2.1. Description of the building
The Janata Bank building, located on the campus of Khulna University of Engineering and Technology in Khulna, was selected as the case study building. The structure is a well-equipped, two-story office building used for banking activities, situated in Khulna, Bangladesh. Autodesk Revit was used to model the as-built data and the floor plan is displayed in Fig. 1. Some photos of exterior and interior view of the existing office building are shown in Fig. 2. The office building was south faced. The selected building was visited and investigated several times to collect information about occupant number, occupancy schedule, lighting schedule, office equipment schedules other thermal comforting factors for rendering actual analytical model into Autodesk Revit. Table 1 provides important information about occupancy, schedules, materials, floor area, HVAC systems, and other critical factors required for energy simulations in each zone. The actual location in Khulna was selected to generate the weather report in Green Building Studio (GBS) and afterwards the energy simulation was conducted.
Table 1
Analytical Properties of Building
Building Type | Office |
Floor Area | 249.7138 m2 |
Occupancy Schedule | 12/5 facility |
Occupant Number | 17 |
Shading device | No shading devices |
Window Type | U value of Large single-glazed: 0.9795 W/m2K |
HVAC System | Residential 17 SEER split unit |
Average Lighting Power Density | 10.76 W / m² |
Average Equipment Power Density | 13.99 W / m² |
Specific Cooling | 7 m² / kW |
Specific Heating | 5 m² / kW |
Total Cooling Capacity | 59 kW |
Total Heating Capacity | 86 kW |
Wall | U value: 2.39 W/m2K |
Roof | U value: 3.99 W/m2K |
Window to wall ratio | 20% |
Using Building Information Modeling (BIM) tools, building in Revit entails producing a 3D model of a structure. Utilizing Revit's Building Information Modeling (BIM) tools, the project was designed and walls, floors, roofs, doors, windows, and other elements were added before starting to create a model. After the model was finished, Revit’s rendering environment was used to set up realistic materials, lighting, and views. Parameters like quality and resolution was adjusted to get the desired effects. The annual energy modeling for the building was done with Green Building Studio. This involved adding hourly meteorological data particular to each site to the prototype model. The current study used energy simulation tools, specifically sensitivity analysis, together with analytical methods enabled by Auto Desk Revit and Green Building Studio (GBS). The overall flowchart of the methodology is detailed in Fig. 3.
2.1.2. Validation of energy simulation
Validation of simulated results is an essential step in building energy analysis, ensuring that simulation models accurately represent real-world building behavior and provide reliable insights for decision-making in building design, operation, and retrofitting. The actual location in Khulna was selected to generate the weather report in Green Building Studio (GBS) and afterwards the energy simulation was conducted. The simulation yielded annual energy consumption of the virtual model was 53980 kWh whereas the actual real-life energy consumption was 57514.3 kWh which was obtained from Janata Bank manager. This means the virtual model’s consumption results differed from actual consumption data by 6.15%. The range of acceptable deviations between the simulated and real consumption data is roughly ± 15%. According to [30], our simulated results are reasonable. Furthermore, validation facilitates the discovery of inconsistencies or constraints in the simulation model, enabling its optimization and enhancement. It helps those who use simulation to comprehend the model's advantages and disadvantages and make the required modifications to improve accuracy and dependability.
