Evaluation of the Impact of Input-Data Resolution on Building-Energy Simulation Accuracy and Computational Load—A Case Study of a Low-Rise Office Building

Abstract

Building-energy consumption is the primary aim of urban energy consumption, which can aid in optimization of building operation and management techniques, creating sustainable building and built environments. However, modellers’ understanding of the relationship between building-energy modelling (BEM) accuracy and computational load is still qualitative and deprived of accurate quantitative study. Based on a bottom-up engineering methodology, this study aims to quantitatively explore the effects of building-model input data with different resolution accuracies on energy simulation results, including evaluation of computational load. According to the actual parameters of the case-study building, 108 models with varying input resolution levels were developed to estimate hourly energy usage and annual mean ambient temperature. The results demonstrated that with input parameters at low resolution levels, geometric parameters such as exterior windows, interior windows, and shading exhibited significantly lower computational loads, resulting in reduced errors in the final simulation performance, whereas the occupancy schedule, thermal zoning, and HVAC configuration parameters exhibited significant declines in simulation performance and accuracy. This study presents a methodology applicable to the majority of low-rise, rectangular office structures. Future work would concentrate on carrying out comparison tests for different building forms and types while gradually improving the automation of the process to enable use of the appropriate accuracy level in assessing the crucial issue of energy-modelling input.

1. Introduction

1.1. Motivation

As a result of rapid population growth, economic development, and urbanization over the past few decades, global energy consumption has rapidly increased [1], resulting in a substantial increase in carbon-dioxide emissions and worsening of global warming. In 2021, global CO₂ emissions reached an all-time high of 36.3 billion tonnes [2], up 6% from the previous year. More than 40% of the total primary energy consumption in the United States and the European Union is attributable to buildings [3], creating a great potential and obligation to achieve sustainable goals.

Researchers from around the globe are developing building-energy models and energy-efficient ways to reduce the energy consumption of buildings [4]. Retrofitting and optimizing geometry and passive as well as active strategies for new and existing buildings [5] are the chief methods to reduce building-energy costs. In this case, building-energy modelling (BEM) is required to test these energy-saving methods. Recently, it has emerged as an interesting paradigm for evaluating building-energy efficiency, as it plays a crucial role in optimizing every phase of the building system’s life cycle, from predesign to operation. For single-building-scale BEM software, such as DOE-2 [6], EnergyPlus [7], and DeST [8], the underlying premise is a physics-based engineering model, with detailed energy consumption, modelled from 2D building drawings and occupancy schedules. To achieve energy efficiency in urban buildings, BEM will eventually need to be extended to urban building-energy modelling (UBEM). Energy simulation of buildings at the urban scale is challenging due to the complexity of creating and running simulation models for an enormous number of buildings. CitySim [9], UMI [10], CityBES [11], CEA [12], and TEASER [13] are presently the most prominent BEM software for urban buildings. They are based on a simulation method that can be described as a thermal-load simulation of extremely simplified “boxes”. While this reduces calculations and improves modelling speed on a large scale, its accuracy and reliability are low [14].

At present, researchers’ attention to this critical issue almost exclusively stems from the process of using statistical methods for BEM to perform uncertainty and sensitivity analyses. This is often described as “computational cost” and is included as one of the main considerations in selecting appropriate statistical submethods for BEM [15,16,17]. However, the core concentration of this study is the accuracy and computability issues of the BEM process itself due to insufficient past studies systematically focused on it. As stated by Hao et al. [18] in their study, building-energy models must strike a balance between accuracy and computational load to be applicable. Existing mainstream methods have a research gap in this aspect, which is inadequate for achieving balance.

