BIM-Based Machine Learning Application for Parametric Assessment of Building Energy Performance

Abstract

The energy performance of buildings has become a main concern globally in response to increased energy demand, the environmental impacts of energy production, and the reality of energy poverty. To improve energy efficiency, proper building design should be secured at the early design phase. Digital tools are currently available for performing energy assessment analyses and can efficiently handle complex and technically demanding buildings. However, alternative designs should be checked individually, and this makes the process time-consuming and prone to errors. Machine learning techniques can provide valuable assistance in developing decision support tools. In this paper, typical residential buildings are considered along with eleven factors that highly affect energy performance. A dataset of 337 instances of such parameters is developed. For each dataset, the building energy performance is estimated based on BIM analysis. Next, statistical and machine learning techniques are implemented to provide artificial models of energy performance. They include statistical regression modeling (SRM), decision trees (DTs), random forests (RFs), and artificial neural networks (ANNs). The analysis reveals the contribution of each factor and highlights the ANN as the best performing model. An easy-to-use interface tool has been developed for the instantaneous calculation of the energy performance based on the independent parameter values.

1. Introduction

The increase in population and human activities, followed by the expansion of the built environment, have intensified sustainability problems, among which is the need for significant amounts of energy in contemporary buildings. According to the Global Status Report for Buildings and Construction [1], buildings consume about 36% of annual global energy resources and are responsible for 60% of global electricity production, whilst they contribute 39% of global annual CO₂ emissions.

To facilitate residents’ well-being, modern building design requires increased energy consumption. In addition, existing building structures are highly energy-intensive and environmentally unfriendly [2]. Previous studies and construction field experience have demonstrated that energy consumption in buildings can be effectively reduced through simple design solutions, such as the selection of appropriate building shapes and orientations. Thus, one of the challenges of the construction industry is the sustainable design of new buildings and adaptation of existing ones. In this direction, it is imperative to reassess the conventional design and evaluation procedures of structures, particularly through an integrated parametric approach that considers interdependences among leverage factors. In current practice, energy analyses are often conducted separately, as the planning and dimensioning processes are not aligned with the energy assessment process. In fact, most existing buildings were designed in the past with outdated regulations οr without any sort of provisions concerning their energy efficiency [3]. Additionally, structure complexity, multiple influential parameters and interdependencies, uncertainties, and limited data often result in assumptions and simplifications, which underestimate or distort the influence of these parameters, leading to inaccurate results and inadequate designs.

The energy efficiency of buildings is affected by several factors related to a building’s structure, location, and usage. These factors can be classified into two main classes: physical and human-influenced ones [4]. The first class encompasses technical and physical factors, such as the building envelope, local climate, heating, ventilation, and air conditioning systems. The second category includes human-related factors, such as occupant behavior and energy usage patterns, which exhibit significant variability and uncertainty in energy consumption. Finally, there are social factors, such as energy prices, that have social consequences on energy efficiency.

An effective way to estimate building energy efficiency is the utilization of building energy modeling (BEM) software [5]. However, partially using BEM software without synchronizing it with the global project design and construction process can result in ineffective decisions, which are partly due to discontinuous information flow among digital design models. In addition, the process is time-consuming, tedious, and susceptible to errors, as energy efficiency data must be entered into BEM software manually. The introduction of BIM technology allows for options that can overcome the limitations of conventional energy efficiency assessment processes through BEM, as currently, BEM can act as a subsystem of BIM. The use of BIM can result in accurate and systematic calculations and provide the ability to monitor and control energy use during the operation phase of buildings [6]. The application of BIM technology combined with simulations from the initial stages of project design allow for a wide exploration of alternative design solutions [7]. In fact, this technology provides the ground for understanding and quantifying the actual conditions of a building as well as simulating real-world user behavior, thus improving energy efficiency analyses.

While BIM represents the foundation for performing several types of analyses in building design, these analyses are usually multifactorial and require comparative evaluation of alternative solutions, with user involvement in setting these solutions up and comparing the results. The partial automation of such analyses results in increased time and effort requirements. Instead, one can employ a parametric analysis and use machine learning techniques to easily capture the influence of energy consumption factors and comparatively evaluate alternative building configurations in terms of energy efficiency. The present study first analyzes several factors that predominantly affect the energy efficiency of a typical residential building using BIM technologies. Based on this analysis, a dataset relating the factors influencing energy to their corresponding energy consumption is developed. Machine learning techniques are then developed and evaluated as an instrument for instantaneous energy assessment analyses that can assist in decisions on initial building designs or retrofitting.

2. Background

Existing research has dealt with several aspects of building energy analysis. In the direction of sustainable design, key architectural design variables (ADVs), which influence sustainability characteristics during early building design and integrate computational simulations and stakeholder insights, were examined [8]. This research underscores the importance of including stakeholder perspectives in developing design optimization tools, aligned with real-world needs for effective sustainable design.