2.2. Weather conditions
This research was conducted in the contexts of Bangladeshi climate which is located in tropical region in Southeast Asia. Heating system is not required for most of the regions in tropical climate. Moreover, geographic position of Bangladesh is in subtropical monsoon climate which is distinguished by wide seasonal variations in rainfall, high temperature and humidity. Bangladeshi weather stays hot and humid most of the time in a year. April and May are the hottest months, with highs frequently reaching 30°C (86°F) while November through February is the colder month, with slightly milder temperatures between 10°C and 20°C (50°F and 68°F) and also from June to October, there is a monsoon season that brings with its heavy rains and sporadic cyclones [31]. The southwest monsoon winds that originate in the Bay of Bengal also cause extensive flooding. On the other hand, little rain falls during the dry season, which runs from November to May, but humidity levels are still high, especially near the coast. Maximum office buildings in Bangladesh have installed air conditioning system only for cooling. The weather data for the key cities, which included the current location of Khulna (existing), Chattogram, Dhaka, Rajshahi, Sylhet, Barisal, and Rangpur, was generated using Green Building Studio. Autodesk Green Building Studio stations are based on recent actual year weather data rather than TMY2 and CZ2 stations, which are based on 30-year averages of weather data. The meteorological data recorded for the 7 cities was the was the data basis for modeling analysis of the simulation. These data were used as input for ANN modelling. The mean annual weather information of the chosen cities is shown in Table 2. The locations and their respective locational coordinates are given specifying the latitudes and longitudes. Dew point temperature, dry bulb temperatures, relative humidity and atmospheric pressure data have been collected from GBS and used for the simulation process to calculate the annual electricity consumption and annual electricity demand. In energy simulation, it is essential to comprehend atmospheric pressure, relative humidity, dew point temperature, and dry bulb temperature in order to precisely evaluate environmental factors and how they affect building performance.
Table 2
Mean annual weather input parameters of selected cities
City | Location | Dry bulb (degree Celsius) | Dew point (degree Celsius) | Relative Humidity (%) | Pressure (mb) |
Khulna | Latitude = 23.0500, Longitude = 89.6333 | 25.409 | 17.815 | 65.433 | 1006.03 |
Chittagong | Latitude = 22.4000, Longitude = 91.8833 | 24.989 | 19.758 | 74.17 | 1006.482 |
Dhaka | Latitude = 23.7333, Longitude = 90.3833 | 25.542 | 17.993 | 65.407 | 1006.772 |
Sylhet | Latitude = 24.9000, Longitude = 91.8333 | 24.879 | 17.7192 | 66.026 | 1007.837 |
Barisal | Latitude = 22.7000, Longitude = 90.3167 | 25.506 | 18.519 | 67.472 | 1007.01 |
Rajshahi | Latitude = 24.3833, Longitude = 88.6500 | 25.256 | 16.53 | 61.399 | 1005.162 |
Rangpur | Latitude = 25.7333, Longitude = 89.2833 | 24.785 | 16.0956 | 60.931 | 1003.89 |
Because location affects many different aspects of a building's performance, it is an important consideration in energy analyses of buildings. Variations in temperature, humidity, wind direction, and solar radiation all have an immediate impact on how much energy is needed for lighting, heating, and cooling. Variations in solar exposure can influence energy efficiency by helping or hindering passive solar design strategies.
2.3. Design parameters
The building's conceptual mass was reconstructed in order to analyze energy utilization. In order to conduct an energy assessment, the required modifications were made to the facade features, including changing the Window to Wall Ratio (WWR), changing wall types and insulation, and adjusting window glass materials and roof materials by using Autodesk Revit. The analytical building model was simulated using GBS, adjusting the WWR, wall types, window types, and roof types to see how each affected energy use. The energy simulation was carried out by modifying the window to wall ratio (WWR) and the design characteristics of the walls, roofs, and windows of the modelled building using the different weather data. The various window, wall and roof materials utilized along with their corresponding U-values are considered in the study. The U-value evaluates the rate of heat gain or loss through all of the combined