1.2. Input Parameters for Building-Energy Modelling (BEM)

The fenestration of a building has a considerable impact on its performance, and the window-to-wall ratio (WWR) is a crucial statistic that describes the proportion of windows on an external wall [19]. In general, different climate zones and orientations must be considered, in designing the ideal WWR for each façade of a structure, as optimal design parameters for cooling and heating loads [20]. Current building-energy models for commercial deliveries and academic research, such as DesignBuilder, almost always assume a common WWR value to simulate building openings and produce final energy-consumption statistics. Thermal-bridging-based losses between windows and walls are a substantial source of heat loss, which is concerning. Gustavsen et al. [21] used a typical 160 m² Norwegian home as an example in which the window-to-wall interface accounted for about 40% of the total heat loss, owing to thermal bridging. Misiopecki et al. [22] examined the thermal performance of window-to-wall connections through testing multiple wall constructions and windows with varying insulation properties. It was demonstrated that the position of a window has a significant impact on the thermal bridging effect, with windows placed in the most energy-efficient positions reducing thermal bridging losses by more than 50%. Thus, even when the overall window-to-wall ratios (WWRs) and all other criteria (U-value, infiltration rate, on/off profile) are identical, variances in positions of window openings will likely have a significant impact on a building’s total energy consumption.

Solar shading of buildings is a major component of their overall energy usage, which can have a direct impact on a building’s solar heat gain. The most effective way to filter solar radiation before it reaches the glazed area, according to the ASHRAE [23], is to employ an exterior shade device. In certain circumstances, shade elements could block up to 80% of solar heat gain to a glazed surface. In addition, the performance of shading is also affected by the position and climate of the area. Kirimtat et al. [24] further stated that simulation tools are commonly used to determine optimal shading elements through identifying shading performance. Previous studies have focused mainly on solar-shading performance, and the effect of model accuracy on solar-shading performance has not been discussed.

An occupancy schedule is used to explain occupant behaviour (OB), which is typically seen as a quantitative depiction of direct and indirect influences of occupants on building performance. Specifically, occupant behaviour and activities are one of the six determining factors of building-energy consumption [25]. A substantial effect on the accuracy of building-performance simulation results from it [26]. Heydarian et al. [27] discussed the significance of OB, including how a building’s occupants contribute as a source of heat. Furthermore, interaction between occupants and building systems affects the energy consumption of the HVAC, lighting, and equipment loads. Therefore, it is essential that the occupancy schedule is accurate. Typically, occupancy-scheduling modelling can be classified as stochastic or deterministic. The occupancy schedule adheres to a deterministic rule in a deterministic model. According to the stochastic model, the occupancy schedule is a random output from an occupancy simulator. In general, replacing a deterministic model with more appropriate stochastic processes would improve accuracy [28].

Thermal zoning of a building is an important definition in building-energy modelling, including thermal zoning in design of HVAC systems as well as thermal zoning in building-performance simulation to control the thermal environment of the building through individual set points and schedules for individual thermostat sensors. It is generally accepted that the earlier multizone thermal model simulations are used in the design process, the more impact they will have on crucial design decisions [29]. Shin et al. [30] indicated in their assessment of thermal zoning for energy simulation that while thermal zoning techniques based on standard practices exist, there is no standard approach to thermal zoning for all building types that can be studied using a whole building-energy simulation program.

A recent study found that depending on HVAC system choice and geographic location, HVAC accounts for between 23.8% and 72.9% of total building-energy usage [31]. Consequently, modelling of HVAC is crucial for promoting its continued effective regulation and enhancing energy efficiency. Stadler et al. [32] explained that simulation tools, such as EnergyPlus, could estimate the energy consumption of a building and examine the load performance to evaluate operating-system strategies based on standard procedures and algorithms. In addition, the algorithms would require input, including component capacity and flow rate. DesignBuilder categorises input data into three kinds, namely default, autosize, and custom, according to its official website [33]. In particular, EnergyPlus calculates autosize input based on existing conditions such as weather data, zone information, and set-point value. However, Ahn et al. [34] noted that excessive use of the default and autosize parameters in EnergyPlus could significantly increase the uncertainty of results, compromising objectivity and reliability. There is currently a lack of comprehensive quantitative analysis of the effects of automatic and custom dimensions on the energy consumption of buildings.

