (2010). A hybrid decision support system for sustainable office building renovation and energy performance improvement. Energy and Buildings, 42(3), 290–297.

Karatasou, S., Santamouris, M., & Geros, V. (2006). Modeling and predicting building’s energy use with artificial neural networks: methods and results. Energy and Buildings, 38(8), 949–958. Kavousian, A., Rajagopal, R., & Fischer, M. (2013). Determinants of residential electricity consumption: using smart meter data to examine the effect of climate, building characteristics, appliance stock, and occupants’ behavior. Energy, 55, 184–194. Korolija, I., Marjanovic-Halburd, L., Zhang, Y., & Hanby, V. I. (2013). UK office buildings archetypal model as methodological approach in development of regression models for predicting building energy consumption from heating and cooling demands. Energy and Buildings, 60, 152–162. Korolija, I., Zhang, Y., Marjanovic-Halburd, L., & Hanby, V. I. (2013). Regression models for predicting UK office building energy consumption from heating and cooling demands. Energy and Buildings, 59, 214–227. Lam, J. C., Wan, K. K. W., Liu, D., & Tsang, C. L. (2010). Multiple regression models for energy use in air-conditioned office buildings in different climates. Energy Conversion and Management, 51(12), 2692–2697. Li, X., & Wen, J. (2014). Review of building energy modeling for control and operation. Renewable and Sustainable Energy Reviews, 37, 517–537. Li, H., Liu, X., & Fang, G. (2010). Preparation and characteristics of n-nonadecane/cement composites as thermal energy storage materials in buildings. Energy and Buildings, 42(10), 1661–1665. Lund, H., Marszal, A., & Heiselberg, P. (2011). Zero energy buildings and mismatch compensation factors. Energy and Buildings, 43(7), 1646–1654. Ma, Z., Cooper, P., Daly, D., & Ledo, L. (2012). Existing building retrofits: methodology and state-of-the-art. Energy and Buildings, 55(0), 889–902. Ma, J., Qin, J., Salsbury, T., & Xu, P. (2012). Demand reduction in building energy systems based on economic model predictive control. Chemical Engineering Science, 67(1), 92–100. Magnier, L., & Haghighat, F. (2010). Multiobjective optimization of building design using TRNSYS simulations, genetic algorithm, and Artificial Neural Network. Building and Environment, 45(3), 739–746. Mathews, E. H., Botha, C. P., Arndt, D. C., & Malan, A. (2001). HVAC control strategies to enhance comfort and minimise energy usage. Energy and Buildings, 33(8), 853–863. Mba, L., Meukam, P., & Kemajou, A. (2016). Application of artificial neural network for predicting hourly indoor air temperature and relative humidity in modern building in humid region. Energy and Buildings, 121, 32–42. Menezes, A. C., Cripps, A., Bouchlaghem, D., & Buswell, R. (2012). Predicted vs: actual energy performance of non-domestic buildings: using post-occupancy evaluation data to reduce the performance gap. Applied Energy, 97(0), 355–364. Moral-Carcedo, J., & Vicéns-Otero, J. (2005). Modelling the non-linear response of Spanish electricity demand to temperature variations. Energy Economics, 27(3), 477–494. Mukherjee, S., Birru, D., Cavalcanti, D., Das, S., Patel, M., Shen, E., et al. (2010). Closed loop integrated lighting and daylighting control for low energy buildings. In ACEEE summer study on energy efficiency in buildings. Neto, A. H., & Fiorelli, F. A. S. (2008). Comparison between detailed model simulation and artificial neural network for forecasting building energy consumption. Energy and Buildings, 40(12), 2169–2176. Pérez-Lombard, L., Ortiz, J., & Pout, C. (2008). A review on buildings energy consumption information. Energy and Buildings, 40(3), 394–398. Pandharipande, A., & Caicedo, D. (2011). Daylight integrated illumination control of LED systems based on enhanced presence sensing. Energy and Buildings, 43(4), 944–950. Pardo, A., Meneu, V., & Valor, E. (2002). Temperature and seasonality influences on Spanish electricity load. Energy Economics, 24(1), 55–70. Ramesh, T., Prakash, R., & Shukla, K. K. (2010). Life cycle energy analysis of buildings: an overview. Energy and Buildings., 42(10), 1592–1600. Roche, P., & Milne, M. (2005). Effects of combining smart shading and ventilation on thermal comfort. In Solar energy national conference. Rubinstein, F., Neils, D., & Colak, N. (2001). Dimming daylighting, and the electricity crisis in california. In Invited paper for the 2001 IESNA national conference. Sözen, A. (2009). Future projection of the energy dependency of Turkey using artificial neural network. Energy Policy, 37(11), 4827–4833. Safa, M., & Allen, J. (2014). Safa M. predicting energy usage using historical data and linear models. In 31 st ISARC (pp. 531–535). Safa, M., & Samarasinghe, S. (2011a). Determination and modelling of energy consumption in wheat production using neural networks: a case study in Canterbury province, New Zealand. Energy, 36(8), 5140–5147. Safa, M., & Samarasinghe, S. (2011b). Determination and modelling of energy consumption in wheat production using neural networks: a case study in Canterbury province, New Zealand. Energy, 36(8), 5140–5147. Samarasinghe, S. (2007). Neural networks for applied sciences and engineering: from fundamentals to complex pattern recognition. Boca Raton, FL: Auerbach. Santamouris, M. (2013). Energy and climate in the urban built environment. Routledge. Sholahudin, S., & Han, H. (2016). Simplified dynamic neural network model to predict heating load of a building using Taguchi method. Energy. Suganthi, L., & Samuel, A. A. (2012). Energy models for demand forecasting—A review. Renewable and Sustainable Energy Reviews, 16(2), 1223–1240. Thomsen, K. E., Schultz, J. M., & Poel, B. (2005). Measured performance of 12 demonstration projects—IEA Task 13 advanced solar low energy buildings. Energy and Buildings, 37(2), 111–119.