thicknesses of the parts that go into making up a wall, floor, or roof in a building. The U-value, sometimes referred to as the thermal transmittance or heat transfer coefficient, is a term used in building energy analysis that describes the rate of heat transfer per unit area per unit temperature difference between the indoor and outdoor environments through a building element (such as walls, roofs, windows, or doors). The higher the energy efficiency when it comes to building components, the lower the U-value [32]. In this paper we used U-values as a way to differentiate different types of walls, roofs and windows and how it affects the energy consumption in the building. Several considerations must be made in order to ensure the best possible thermal performance and energy efficiency when choosing wall materials for energy analysis in building design and construction. Selecting the right insulation materials is important because they reduce heat transfer and keep interior temperatures comfortable. The various U values (W/m2/K) that were employed were 2.39, 3.99, 0.88, 0.46, 0.32, 0.19, 0.15, 1.21 for walls. Different types of walls materials used were uninsulated wall, R13 metal wall, R13 wood wall, R13 + R10 metal, ICF walls, R38 wood, R2 CMU and SIP. The u values of roof materials are existing = 2.52, R10 = 0.45, R19 = 0.33, R38 = 0.13, R60 = 0.08, SIP = 0.15, R15 = 0.34. The degree of insulation, solar heat gain, daylight penetration, and ventilation that different window options offer varies, and these factors are important in determining how much energy a building uses overall as well as the quality of its indoor environment. Single pane windows, which have only one layer of glass, are less expensive but have poor thermal performance, increasing heat gain and loss as well as the energy needed for heating and cooling. On the other hand, double pane windows have two glass layers that are divided by an insulating gas. This results in improved thermal insulation and a reduction in heat transfer, which enhances energy efficiency. With three layers of glass, triple pane windows further improve thermal resistance and work especially well in areas with drastic temperature swings. The different window types and their U values (W/m2 K) are existing = 0.9795, single clear = 6.17, double clear = 2.74, Double LoE = 3.49 and Triple LoE = 1.55.). The WWR is the ratio of the area of windows on a building's façade to the area of walls overall. It has a big impact on things like solar heat gain, daylight penetration, heating and cooling loads, and visual comfort. The WWR has an impact on how much solar radiation enters the building through the windows, which has an immediate effect on the loads for heating and cooling. The used WWR in this paper includes 0%, 20%, 30%, 65% and 95%. Table 3.3 shows the design parameters used for creating the analytic model.
Table 3
Input design parameters of simulation
Input Name | Parameter Values |
Wall Types U value (W/m²K) | Existing = 2.39 |
Uninsulated = 3.99 |
R13 metal = 0.88 |
R13 wood = 0.46 |
R13 + R10 metal = 0.32 |
ICF = 0.19 |
R38 wood = 0.15 |
R2 CMU = 1.21 |
SIP = 0.15 |
Window Type U value (W/m²K) | Existing = 0.9795 |
Single clear = 6.17 |
Double Clear = 2.74 |
Double LoE = 3.49 |
Triple LoE = 1.55 |
Roof Types U value (W/m²K) | Existing = 3.99 |
Uninsulated = 2.52 |
R10 = 0.45 |
R19 = 0.33 |
R38 = 0.13 |
R60 = 0.08 |
SIP = 0.15 |
R15 = 0.34 |
Window to wall ratio (%) | 0 |
20 |
30 |
65 |
95 |
2.4. Simulation results sample database
Table 4 presents a sample of the database generated which includes input parameters and the Annual Electricity Usage (kWh) and Annual Electricity Demand (kWh), which is considered for this study for building the database of the ANN model. Large volumes of data are needed for machine learning models to identify patterns and generate precise predictions. This data can be stored and arranged in an organized manner using databases, which makes it available for model training.
Table 4
Energy consumption database sample
Dry bulb (degree Celsius) | Dew point (degree Celsius) | Relative Humidity (%) | Pressure (mb) | Window to wall ratio (%) | Window Type U value (W/m²K) | Wall Types U value (W/m²K) | | Annual Electricity Usage (kWh) | |
Roof Types U value (W/m²K) | Annual Electricity Demand (kWh) |
25.409 | 17.815 | 65.433 | 1006.03 | 20 | 0.9795 | 2.39 | 3.99 | 53980.22 | 66597.95 |
25.409 | 17.815 | 65.433 | 1006.03 | 95 | 3.49 | 2.39 | 3.99 | 55264.35 | 68725.39 |