The role of BIM in facilitating the LEED^® certification process was explored in [9]. The findings underscore BIM’s potential to save time and resources, thus enhancing the feasibility of sustainable project management. Further, the implementation of national and international energy standards in low-rise residential buildings in Jordan utilizing BIM models was examined in [10]. The use of BIM tools and regression models to optimize energy performance in early-stage building design was demonstrated in [11]. By simulating various prototype building shapes, the study identified energy-efficient geometries, noting that triangular designs consumed the least energy. A framework combining BIM with a life cycle cost (LCC) analysis and orthogonal testing methods to identify optimal design parameters for energy efficiency in green buildings was introduced in [12]. The study revealed significant annual energy reductions, emphasizing enhancements in building envelope features such as insulation. A BIM-driven approach aiming at net-zero energy in tall buildings within the Malaysian construction sector was the focus of [13]. The results demonstrate significant correlations between BIM usage and factors such as early design integration and predictive energy performance.

Moving on to machine learning applications for energy consumption prediction, the effectiveness of deep learning (DL) methods for short-term cooling load prediction in buildings was explored in [14], emphasizing their potential to outperform traditional physic-based models. A comprehensive review of methods for predicting building energy demand, with a focus on data-driven techniques, like support vector machines (SVMs) and artificial neural networks (ANNs), was presented in [15]. The study highlights the strengths of these methods, including achieving a high predictive accuracy. It further notes that their effectiveness can be hindered by limited data availability. Another review of data-driven models for predicting building energy consumption, emphasizing their role in energy management, conservation, and planning, was performed in [16]. The study recommended focusing on big data analytics and incorporating occupant behavior insights to enhance model accuracy and reliability, paving the way for more tailored prediction models. An analysis of building energy prediction models was performed in [17], categorizing them into white-box, black-box, and grey-box types and highlighted that black-box models, which include machine learning techniques, like SVMs and ANNs, can handle complex data relationships. However, they face challenges related to human behavior and climate variability. A hybrid predictive model that combines empirical mode decomposition (CEEMDAN), deep learning, and ARIMA to enhance short-term heating load forecasts was introduced in [18]. The study demonstrated the potential for hybrid models to address complex data and improve forecast reliability, suggesting future research on dynamic weighting strategies to enhance adaptability. Deep transfer learning (DTL) strategies were evaluated in [19] to overcome data scarcity in cross-building energy prediction. The study concluded that this approach is effective for improving prediction accuracy across various building types and highlights the need for further explorations into diverse application contexts. The use of ANNs for national-level hourly heat demand prediction, focusing on simplifying input requirements while maintaining accuracy, was explored in [20]. The study identifies temperature and radiation as critical factors, allowing for reduced input complexity.

Focusing more closely on optimized and hybrid ML approaches, a comprehensive comparison of various machine learning (ML) techniques to predict annual energy consumption in residential buildings was provided in [21], with a particular focus on aiding early design decisions to enhance energy efficiency. The authors advocate for the use of DNN models for early-design-stage energy consumption forecasting and recommend further studies incorporating larger datasets and exploring ensemble algorithms to enhance predictive reliability. A hybrid machine learning model designed to enhance the accuracy of heating energy consumption predictions in residential buildings was presented in [22]. The study analyzed the support vector regression (SVR) and six meta-heuristic algorithms, including battle royale optimization (BRO) and particle swarm optimization (PSO). The authors recommend extending this method to various climates and building types to verify its broader utility. Energy demand forecasting across seven sectors in Iran was explored in [23], employing machine learning algorithms (ANN, LSTM, and ARIMA models), enhanced with optimization techniques (PSO and Grey Wolf Optimizer). The work concludes that this framework provides a foundation for strategic energy planning and management. An AutoML framework for predicting residential heating and cooling loads was introduced in [24], which automates the feature engineering and model optimization process. The study concludes that the use of SHapley Additive exPlanations (SHAP) values in model explainability enhances the framework’s practical utility for early-stage building design. Finally, the integration of evolutionary algorithms with conventional neural networks for residential energy consumption forecasting was proposed in [25], leading to the recommendation of using hybrid techniques for future optimization.

To relate energy analyses and socio-economic factors, artificial neural networks (ANNs) have been utilized as tools for policymakers to identify and assist vulnerable populations, as part of an energy poverty study in Greece, with targeted energy efficiency initiatives and financial aid [26]. The analysis incorporates various socio-economic and geographical factors, including house age and ownership status, to predict seven energy poverty indicators.