As for the general BEM process, previous research has almost exclusively focused on the effects of these model input parameters on their own (e.g., occupant behaviour [35,36,37], HVAC system [38], thermal zoning [39,40,41]) on building-energy-simulation results, with few studies focusing on the possible interactions between these model input parameters and their impacts on final simulation results. Other, similar studies focused on the calibration process of BEM, while some effort was devoted to studying resolution of the metered parameters required for BEM calibration [42,43]. Conclusively, there is no previous article that has quantitatively assessed which BEM input parameter and which level of resolution have significant effects on model fidelity. Alternatively, a few studies recommended sensitivity analysis to improve building-energy assessments. For instance, Kristensen et al. [44] identified and ranked the influence of model input parameters using three sensitivity-analysis techniques; however, they neglected to account for computational load in their research. Nonetheless, the focus of current research is still mostly on the applicability of various sensitivity-analysis methodologies in BEM, with particular examples [45,46,47,48]. In addition, there were also very few prospective studies on accuracy of simulation results or simulation duration.

1.3. Aim and Contribution of this Research

Prior to officially conducting this research, the following key concepts, which can help readers better understand the innovation of this paper, were precisely defined:

• Input-Parameter Resolution: This refers to the accuracy of a certain input parameter of the building-energy model, along with the level of detail included. In this article, input-parameter resolution is manually graded.
• Computational Load: This refers to the consumption cost of the computer hardware required to complete BEM simulation based on a certain input-parameter resolution, which can be measured in specific ways.
• Simulation Performance: This refers to the efficiency of the BEM simulation. Higher accuracy with a lower calculation load results in better simulation performance.
• Model-Error Metrics: This refers to an indicator that allows quantitative evaluation of the error size of the model results, which is derived from official regulations.
• Sensitivity Analysis: In this study, this refers to the use of various statistical methods to find the degree of influence of different input-parameter resolutions on simulation performance.

In general, with respect to the literature discussed above, the aim of this research was to examine and rate the resolutions of various major geometric and non-geometric building-model input parameters that may affect overall BEM performance, as well as the computational load they impose. The core contributions of this research to the current knowledge are the following:

• Simultaneous consideration of the several most common BEM input parameters and grading their resolution levels scientifically;
• Creative permutation of different input-parameter resolutions into 108 models to test the effects of all input parameters on their own and based on their interactions;
• Measuring the computational load through the CPU monitoring method, which is more accurate, instead of the traditional timing method.

In addition, this research combined multiple statistical methods (the Morris method for sensitivity analysis, the normality test, U-testing, etc.) to analyse the relationships between input-parameter resolution, simulation accuracy and computational load. Moreover, the final results enhanced the modellers’ understanding of the connections between model input parameters and their simulation performance, driving BEM and UBEM toward better accuracy and computing efficiency.

This article comprises five sections. Section 2 explains the methods employed in the BEM input-parameter resolution-level division, computational-load metering and sensitivity analysis. Section 3 presents the results, followed by the relevant discussion. Recommendations for future studies and findings are presented in Section 4 and Section 5.

2. Methodology

2.1. Case-Study Building

Prototypical building-energy models represent limited types of buildings within a certain geographic area [49]. In taking Shanghai as an example, office buildings account for more than 25% of the existing commercial building stock [50]. In addition, to the authors’ knowledge, about half of the office buildings in major cities in China are low-rise office buildings. In 2020, Hong et al. [51] systematically surveyed and collected data from 136 individual low-rise office buildings in 10 office parks in Minhang District, Shanghai, China. Through a performance-index system (PIS) they developed, they clustered and analysed the six key indicators and refined them into four prototypical buildings.

Low-rise office buildings were chosen as the case-study buildings because they themselves not only account for considerable energy consumption but also have great significance to formulation of energy-saving measures for other commercial buildings of the same type. As shown in Figure 1 below, the campus of the University of Nottingham, Ningbo, China, is situated in the eastern coastal city of Ningbo, where the climate is typically subtropical monsoon, with hot summers and cold winters. The Science and Engineering Building (SEB) of the campus (lat. 29°48′ N, long. 121°33′ E), with a total building area of 13,975.30 square meters and four stories aboveground, was completed in 2010, meaning it can be classified into the C2W3F4 prototypical buildings suggested in Hong et al.’s study. As it was a typical low-rise office building, we were able to generalise similar building stock from the SEB; thus, it was finally selected as the case-study building for this research.

The SEB utilises a reinforced concrete frame construction, an aerated concrete self-insulating outer wall filled with thermal-bridge polymer thermal-insulation mortar, extruded polystyrene boards for the roof, and heat-insulating metal-profile window frames for the outside windows.