25.409 | 17.815 | 65.433 | 1006.03 | 30 | 2.74 | 2.39 | 3.99 | 53987.18 | 66622.98 |
25.409 | 17.815 | 65.433 | 1006.03 | 20 | 0.9795 | 0.15 | 3.99 | 51960.22 | 63284.79 |
25.409 | 17.815 | 65.433 | 1006.03 | 20 | 0.9795 | 2.39 | 2.52 | 50945.5 | 61345.08 |
24.989 | 19.758 | 74.17 | 1006.482 | 95 | 1.55 | 2.39 | 3.99 | 52533.92 | 60359.58 |
25.542 | 17.993 | 65.407 | 1006.772 | 65 | 3.49 | 2.39 | 3.99 | 54955.11 | 67527.14 |
25.542 | 17.993 | 65.407 | 1006.772 | 20 | 0.9795 | 2.39 | 0.13 | 43464.29 | 46778.44 |
25.542 | 17.993 | 65.407 | 1006.772 | 20 | 0.9795 | 2.39 | 0.08 | 43211.87 | 46290.38 |
25.256 | 16.53 | 61.399 | 1005.162 | 0 | 0.9795 | 2.39 | 3.99 | 53095.11 | 66403.98 |
25.256 | 16.53 | 61.399 | 1005.162 | 20 | 0.9795 | 2.39 | 2.52 | 50461.63 | 62049.01 |
25.256 | 16.53 | 61.399 | 1005.162 | 20 | 0.9795 | 2.39 | 0.45 | 43614.17 | 48076.8 |
24.879 | 17.7192 | 66.026 | 1007.837 | 20 | 0.9795 | 2.39 | 3.99 | 52002.43 | 61804.98 |
24.879 | 17.7192 | 66.026 | 1007.837 | 65 | 6.56 | 2.39 | 3.99 | 52787.38 | 63231.61 |
24.879 | 17.7192 | 66.026 | 1007.837 | 65 | 6.17 | 2.39 | 3.99 | 52609.16 | 62993.84 |
24.879 | 17.7192 | 66.026 | 1007.837 | 30 | 3.49 | 2.39 | 3.99 | 52036.68 | 61851.91 |
24.879 | 17.7192 | 66.026 | 1007.837 | 30 | 1.55 | 2.39 | 3.99 | 51753.96 | 61398.26 |
24.785 | 16.0956 | 60.931 | 1003.89 | 95 | 3.49 | 2.39 | 3.99 | 53909.95 | 68162.24 |
24.785 | 16.0956 | 60.931 | 1003.89 | 95 | 1.55 | 2.39 | 3.99 | 52984.15 | 66663.65 |
24.785 | 16.0956 | 60.931 | 1003.89 | 20 | 0.9795 | 2.39 | 0.13 | 42235.1 | 46966.15 |
24.785 | 16.0956 | 60.931 | 1003.89 | 20 | 0.9795 | 2.39 | 0.08 | 41991.39 | 46503.12 |
2.5. Artificial Neural Network (ANN) Model
The proposal of an ANN model for the purpose of forecasting Bangladeshi office building energy consumption is one of the study's objectives. The ANN model is presented as a solution to the nonlinear fitting problem, taking into account intricate relationships between inputs and outputs. The ANN model, a black box model with a strong capacity for self-learning, is widely used for high-accuracy prediction and quick computation. The modeling process is significantly streamlined by using an ANN model as opposed to a white box model, which necessitates intricate parameter setting and laborious model construction. In terms of model development, the ANN model is trained and tested using the MATLAB toolbox. The following lists the stages involved in creating an ANN model. In order to train and validate the simulated building and represent the investigated problem in any general form or condition, an appropriate database must be implemented in order to develop the ANN model. Seventy percent of the 434 simulated datasets were used for training, fifteen percent were used for selection, and fifteen percent were used for testing. The neural network takes eight inputs: pressure, WWR, relative humidity, dry bulb temperature, dew point temperature, window type, wall type, and roof type. Annual energy consumption and annual electricity demand were the results. The Levenberg-Marquardt (LM) algorithm was used to implement the neural model, which was designed to approach second-order training speed without requiring the computation of the Hessian matrix. Neurons in an ANN model is assembled to simulate the structure and functions of the human brain. The transfer of signals depends on the function of neurons, which are the fundamental components of the ANN model. With the Levenberg Marquardt learning algorithm, the hidden layer and output layer neurons use the hyperbolic tangent sigmoid and pure linear transfer function, respectively. The continuously differentiable hyperbolic tangent sigmoid function takes inputs from [- ∞, +∞]and converts them to [− 1, 1]. A pure linear function has equivalent inputs and outputs. The ANN model can produce any value thanks to its pure linear function. The goal of the ANN model's training process is to minimize the error between the target and output by continuously adjusting the weight and bias. The performance function in use is MSE. Figure 4 depicts the ANN model's convergence process. The training error has an initial value of 0.7682 and a final value of 0.0071 after 74 epochs. The selection error has a starting value of 1.0896 and ends up at 0.0082 after 74 epochs. Having low selection error—that is, effectively generalizing to new data—rather than merely learning the training set by heart is the ultimate aim of a machine learning model.