The existing literature has dealt with several aspects of the energy efficiency of buildings. However, the analyses mainly focus on fragments of the whole process while the outcomes remain within a rather scientific framework without the appropriate connection to engineering practice and design tools. There are also several research works using ML methods for assessing energy consumption. According to [16], 84% of such studies focus on short-term energy consumption prediction because of its direct relation to the day-to-day operations of buildings, while only 12% of the studies focus on long-term (yearly) energy consumption prediction. The first class refers to energy efficiency improvements, in terms of temperature or lighting, following certain interventions, such as cell insulation upgrades. An indicative study of this kind [27] focuses on an energy simulation of an academic building and aims at evaluating the BIM’s potential to improve the building’s energy efficiency through the simulation and optimization of design parameters, such as building orientation or insulation characteristics, individually or collectively. A long-term energy efficiency goal refers to a multi-parametric evaluation of energy-efficient building design in a structure’s life cycle [3]. Within the latter class, the present work combines BIM modeling and machine learning techniques to evaluate the influencing factors of building energy performance and provides a practical and easy to use tool for energy efficiency analyses of typical residential buildings. The adoption of BIM provides an unlimited capability to simulate the structure and develop energy models for any type and size of building, design characteristics, environmental factors, etc.

3. Methodology

The traditional methods of building energy performance assessment, via analytical calculations, are quite tedious and of low accuracy. This is because they rely on simplified assumptions and limited data that do not fully reflect the linkage of various factors with energy consumption. To address these limitations, a more comprehensive and integrated approach is needed, which considers all relevant factors with their interrelationships. The advancement of building information modeling (BIM) can significantly contribute to energy performance assessment upgrades, the optimal design of new buildings, and the retrofitting of existing ones.

While BIM tools can effectively carry out the desired analysis, the process of comparatively assessing alternative building designs requires the development of each individual BIM design. There are many of these designs and they can be quite different in terms of their characteristics, making the whole process arduous and time-consuming. On the other hand, the existence of a ready-to-use digital solution that could instantly provide an estimate of the building energy performance, based on some prevailing characteristics, without the need to develop each BIM model, could provide a great decision support in designing high-energy-performance buildings. To obtain such a solution, a two-step process is required, which includes the identification of the prevailing parameters and their interrelationships, as well as the energy performance assessment based on those parameters.

In the present study, a set of parameters that influence the building energy performance are developed based on research findings and practical experience. Then, an experiment is performed with alternative sets of these parameters, and the corresponding energy performance is assessed via a BIM application. The input–output set of this process is used to develop artificial prediction models, employing alternative methods of machine learning. A typical group of residential buildings is examined. The energy performance of buildings depends on several parameters with varying levels of contributions. In this work, the ones that have been highlighted in the literature and in practice as having a significant impact on the energy performance of structures are considered. These factors are generally related to the architectural design, the characteristics of the building shell, and the climatic conditions in the area and are in agreement with previous studies [28,29,30]. The analysis involves eleven main parameters that considerably influence the energy performance of such buildings. These parameters are codified in

Table 1. Description of design factors and corresponding values.

	Description	Parameter Values	Model Values
X1	Masonry thickness	15, 20, 25 cm	15, 20, 25
X2	Wall insulation thickness	2, 3, 5 cm	2, 3, 5
X3	Slab thickness	15, 20 cm	15, 20
X4	Slab insulation thickness	2, 3, 5 cm	2, 3, 5
X5	Window-to-wall ratio	15%, 20%, 30%	15, 20, 30
X6	Window glass type	Single, double	1, 2
X7	Climate zone at building location	Zone 1, 2, 3, 4	1, 2, 3, 4
X8	Building orientation	South, North, East, West	1, 2, 3, 4
X9	Roof existence	Yes, No	1, 2
X10	Basement existence	Yes, No	1, 2
X11	Building size	1 floor, 2 floors	1, 2

and schematically illustrated in Figure 1. The location parameter corresponds to four different climatic zones, in terms of the average temperature, as is further described in the case study section.

Energy Efficiency Analysis of Buildings Utilizing BIM

This section describes the methodology to assess the energy efficiency of buildings using building information modeling (BIM) technology. This assessment is based on simulated data that have come from full building modeling. The energy efficiency assessment is performed via the quantitative parameter of energy consumption. The methodology can be used in the early design process of buildings for providing decision support for selecting appropriate features for a design.

The BIM design of a building starts with architectural modeling, in which the building elements with their characteristics (geometric dimensions, materials, supplier info, etc.) are entered in the 3D digital model. Then, the building structural characteristics are entered in the model and the building/shell design is created. Finally, the HVAC (heating, ventilation, and air conditioning) design is integrated. The whole building design process is dynamic and subject to continuous revisions and improvements. The BIM model, due to its high interoperability, is continually updated and provides a powerful model and a vivid picture of the alternative designs to all stakeholders. Figure 2 illustrates the BIM modules and their involvement in building energy modeling. Specifically, 3D-BIM is used for modeling the design characteristics of the building (Figure 2a). The solar path and the sunlight radiance components, which are shown in Figure 2b and Figure 2c, respectively, are tools that determine the sun orbit and the intensity and distribution of natural lighting at the building’s location as well as their effect on energy consumption. Finally, the energy model is a BIM tool that calculates the energy consumption of a building with specific characteristics (Figure 2d).

BIM technology provides significant benefits in interpreting and quantifying the actual condition of a building, while it further allows for the simulation of user behavior. By incorporating energy simulations into the BIM model, alternative designs can be explored, and energy-efficient solutions can be identified. In this direction, BIM-based building energy assessment is crucial in promoting sustainable construction.