Table 1. Basic building characteristics of the SEB.

Facade and Orientation	Front Elevation Southern Face
Floor Shape	Rectangular
Floor Dimensions	64.7 m × 70.1 m
Total Height	17.2 m
Window-to-Wall Ratio	0.35 (east), 0.4 (south), 0.33 (west), 0.34 (north)
Thermal-Performance U-Value (W/m2K)	0.67 (roof), 1.0 (wall), 3.4 (windows)
Total Energy-Transmittance Coefficient of Windows G-value	0.691

, which follows, outlines the fundamental properties of the SEB. Among them, the U-value and G-value of the building were set as constants according to the real situation of the SEB in this study. This is due to the fact that, unlike the WWR, they could not be input at different resolutions, although they significantly affected the overall thermal performance of the building. There have been few studies on sensitivity analysis of building-performance simulations based on changing U-values and G-values, which are not within the scope of this study.

2.2. Building-Energy-Modelling Simulation Process

EnergyPlus inherited and combined the best features of the BLAST and DOE-2 programs with unmatched advantages in connectivity and extensibility, including third-party interface and module development [52]. DesignBuilder (DB) is simulation software for building performance, designed specifically to execute the EnergyPlus simulation virtual building model [29]. It provides a friendly graphical interface for easy creation of EnergyPlus models. The joint use of the two allows customization of simulation parameters and boundary conditions as well as execution of dynamic energy simulation [53]. Python is a powerful, elegant programming language [54] that bridged the gap between EnergyPlus and other available BEM optimization tools to enable building-energy optimization in a computationally efficient manner [55]. More importantly, the EnergyPlus program can be loaded via Python scripts in batches, enabling automatic operation. Considering that this research required plenty of repeated editing of the geometric and nongeometric configurations of the case-model buildings and a large number of simulation-file-running tasks, this research finally applied DB (v7.0.0) and EnergyPlus (v9.4.0) as the primary BEM tools. Figure 2 below clearly shows the modelling, simulation, and analysis processes of this research.

2.3. Model Input-Parameter Resolution

After students majoring in built-environment engineering at the University of Nottingham were spoken with, the findings revealed that the openings, shading, and thermal zone division in their building-energy simulation were nearly identical to the original building’s geometric features as depicted in the engineering CAD. Although this approach appeared to be in perfect accord with the actual building and seemed to ensure complete accuracy, no systematic research had been conducted to demonstrate this notion. Furthermore, this method of simultaneously measuring BEM performance and computational load tends to generate hundreds of different room partitions with a high level of detail, which poses significant obstacles for BEM tasks, including the effort of the modeller and the load on the computer.

Based on the research analysis in Section 1, this study creatively classified building openings, shading, scheduling, thermal zones, and HVAC configurations into two to three distinct detail levels. All input-data-resolution levels and diagrams are displayed in

Table 2. Details of model input-data-resolution levels.

Item		Resolution Description		Annotation No.	Image
Openings	External Windows (Two Levels)	/	Photo	/
		A	A WWR value entered for the entire building (0.355)	①
		B	Layout based on the actual situation	/
	Internal Windows (Three Levels)	/	Photo	/
		A	No internal windows	/
		B	Utilises DB’s internal glazing templates.	②
		C	Layout based on the actual situation	/
Solar Shading (Three Levels)		/	Photos	/
		A	No solar shading	/
		B	Utilises DB’s default overhang shadings	③
		C	Layout based on the actual situation	/
Occupancy Schedule (Two Levels)		A	Default typical activity template provided with DB	④
		B	Uses high-resolution occupancy profile	⑤
Thermal Zoning (Two Levels)		A	Divides the building into internal and external zones based on 6.86 m depth	⑥
		B	Based on the real layout; the rooms are merged with the same orientation and type	⑦
HVAC System (Two Levels)		A	Default autosized VRF AC model provided with DB	/
		B	Custom VRF AC: heating/cooling capacity and airflow rate for VRF AC system set based on actual layout and configuration	⑧

, while

Table 3. Annotation of the energy-model resolution division.