M. Safa et al. / Sustainable Cities and Society 29 (2017) 107–117 Thyholt, M., & Hestnes, A. G. (2008). Heat supply to low-energy buildings in district heating areas: analyses of CO2 emissions and electricity supply security. Energy and Buildings, 40(2), 131–139. Üc¸tu˘g, F. G., & Yükseltan, E. (2012). A linear programming approach to household energy conservation: efficient allocation of budget. Energy and Buildings, 49(0), 200–208. Vakiloroaya, V., Su, S. W., & Ha, Q. P. (2011). HVAC integrated control for energy saving and comfort enhancement. In Int symp automation and robotics in construction (pp. 245–250). Wang, W., Zmeureanu, R., & Rivard, H. (2005). Applying multi-objective genetic algorithms in green building design optimization. Building and Environment, 40(11), 1512–1525. Xu, H., & Niu, J. (2006). Numerical procedure for predicting annual energy consumption of the under-floor air distribution system. Energy and Buildings, 38(6), 641–647. Xue, Y., & Chen, Q. (2014). Influence of floor plenum on energy performance of buildings with UFAD systems. Energy and Buildings, 79, 74–83. Yang, J., Rivard, H., & Zmeureanu, R. (2005). On-line building energy prediction using adaptive artificial neural networks. Energy and Buildings, 37(12), 1250–1259.

117

Yilmaz, Z., & Basak Kundakci, A. (2008). An approach for energy conscious renovation of residential buildings in Istanbul by Trombe wall system. Building and Environment, 43(4), 508–517. Yu, Z., Haghighat, F., Fung, B. C. M., & Yoshino, H. (2010). A decision tree method for building energy demand modeling. Energy and Buildings, 42(10), 1637–1646. Yu, Z., Fung, B. C., & Haghighat, F. (2013). . pp. 207–222. Extracting knowledge from building-related data—A data mining framework. Conference Extracting knowledge from building-related data—A data mining framework (Vol. 6) Springer. Yu, Z., Haghighat, F., & Fung, B. C. M. (2016). Advances and challenges in building engineering and data mining applications for energy-efficient communities. Sustainable Cities and Society, 25, 33–38. Zhang, J., & Haghighat, F. (2010). Development of Artificial Neural Network based heat convection algorithm for thermal simulation of large rectangular cross-sectional area Earth-to-Air Heat Exchangers. Energy and Buildings, 42(4), 435–440. Zhang, Y. P., Lin, K. P., Yang, R., Di, H. F., & Jiang, Y. (2006). Preparation, thermal performance and application of shape-stabilized PCM in energy efficient buildings. Energy and Buildings, 38(10), 1262–1269.