In the following analysis, different pieces of terminology are used in terms of the aim and the output variable of the model. “Energy efficiency” refers broadly to the capability of the building to provide comfortable temperature conditions for inhabitants at relatively low energy consumption rates. “Energy consumption” represents the quantitative assessment of the required energy to attain acceptable levels of comfort and represents the model output variable. “Energy estimation or prediction” highlights the process of calculating the energy consumption level of an existing building or forecasting the energy consumption of a newly designed or reconstructed building.

4. Model Development

The energy consumption of a building is affected by several parameters associated with the characteristics of the building and its surroundings. To obtain a generic energy assessment model, the actual building design and energy performance should be analyzed on a parametric basis. This objective can be fulfilled if a representative set of input parameters are developed, and the corresponding energy consumption outputs are computed. Next, a machine learning technique can be employed to develop the corresponding artificial model.

Two general approaches for developing the artificial energy prediction model are examined in this paper. The first includes statistical methods in the form of curve fitting by means of statistical regression analysis (SRM model). The second class includes machine learning techniques, namely decision trees (DT model), random forest (RF model) and artificial neural networks (ANN model). All these methods are particularized as alternatives to energy performance assessments based on BIM modeling and aimed to involve low-effort calculations and provide wide applications without restrictions (Figure 3).

4.1. Multiple Nonlinear Regression Model

The approximation of the relationship between energy consumption and the factors that leverage it on a statistical ground can be elaborated through regression. The use of multiple nonlinear regression requires the selection (or assumption) of the nonlinear relationship type between input and output data. In many cases, however, the use of multiple nonlinear regression is not sufficient (for strong nonlinear problems), as the relationship type is confined within the available software libraries. Stata software allows for switching among different parameter combinations and exponents (Equation (1)), thus facilitating the determination of functional relationships. Equation (1) describes the general form of the SRM model [31], where Y is the output parameter (e.g., in kWh/m² or EUR/m² per year), X is the vector of the independent variables, and B_i and P_i are the model coefficients and exponents.

$$Y=B_0+B_1\cdot X^{p_1}+B_2\cdot X^{p_2}+\dots$$

(1)

4.2. Machine Learning Model

4.2.1. Decision Trees (DTs)

In terms of machine learning techniques, the first method employs decision trees (Figure 4). Decision trees (DTs) provide a graphical representation of problem solving through gradual decisions. They can be used in problems with numerical and classification data, being more effective in resolving problems of the latter type. A decision tree algorithm is a way of creating a tree-like model which makes decisions based on the features of the data. The tree starts with a single node known as the root node, which represents the entire dataset. The root node is then split into two or more child nodes based on the values of one of the features in the dataset. This process is repeated for each child node retrospectively, creating a hierarchical structure of nodes that represents the decision process. The algorithm selects the feature that best separates the data into subsets with similar target values to split the data at each node. This is accomplished by gauging the information gain of each feature. The information gain is a measure of how well a feature separates the data into subsets with similar target values. Eventually, the feature with the highest information gain is chosen to split the data at that node. The process of splitting the data continues recursively until it reaches a stopping criterion, i.e., a set of rules that determine when the algorithm should stop creating new nodes. This could be the maximum tree depth, the minimum number of samples per leaf, or the minimum information gain. Once the tree is built, it can be used to make predictions. Each leaf node of the tree represents a prediction for the target variable. The structure of the tree allows one to represent the decision process and the decision boundary in a readable format.

In the current application, the parameters related to the architectural design, the building frame characteristics, and the climate parameters are set up in a question form and the energy efficiency is set up as a decision output (assessment). These components are combined through a learning process to create a decision tree that can be used to estimate energy costs. However, in many cases, if the decision-making process consists of several steps, the decision trees can grow extensively. Still, due to the ease and supervision they offer, DTs are popular and practical tools in decision-making processes.

4.2.2. Random Forests (RFs)

A combination of DT algorithms is often developed to form a random forest algorithm (RF). Random forests (RFs) use a series of different decision trees combining the individual predictions to enhance the accuracy of the output. Figure 5 indicates such a structure with N decision trees. The RF algorithm is powerful and accurate for both problem classes, i.e., regression and classification.

An RF algorithm builds multiple decision trees (also known as “forests”) and combines their individual predictions to a final decision. A random sample is selected from the dataset for training each decision tree. A random subset is selected for each decision tree for splitting the data at each node. Then each decision tree is trained on its sample of data and selected features. The predictions from all decision trees are finally combined by taking, for instance, the majority of votes for classification problems, or averaging the predictions for regression problems. An advantage of random forests compared to decision trees is the reduced overfitting outcome.