2.6. Structure of ANN model
The input layer, hidden layer, and output layer make up the ANN model's structure. An initial backpropagation ANN model is built with an input layer that represents the eight parameters previously described, a hidden layer that contains three neurons, and an output layer that represents the annual electricity demand and consumption (kWh). The number of inputs and outputs, respectively, determines the number of neurons at the input and output layers. To get the best model accuracy, the number of neurons at the hidden layer must be frequently changed. The input signal travels through the hidden layer, input layer, and output layer. The ANN model's capacity for information processing and adaptation is significantly increased by the parallel computation of multiple neurons, enabling the solution of nonlinear problems. The trial-and-error approach is advised in regards to hidden neurons [33]. Training results showed that an ANN model with four hidden neurons has the highest accuracy. The ANN model's structure is displayed in Fig. 5. If, after a certain number of iterations, the Mean Square Error (MSE) stabilizes, the training is said to have achieved convergence. algorithms for selecting neurons that are used to determine the ideal quantity of hidden neurons in a network. The growing neurons algorithm is the one we've chosen to determine the ideal neuron count for this application. This method adds a certain number of neurons in each iteration after starting with a minimum number. There was a minimum of one hidden perception to be evaluated and a maximum of ten. One hundred was chosen as the maximum number of iterations at which the selection error increases. There was a maximum of 1000 iterations required to complete the algorithm. Every neural network had three trials. Choosing a model whose complexity is optimal for generating a satisfactory data fit yields the best results. Four neurons are required to achieve the lowest selection error, as the Fig. 4 illustrates. The ideal selection error at 4 neurons was 0.00551, and the optimal training error was 0.00545. Our neurons selection algorithm has determined the ideal number of neurons for this new neural network architecture. The new neural network has 8 inputs, 1 output, and 4 neurons in hidden layer or perception layer. We selected this new network to use with our ANN model.
The Fig. 6 depicts the graphical representation of the chosen network architecture. It contains the following layers: Scaling layer with 8 neurons (yellow), perceptron layer with 4 neurons (blue), Perceptron layer with 2 neurons (blue), unscaling layer with 2 neurons (orange) and bounding layer with 2 neurons (purple).
2.7. Prediction performance evaluation indexes
Three statistical indices are introduced to evaluate the prediction accuracy of ANN model comprehensively, which are R2, RMSE and CV-RMSE. They are calculated by the following equations:
$$\:{R}^{2}=1-\left(\frac{{\sum\:}_{i=1}^{n}{({x}_{i}-{y}_{i})}^{2}}{{\sum\:}_{i=1}^{n}{({x}_{i}-\stackrel{-}{x})}^{2}}\right)$$
1
$$\:RMSE=\sqrt{\frac{1}{n}\sum\:_{i=1}^{n}{({x}_{i}-{y}_{i})}^{2}}$$
2
$$\:CV-RMSE=\frac{RMSE}{\stackrel{-}{y}}$$
3
Where xi and yi represent the actual and predicted values of the dependent variable, respectively. Whereas n represents the number of samples.\(\:\:\stackrel{-}{y}\) represents mean of the observed (simulated) values. RMSE (Root Mean Square Error) and CV-RMSE (Coefficient of Variation of Root Mean Square Error) are both metrics used to evaluate the performance of a regression model, particularly in terms of how well it predicts values relative to the observed or true values. RMSE directly measures the average magnitude of errors in the predicted values, CVRMSE provides a relative measure of error that's scaled by the mean of the observed values, making it easier to interpret and compare across different datasets or scenarios. The coefficient of determination, often denoted as R2 is a statistical measure that represents the proportion of the variance in the dependent variable that is predictable from the independent variables in a regression model. R2 helps to assess how well the model fits the data.