Annotation No.	Remarks
①	The WWR value of the four facades of the SEB was calculated to be 0.355 based on the existing conditions.
②	Utilised the 20% internal glazing template provided with the DB software.
③	Utilised the 1.0 m overhang shade provided with the DB software.
④	Utilised the standard office-building activity file provided with the DB software.
⑤	Utilised the high-resolution occupancy simulator developed by Chen et al. [56]
⑥	Standard thermal zoning strategies specified in Appendix G of ASHRAE Standard 90.1-2016 [57]. Specifically, Appendix G states that due to the strong interaction of building-perimeter spaces with the outdoors as a result of solar penetration, the floor height is defined as being within 15 feet (4.6 m) of windows for a typical office ceiling height of 9 to 10 feet (2.7 m to 3 m). The floor height of the building in this case (about 4.5 m) exceeded the ASHRAE standard. The computation indicated that the critical depth was 22.5 feet (6.86 m).
⑦	Based on hourly load calculations and building geometry, the HVAC system combined areas/rooms of the building into thermal zones. Specifically, two zones/rooms were considered to be merged into the same hot zone of the HVAC system if their hourly load calculations showed sufficient approximations in the time series [58].
⑧	Customised the parameters of the VRF system using HVAC layout drawings from the engineering construction office.

has special annotations.

All classes of building-energy modelling input parameters were numbered and organised as A, B, and C to indicate increasing detail, based on the principle of free permutation. As a result, there were 144 models in total. Since class A for thermal zoning and class B for the HVAC system were incompatible and cannot coexist, there were 108 models in total. The names of all of these models can be found in the Appendix A at the end of the article.

2.4. Model-Error Metrics

Several criteria are necessary for quantitative validation of building-energy model simulations, which is the first step in the workflow of the UBEM standard for model calibration. Common benchmarks or metric measuring systems currently exist for BEM in order to match building-energy simulation models to measured data.

In general, the normalised mean bias error (NMBE) and the coefficient of variation of the root mean square error (CVRMSE), which evaluate the variability of errors between simulated and measured values, were used to determine the confidence in particular building-energy models. The American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) Guide 14 [59] and the Federal Energy Management Program (FEMP) Measurement and Verification (M&V) Guide set goal values [60] for the NMBE and the CVRMSE at the individual building level, as did the International Performance Measurement and Verification Protocol (IPMVP) [61].

The NMBE (normalised mean bias error) was utilised to make the MBE (mean bias error) results comparable. In accordance with the principle of normalization, the global difference between the measured and predicted values was calculated via dividing the NMBE by the mean of the measured values ($\overline{m}$) (shown in the equation below). Particularly, Reddy et al. [62] noted that the recommended calibration value for the adjustable model parameter (p) is zero. However, Coakley et al. [63] noted that a cancelling effect may occur when the NMBE is used alone, which may increase uncertainty. Therefore, the NMBE cannot be employed independently.

The CVRMSE is used to calibrate building-performance models. This metric indicates whether the observed relationship between variables during a baseline period is unstable. It is the coefficient of variation between the expected and observed input series. Specifically, the suggested value for the model parameter (p) was 134. Unlike the NMBE, the CVRMSE is not affected by cancellation errors. Therefore, the ASHRAE, the FEMP, and the IPMVP use the NMBE and the CVRMSE to validate the accuracy of models [64].

R² (the coefficient of determination) is another evaluative metric (the equation shown below). This indicator, according to Moriasi et al. [65], represents the proportion of the measured data’s variation that a model predicts. In addition, the result of R² is confined to a range between 0 and 1, and the maximum value indicates that the simulated result precisely matches the measured value. In Reddy et al.’s [62] report, it was demonstrated that R² is not a typical indicator for calibrated models. However, both the ASHRAE and the IPMVP recommended that this indicator’s calibration value should not be less than 0.75 [66,67]. The formulas of the three evaluation metrics used in this study are as follows:

$$NMBE=\frac{1}{\overline{m}}·\frac{{\displaystyle {\sum }_{i=1}^n}(m_i-s_i)}{n-p}\times 100(\%)$$

$$CV(RMSE)=\frac{1}{\overline{m}}\sqrt{\frac{{\displaystyle {\sum }_{i=1}^n}{(m_i-s_i)}^2}{n-p}}\times 100(\%)$$

$$R^2={\left(\frac{n·{\displaystyle {\sum }_{i=1}^n}{m}_i·si-{\displaystyle {\sum }_{i=1}^n}{m}_i·{\displaystyle {\sum }_{i=1}^n}{s}_i}{\sqrt{(n·{\displaystyle {\sum }_{i=1}^n}{m_i}^2-{({\displaystyle {\sum }_{i=1}^n}{m}_i)}^2)·(n·{\displaystyle {\sum }_{i=1}^n}{s_i}^2-{({\displaystyle {\sum }_{i=1}^n}{s}_i)}^2})}\right)}^2$$

where

$m_i$ is the measured (baseline) data for each model instance, “i”;