4.2.3. Artificial Neural Network (ANN)

Artificial neural network (ANN) methods are particularly useful in cases in which the cost of implementing an analytical computational process is high or if the way that influencing factors interact is not well known. The ANN structure consists of several input nodes (which refer to the input parameters) and intermediate neurons arranged in layers to process the input data, as well as the output node (Figure 6). The efficiency of an artificial neural network is determined by its architecture and the learning process. The selection of the number of neurons, their layers, and the way that they are connected depend on the problem’s nature and characteristics. ANN learning is the process of acquiring knowledge through training examples, which is then recalled, providing an output in similar problems. Neural network operation is based on the transmission of a signal from the initial neurons (input layer) to the final neuron (output layer). Depending on the neuron connection characteristics, the signal is attenuated or amplified during its propagation, so that the physical quantity in the last neuron is calculated.

When developing machine learning models, a balance between model and actual problem complexities should be retained. In this sense, it is not always effective to use large-scale (several input parameters) or complex (in architecture) models. The employment of such a model does not necessarily guarantee more precise predictions. This is because the increase in size and complexity makes the model hypersensitive [32] to such an extent that it cannot work for out-of-training data due to overfitting. On the contrary, a simple model may exhibit high deviation between simulated and actual output data due to underfitting. The above inefficiencies can be regulated by the two complementary error types, variance and bias, which indicate overfitting and underfitting conditions, respectively. An effective model should retain balance between the two errors and be obtained by a trial-and-error development process.

In summary, four compact models are developed based on non-linear multivariate regression, decision trees, random forests, and artificial neural networks. More specifically, the statistical regression modeling (SRM) approach is employed, utilizing Stata 18 software to develop a non-linear and multivariable regression model. The decision tree model consists of 7 levels and 24 terminal nodes and has been trained using the C4.5 algorithm (https://en.wikipedia.org/wiki/C4.5_algorithm (accessed on 10 December 2024)). The random forest model includes five trees (with the same form as that in DT model) and has also been trained using the C4.5 algorithm. The root mean square error (RMSE) statistic is used as the objective (loss) function. The number of epochs in both models (DT and RF) is in the order of 1000.

A multi-layer perceptron (MLP) architecture is implemented for neural network development. The MLP structure consists of 11 input neurons corresponding to the input variables, a hidden layer of 11 neurons, and an output neuron. An activation function of a sigmoid form is employed for the hidden and the output layers. The specific design efficiently serves as a trade-off between model accuracy and simplicity, following experimentation on alternative designs to prevent overfitting and underfitting issues. The training process has been carried out using the Levenberg–Marquardt algorithm, which is a variant of the back-propagation method. The number of epochs is in the order of 5000.

5. Data Management and Development Software

The model’s effectiveness highly depends on the size and quality of the available data. In this work, the necessary data have been developed from simulation and parametric analysis of buildings through BIM application. The input parameter selection was made in a way that data are rather representative of all eleven design parameters in Table 1 within a reasonable range of values. The generated sample includes 337 arrays of input–output data. The dataset is randomly divided into training, validation, and testing subsets in proportions of 70%, 15%, and 15%, respectively.

According to [33], the training sample size in such type of analysis should be above an approximate level of fifty plus eight times the number of independent parameters. Further, the work in [34] proposes that 10 to 15 observations per parameter are required to avoid overfitting in statistical and machine learning models. In the present work, the training sample size is roughly twice as high as these levels. As such, it can assure that the model complexity (overfitting) is controllable, and the estimates are statistically reliable and representative of the population, thus avoiding underfitting.

The energy consumption assessment is performed using Autodesk Revit and Green Building Studio (GBS) software (https://gbs.autodesk.com/gbs). Revit has been used to develop the building model while GBS provides energy analysis and consumption optimization. The framework for the evaluation of energy assessment and decision making is illustrated in Figure 7.

6. Case Study

6.1. Input Data

A typical residential building is analyzed as part of the case study. The structure consists of two or four apartments, as per the architectural form in Figure 8, over one or two floors. Building type A includes two apartments each with an area of 212 m² that are placed either side by side (single floor) or on top of each other (two-floor configuration). Building type B includes four apartments each with an area of 212 m² that are placed either side by side (single floor) or on top of each other (two-floor arrangement). The exact building formations are shown in Figure 9.

Each apartment is a typical four-person family flat. Rooms are cooled and heated via an HVAC system. Rooms are classified into three common typical thermal zones. The first includes the kitchen and the living room, the second comprises the bedrooms, and the third the corridors and bathrooms. Two thermostats—one for heating and one for cooling—are considered. The set points for heating and cooling are at 24 °C and at 21 °C, respectively. The electric loads (for lighting and equipment) are determined based on the number of residents and defined as Watt (W) per square meter (m²).

Regarding the location parameter (X7), four different climatic zones are considered, based on ambient conditions in Greece and represented by four cities, as shown in Figure 10. These cities are Chania, Patras, Ioannina, and Florina with average annual temperatures of 17.2 °C, 15.5 °C, 12.8 °C, and 10.4 °C, respectively (https://en.climate-data.org).

