1. Introduction to Residential Building Energy
Due to the rapid growth in economy and population during last few decades, the consumption of electrical energy has been rapidly increasing day by day [
1]. In 2018, the International Energy Agency (IEA) reported that most of the electrical energy is spent in residential buildings, and the demand for energy is rising every year due to excess usage of energy appliances as shown in
Figure 1. Existing studies reveal that residential buildings consumed more energy in summer and winter seasons, which totally depends on the building architecture and occupied area [
2].
Internal and external environment temperatures also have an effect on the total energy consumption in a building [
3]. Therefore, precise prediction of HL and CL is important in order to provide a luxurious life for occupants [
4,
5]. HL is described as the total amount of required heat energy to keep the room temperature normal, while CL is the sum of thermal energy necessary to be eliminated from a cooling area in order to keep the temperature at an appropriate level [
2].
Before evaluating the thermal load, it is important to know the infrastructure of buildings, because energy consumption is reliant on their physical attributes. Basically, four tools are employed to predict the CL and HL of buildings: simulation modeling, engineering calculations, statistical models, and ML models [
6]. The simulation model is commonly used to simulate energy efficiency based on prior information, but it is a very difficult and time-consuming model because it requires more skill to operate. For instance, Bagheri et al. [
7] considered the simulation methods in terms of its applications and limitations in the domain of energy performance. The second tool utilized complicated mathematical formulas according to its principles to efficiently predict energy load. Next, a statistical tool is used to evaluate linear regression models for residential energy consumption prediction, and later, the performance of the model is enhanced by modifying different parameters. The final tool is ML, which is a subset of statistical techniques, but it has the potential to learn from real data and predict the desired outputs. Further, it assists civil engineers in evaluating the ingredients used in the building design. For instance, support vector regression (SVR), clustering, and Gaussian-based regression are active ML approaches in energy predictions [
6].
ML algorithms can be broadly categorized into two main groups (i.e., supervised and unsupervised) based on diverse learning style. The predicted output variables are available in the case of supervised learning, while the unique labeled output does not occur in an unsupervised learning strategy. The current study focuses on a supervised learning approach because an energy efficiency dataset has labeled data. Artificial neural networks (ANNs) have gained attention among supervised learning techniques due to nonlinear relationships within the data. Moreover, the activation function of ANN can predicted the desired outputs, which indicate the nonlinearity with various input attributes [
6]. Numerous ANN architectures, including recurrent networks, radial basis function, and feedforward, are used for energy prediction. Besides ANN, researchers have mostly implemented the multilayer perceptron (MLP) model, where information flows in a single direction with multiple layers. The MLP model comprises three basic layers (input, hidden, output) consisting of neurons with weighted functions. In case of complex data processing, the existing model is altered by increasing the number of neurons and hidden layers.
Managing huge and complicated energy consumption data is formidable for ANN, while researchers have criticized this network due to low transparency in the model [
8]. Sensitivity analysis (SA) is broadly applied to analyze the relationship between variables. For precise energy forecasting, ANN shows better performance if irrelevant inputs are removed [
9]. Sensitivity analysis about the mean (SAAM) is one of the conventional strategies, where changes of dependent variables are recorded while independent variables are kept in a specified range by computing the mean [
10]. The key benefits of SAAM are simple interpretation, easy implementation, and application, along with statistical analysis [
11]. In addition, state-based sensitivity analysis is a global SA method in which separate variables are varied independently and the rest of the variables are changed concurrently to obtain the reliant attributes [
8].
In residential buildings, there are various factors that influence energy consumption, such as consumer’s behavior and building architecture. Therefore, building-structure-related data play a key role in developing an efficient energy model. Moreover, the height of buildings, construction materials, and areas such as wall, roof, and glazing are the main attributes in the current research. The simulated method in [
12] performs a pivotal part in improving building constructions, and it can also accurately depict real assessments of different building designs to predict HL and CL [
13]. On the other hand, most of the researchers get full advantages by applying DL models on different domains, such as movie and video summarization [
14,
15], energy forecasting [
16], biological data analysis [
17], violence detection [
18,
19], and action recognition [
20]. In this study, we explored numerous ML and DL models for the prediction of HL and CL using an energy efficiency dataset. The potential of sequential models for this dataset has not been thoroughly explored till date. Therefore, GRU has an optimal preference to predict HL and CL as there exists an intensely independent relationship between data. We conduct two types of experiments. First, we enhance the existing performance in which HL and CL are predicted separately. Second, a multi-output prediction is performed through the same architecture. The relevance of this work can enable engineers to solve major structural issues when designing an energy-efficient building.