To determine the impact of each influencing factor, a default parameter setting is considered. Each factor is then altered to another value, and the effect on energy consumption cost is computed. A bar graph that represents the effect of each factor on the annual energy cost is estimated for each building type (Figure 11). The graph displays the percentage change in energy consumption per each factor when shifting from the default to a new value. The default parameter values are shown in the graph legend. Positive values indicate an increase and negative values indicate a reduction in energy consumption. The annual cost estimate is based on a unit energy cost of 0.12 EUR/KWh.

The main observations from Figure 11 are as follows:

• The overall view of the two diagrams indicates that individual factors play a similar role in terms of energy consumption or savings in both building types. A rather flat proportion of 1.85 to 1.00 in terms of total energy consumption is observed in these two cases. This is as expected, as type B buildings are more environment-proof, being more condensed than type A buildings. This outcome may highlight an energy performance pattern for buildings of larger sizes.
• The highest influence is observed in terms of the location and the corresponding climatic conditions (parameter X7). For instance, moving from zone 2 (Patras) to zone 4 (Florina), the energy consumption is increased by 13.37% and 12.36% for building types A and B, respectively.
• The presence of adjacent buildings has a positive effect on energy efficiency. This is observed by comparing building configurations 2 and 4 in Figure 9, with the energy savings in the latter case calculated to be 8.80%.
• The energy consumption is increased in one-story compared to two-story structures (parameter X11) by 7.23% and 6.91% for building types A and B, respectively.
• The existence of a roof (parameter X9) within the basic scenario results in energy savings of 6.24% and 6.23% for building types A and B, respectively.
• The window-to-wall ratio (parameter X5) increases from 15% (in the basic scenario) to 30%, leading to increases in energy consumption of 5.08% and 5.25% for building types A and B, respectively.
• Likewise, the thickness variation in the external wall insulation (parameter X2) from one level to another has an influence of about 3% in all cases. The wall thickness itself plays a marginal role in all cases.
• All other parameters have a contribution of less than 2% to energy consumption.

6.2. Models of Energy Performance Prediction

Equation (2) provides the best-fitting relationship for estimating annual energy consumption as a function of the eleven independent factors, as derived from the fractional polynomials in the statistical regression model (SRM).

y = −0.12 X1 − 0.52 X2 + 0.05 X3 − 0.16 X4 + 3.72 X5 − 0.31 X6
+1.06 X7 − 0.99 X8 + 0.98 X9 − 0.10 X10 − 1.10 X11 + 21.30

(2)

Figure 12 presents an illustrative part (due to size limitations) of the developed decision tree (DT) model in a self-explanatory graphical form. The random forest (RF) and the neural network (NN) cannot be easily rendered in a graphical way due to their size. The RF consists of multiple decision trees, which makes it difficult to visualize the entire model. The NN structure is rather standardized and globally known. Nevertheless, typical structures of these models are presented earlier in Figure 5 and Figure 6, respectively. The DT model includes 7 levels and 24 end nodes, with part of it being displayed in Figure 12. The RF model consists of 5 individual decision trees. The ANN model is structured upon 11 input nodes, 11 intermediate nodes, and 168 variables in total.

Table 2. Comparison of results of alternative methods.

	Annual Energy Consumption (EUR/m2)
	Model-SRM	Model-DT	Model-RF	Model-ANN	BIM
1	19.00	18.52	18.39	18.12	18.15
2	15.02	14.49	14.30	14.31	14.33
3	17.45	17.26	17.84	17.14	17.01
4	16.57	16.12	15.92	16.04	15.78
5	17.84	17.39	17.28	17.52	17.75
6	16.77	16.12	15.88	15.92	16.03
7	20.50	22.78	22.19	23.30	22.93
8	17.75	17.37	17.32	17.48	17.27
9	13.26	13.72	14.02	13.37	13.44
10	19.20	18.52	18.71	18.76	18.79

presents the energy consumption estimates, which are derived from the four models for indicative datasets, and the corresponding reference values, which are derived from the BIM model. The results indicate that the model output values do not greatly deviate from each other or from the corresponding reference values. To obtain some quantitative measures, the deviations from the associated reference values are in the ranges of 0.51% to 10.6%, 0.56% to 2.15%, 0.21% to 4.88%, and 0.14% to 1.65% for the SRM, DT, RF, and NN models, respectively. The results highlight the ML model as the best-performing one.

6.3. Model Performance Assesment

Statistical indicators (RMSE, PMRE, and correlation coefficient) are used to evaluate the dependencies between machine learning model output and actual observations. Further, the F-test is a statistical criterion to determine whether model results fit the observed ones. The corresponding mathematical relationships are given in Equations (3)–(5).

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}\cdot {\displaystyle \sum _1^n{\left(x_i-y_i\right)}^2}}$$

(3)

$$\mathrm{PMRE}={\displaystyle \frac{100}{n}}\cdot {\displaystyle \sum _1^n\left|\frac{x_i-y_i}{x_i}\right|}$$

(4)

$${\mathrm{R}}^2={\left({\displaystyle \frac{{\displaystyle \sum _1^n\left(x_i-x_{mean}\right)\cdot \left(y_i-y_{mean}\right)}}{\sqrt{{\displaystyle \sum _1^n{\left(x_i-x_{mean}\right)}^2}}\cdot \sqrt{{\displaystyle \sum _1^n{\left(y_i-y_{mean}\right)}^2}}}}\right)}^2$$

(5)

where x_i is the observed value, y_i is the predicted value, x_mean and y_mean are the corresponding mean values, and n is the sample size.

For in-depth performance appraisal of the developed models (SRM, DT, RF, and ANN), appropriate statistical analysis indexes are computed (

Table 3. Model performance indicators.

Model	Root Mean Square Error (RMSE)	Percent Mean Relative Error (PMRE)	Correlation Coefficient (R2)	F-Test
	All dataset		F 2, 335, 0.05 = 3.03
SRM	0.8329	3.74%	0.8810	171.55
DT	0.4930	2.17%	0.9460	7.06
RF	0.2722	1.15%	0.9836	5.46
ANN	0.1840	0.75%	0.9925	2.71
	Training		F 2, 233, 0.05 = 3.04
SRM	0.8525	3.76%	0.8797	119.56
DT	0.4822	2.02%	0.9498	6.70
RF	0.2302	0.96%	0.9888	5.51
ANN	0.1278	0.54%	0.9965	1.60
	Validation		F 2, 49, 0.05 = 3.19
SRM	0.8190	3.78%	0.8190	23.43
DT	0.5455	2.63%	0.9341	1.15
RF	0.3744	1.74%	0.9688	0.20
ANN	0.2759	1.19%	0.9820	1.01
	Testing		F 2, 49, 0.05 = 3.19
SRM	0.7514	3.63%	0.7514	28.22
DT	0.4871	2.40%	0.9400	2.34
RF	0.3241	1.43%	0.9727	1.54
ANN	0.2692	1.23%	0.9820	0.14

). They include the root mean square error (RMSE), the percent mean relative error (PMRE), and the correlation coefficient (R²). The F-test is further used for contrasting the goodness of fit between the observed and the estimated values. The results indicate that the decision tree (DT), random forest (RF), and artificial neural network (ANN) models exhibit improved performance compared to the statistical regression model (SRM), as indicated by the error indices (RMSE and PMRE) and the correlation coefficient (R2). Among pure ML methods, the ANN model presents the best performance and satisfies the F-test in every dataset. The other two methods (DT and RF) partially satisfy the F-test only in the validation and testing datasets.

More specifically, the RMSE of the ANN model output ranges from 0.1278 to 0.2759 for individual data subsets and is 0.1840 for the full dataset. On the other hand, the SRM model presents much higher RMSE values in the order of 0.7514 and 0.8525 in the data subsets and 0.8329 in the full dataset. Similarly, the PMRE values for the ANN model are between 0.54% and 1.23%, while for the SRM, they are from 3.63% to 3.78%. Finally, the F-test is fully satisfied in all tests of the ANN model, partly satisfied in the cases of the DT and RF models (in validation and training subsets), and not satisfied in any dataset of the SRM model. Interestingly, the performance in the validation and testing subsets occasionally appears to be higher than that resulting from the training subset. This effect may be attributed to the randomness of data segmentation in the three subsets. The observed peculiarity is rather insignificant as the model characteristics are not considerably affected. Additionally, this is an indication that the model does not overfit. The superior performance of ML models is due to their inherent ability to model nonlinear relationships and complex interactions among input parameters. Further, ML models hold the ability to adapt to data of different types and sources, which makes them suitable for use in problems with high complexity.

The model performance comparison can be further visualized in graphical form as in Figure 13, Figure 14, Figure 15 and Figure 16. These figures illustrate the comparison of the model output with the reference (target) values that have been produced by BIM simulation. The main diagrams show the range and dispersion of the dataset output and target points in relation to the diagram diagonal line, which indicates full convergence. The histograms above and to the right show the distribution of target and model values within their respective ranges. In an ideal convergence case, the two histograms would retain the same distribution. The multi-faceted presentation facilitates a better understanding and interpretation of the results, highlighting the accuracy and effectiveness of the proposed models. These diagrams reveal that, in general, there is an acceptable level of convergence between the observed and predicted values with all models. In a more detailed assessment, the ANN (mainly) and the RF present a stronger relationship between the two set values. Finally, the error distributions of the four prediction methods are shown in Figure 17. The ANN model presents the best performance with zero average values and low standard deviation. Instead, other method distributions shift to non-zero average values and present higher standard deviation values.

The research development and the case study results indicate that the proposed methodology can provide energy cost estimates easily and effectively at the very early design stage of typical residential buildings. Additionally, development and testing for specific building types and characteristics have been performed. The experiment shows that this research, in its current form, can have notable potential for scalability and can be extended to different building structures, places, and environments. More importantly, the adoption of BIM provides unlimited capability to develop energy models for any type and size of buildings, design characteristics, environmental factors, etc. Another important aspect is that building energy performance is highly affected by human-related factors, such as occupant number and behavior in relation to energy usage patterns, or operational and maintenance practices, which exhibit significant variability and uncertainty in energy consumption. Although such parameters may be marginally known at the early design phase, they may be incorporated in the modeling in a parametric or stochastic way. The above issues can be investigated as part of future research.