There is no existing work that utilized GRU for this dataset till date. Therefore, in the current study, we utilize the sequence learning model GRU for non-sequential data by examining various parameters. The second limitation is the unavailability of preprocessing methods, including polynomial and min–max normalization for HL and CL. In this study, the simulation data first pass through a preprocessing step where outliers are removed, scattered data is normalized in a specific range, and increase the number of features. Next, the refined data are fed into the GRU network to extract silent hidden patterns. Finally, we evaluate the error in different metrics, such as mean absolute error (MAE), relative mean absolute error (rMAE), mean square error (MSE), relative mean square error (rMSE), root mean square error (RMSE), relative root mean square error (rRMSE), mean average percentage error (MAPE), and relative mean average percentage error (rMAPE). The major contributions of this study are summarized below:
It is a common fact that the performance of a deep model is directly depends on the input data. In this study, energy efficiency dataset is used that contains a limited number of attributes with values in a different range, which cause overfitting and take extra time to converge. To address these issues, first, we pass the input data through a preprocessing layer where the number of features increased using a polynomial equation and min–max normalization process is applied to remove outliers and normalize the data in a particular range.
Existing models in the literature are trained separately for HL and CL prediction, which requires a tedious and time-consuming job. In contrast, the proposed framework has a generalized ability in which the same architecture can be used for both SO and MO that predict HL and CL concurrently.
DL models always reveal a convincing performance compared with traditional ML models. Therefore, we propose a sequence learning model GRU, which learns discriminative features and efficiently predicts the HL and CL. We also conduct a comparative study between ML and DL techniques to show the superiority of DL models.
We verify experimentally that the proposed framework outperforms state-of-the-art techniques using the hold-out and 10-fold methods. To check the effectiveness of the proposed framework, we evaluateit on various metrics, such as MAE, rMAE, MSE, rMSE, RMSE, and rRMSE.
The rest of the paper is categorized into four main sections.
Section 2 briefly discusses the literature study about HL and CL prediction.
Section 3 explains the proposed methodology, followed by comprehensive experiments in
Section 4.
Section 5 concludes this study with future research direction.
2. Literature Review of HL and CL Prediction
The literature study for HL and CL prediction in buildings is mainly divided into four major classes: residential, educational, commercial, and mixed. According to statistics in [
21], 30% of the literature is based on residential building energy. Through Ecotect software, Tsanas and Xifara [
12] simulated 12 distinct building structures to predict HL and CL. After considering all the various permutations of input variables, 768 building designs were generated. During the simulation of building designs, heating, ventilation, and air-conditioning HVAC rules were pursued. Through numerous ML techniques, various researchers analyze these data for precise prediction. Based on the prominent contribution of Tsanas and Xifara [
12], the existing literature is summarized in
Table 1. Although the dataset has been prepared via a simulated tool, but there is lack of data related to building infrastructure and materials. The dataset used in this study is publicly accessible and extensively used for research study by exploring its applications related to energy. Simulated data play a significant role when designing the architectures of a building. The terms used in the existing studies are listed in
Table 1.
Tsanas and Xifara [
12] conducted a detailed statistical study of density and scatterplots. The performance outcomes of the statistical analysis approach are mainly used for nonlinear problems. From
Table 1, it can be observed that few studies have applied ANN on energy efficiency dataset [
24,
32], although others follow the ensemble strategy by integrating different methods [
22,
25,
29,
31]. To the best of our knowledge, only one article exists that applied a deep neural network (DNN) to predict HL and CL, presented by Sekha et al. [
4]. The efficiency of DNN is better as compared with other traditional algorithms, such as Minimax Probability Machine Regression (MPMR) and Gaussian Process Regression (GPR). Moreover, the traditional approaches did not mention the model parameters, such as processing elements, activation functions, and numbers of layers. To achieve a remarkable performance on any models, analysis of data is essential to identify the significant and insignificant inputs. In this regard, Roy et al. [
2] proposed a nonparametric regression model known as Multivariate Adaptive Regression Splines (MARS) that splits the data and fit each interval into a basis function. Principal component analysis (PCA) is also applied for ideal features selection and dimensionality reduction, which eradicates the multilinearity problem. Nilashi et al. [
31] reported that PCA targets four main aspects: retrieving essential information, reducing the dimension of data, simplifying the information, and analyzing architecture-related observations. Most of the articles did not utilize the SA approach for the prediction of HL and CL as shown in
Table 1.
The techniques for quantitative SA are classified into local and global [
37]. For instance, the input instances were impartial with each other; therefore, Ardjmand et al. [
8] defined a regression-based strategy in which conventional SA techniques, such as sampling, regression-based, and variance-based, are expanded to state-based sensitivity analysis (SBSA). This means that modifying one variable value would influence the others; therefore, it is not realistic for fixed values for certain inputs in local SAs. On the other hand, the global SA adjusts the ideal input value, while in the case of multidimensionality, it takes the average number of variable inputs [
38].
The efficiency of a mathematical model is also influenced by many assumptions in order to predict energy HL and CL separately. In majority of the works, the HL and CL are predicted in an SO fashion; however, we develop such a model that can be utilized for both SO and MO. Another primary consideration for enhancing the efficiency of a predictive model is preprocessing of data. Therefore, Kumar et al. [
35] followed the ensemble technique with a proper attribute selection and preprocessing method to efficiently predict energy in real time. To boost the model efficiency, it is necessary to pass the data through the preprocessing stage. Notably, MSE, MAE, RMSE, and MAPE were common evaluation metrics used by researchers for model assessment, but in this research, we also use extra metrics for evaluation, including rMAE, rMSE, rRMSE, rMAPE.