As a final note, the results of this work are qualitatively consistent with those in the literature. However, a comparative quantitative analysis that would enable meaningful comparisons with and evaluations using previous models and research works is not feasible, as the individual studies’ data structures are not fully comparable. Nevertheless, the evaluation of ML model performance is universally performed through the training, testing, and validation processes, which also depend on the size and representativeness of the data used.

In summary, the proposed methodology integrates a BIM simulation, parametric evaluation, and automated estimation of building energy consumption with a simple and effective GUI. The outcome of this effort is the development of ML models that can assess the energy consumption of buildings with distinct characteristics with various levels of performance. These models can effectively assist building designers and researchers in making decisions. The proposed methodology has the advantage of developing energy performance models without the need for the existence or collection of actual data. Instead, it relies on simulated data that can be extracted by consistent BIM-based simulations. Although the current development is confined to a (common) class of residential buildings, the above capability can highly support the repeatability, scalability, and extensibility aims of model development within a practical range of application. In the absence of such models, each individual building design should undergo an energy analysis of several potential solutions based on a multi-parameter evaluation. Another contribution is the development of a simple-to-use GUI (which is described in the next section) for obtaining energy consumption estimates at the early design stage, without the need for comprehensive building modeling and wide energy analyses of alternative building design configurations.

7. Graphical User Interface

The study has developed a method for estimating the long-term energy performance of buildings. BIM technology has been elaborated for the simulation of energy consumption analysis and machine learning techniques to predict building energy efficiency. A drawback of machine learning techniques is their unsuitability for real-life problems. To overcome this obstacle, a user-interface tool has been developed to facilitate the model application (Figure 18). In this application, the user simply inserts the appropriate design parameter values and presses the “Calculation” button. An instant energy consumption calculation is provided based on the selected parameters without any other manual input or calculation.

The integration of the trained artificial neural network (ANN) into the graphical user interface (GUI) has been performed using the synaptic weights of the developed model. The ANN model includes 168 parameters, representing the network weight and bias factors. The model has been implemented in Visual Basic programming language, which is suitable for interface development.

8. Conclusions

The energy performance of buildings has become a main concern globally in view of increased energy demand, the environmental impacts of energy production, and the reality of energy poverty in many areas. To increase building energy efficiency, proper building design should be implemented starting at the early design phase or when retrofitting decisions regarding existing buildings need to be made. Digital tools are currently available for performing energy assessment analyses in the form of building information modeling (BIM) and building energy modeling (BEM), and they can efficiently handle complex and technically demanding buildings. However, alternative designs should be subject to individual analyses, and this makes the process time-consuming and prone to errors. Machine learning techniques, on the other hand, can ameliorate such deficiencies and provide effective and easy-to-use tools for such analyses.

In this paper, representative residential buildings are considered along with eleven factors that affect energy performance. A dataset of 337 instances of such parameters is developed. For each dataset, the building energy performance is estimated based on a BIM simulation. Next, statistical and machine learning techniques are implemented to provide artificial models of energy performance. They include statistical regression modeling (SRM), decision trees (DTs), random forests (RFs), and artificial neural networks (ANNs). The above models are appropriately trained, tested, and validated using the dataset.

The case study results reveal that building energy consumption is mainly affected by the local ambient conditions, the number of building floors (one or two), the existence of adjacent buildings, the presence of a roof, and the window-to-wall ratio. Regarding the prediction models, the evaluation results from the case study indicate that all artificial models present satisfactory (for practical use at the early design phase) levels of convergence to the observed values. The comparative evaluation highlights the ANN as the best-performing model, with high statistical performance indicator values. Calculating the modeling energy efficiency of buildings via machine learning techniques appears to be an attractive approach to overcome the computational cost of creating energy efficiency simulations. Further, to enhance the model’s applicability, a user interface tool has been developed into which the design parameters of the eleven influencing factors are inserted and the energy cost is instantly calculated.

The experiment in the case study indicates that the proposed models can be directly extensible and scalable, at least within a reasonable parameter extrapolation range. For out-of-range cases, the process of developing the required dataset and training machine learning models can be presumably extended to any type and size of building, along with the rest of the influencing parameters. In a similar path, the development can incorporate human-related factors, which highly affect energy consumption.

The proposed methodology integrates BIM simulations, parametric evaluations, and automated estimations of building energy consumption with a simple and effective GUI. In particular, it provides a useful and handy tool for energy consumption estimations of buildings with varying characteristics without the need for acquiring and using actual data, as it develops these data via a BIM simulation. Another contribution is the development of a simple and easy-to-use GUI for calculating the energy consumption and comparatively assessing the energy performance of alternative building design configurations.