$s_i$ is the simulated data for each model instance, “i”;

$p$ is the model parameter (the suggested value is 134);

n is the number of the model sample.

This research’s validating temporal step is hourly. Since the campus administration had not installed any modern metering equipment, it was unable to offer real-time data on energy consumption. Thus, the baseline model was the most accurate and theoretically complex model: No. 144. (BCCBBB). A total of 107 other models were compared to this one. Based on ASHRAE Guideline 14 [68], the FEMP [60], and the IPMVP [69], the

Table 4. Acceptable validation/calibration tolerances.

Calibration Step	Index	Acceptable Values for ASHRAE	Acceptable Values for FRMP	Acceptable Values for IPMVP
Hourly	NMBE	±10%	±10%	±5%
	CVRMSE	30%	30%	20%

below details the acceptable calibration tolerance for the NMBE and the CVRMSE.

In this study, as the simulated values included hourly data on energy usage throughout the year, there were no actual observations with the same temporal resolution. Model 144 (BCCBBB) was deemed the baseline model, while the remaining 107 models served as reference models.

2.5. Computational-Load-Metering Method and Metrics

Another core innovation of this study was a new metering method for computational load. Before objective measurement of a computational load, it is vital to comprehend how EnergyPlus will simulate using the CPU or GPU of the computer. The EnergyPlus simulation procedure was run on the CPU in a single-threaded manner by default. Procmon’s [70] monitoring of the process tree revealed that just one process identifier (PID) has to be tracked for the complete EnergyPlus application to function. Therefore, as shown in Figure 3, identifying the unique PID (energyplus.exe) was sufficient to determine the CPU calculations performed via the EnergyPlus simulation operation.

The script that monitored the CPU calculation of the EnergyPlus simulation was written in Python. The script’s code was broken into two sections: the feedback section and the monitoring section. The monitoring component first monitored and collected the PID of each EnergyPlus action, then performed memory monitoring, data processing, and accumulating the memory gained per second of data to obtain the values of occupied memory and feedback. That part looped in 0.1 s to ensure that EnergyPlus began monitoring within 0.1 s of starting and was then called the monitoring portion. When the EnergyPlus process was completed, the PID was removed and the monitoring was terminated. Moreover, the computation result was a relative value without a unit; this is only needed for proper comparison in following studies.

2.6. Sensitivity Analysis for Simulated Results

In the case of performance prediction with BEM, it was useful to carry out frequent sensitivity analysis (SA) to investigate model behaviour and identify the input factors responsible for the majority of model output variations. Consequently, sensitivity analysis is a broad term for assessing how variability in model outputs is attributed to variability and uncertainty in model input parameters. Currently, numerous researchers have conducted comparisons of different sensitivity-analysis methods, and it can be inferred that no one method could be considered superior, as each method holds its own distinct advantages [71,72]. For this study, the Morris method [73] was the most suitable method, since it allowed more comprehensive evaluation of a model’s behaviour and was sensitive to input parameters.

Relative to other sensitivity-analysis algorithms, the Morris method is fast and involves varying only one parameter at a time and observing the resulting changes in the output of the model (the discretised value). Through repeating this process for each parameter, it is possible to quantify the sensitivity of the model to each parameter and to identify the parameters that had the greatest impact on the model output [15]. In this study, sensitivity analysis of model simulation results, including various submethods and subalgorithms used, was carried out under the framework of the Morris method.

An investigation of model sensitivity to various building-component resolutions was performed in this study to evaluate the reaction of model computational load to diverse input data. As shown in Figure 4, a normality test was conducted on the simulation results of the model. Furthermore, according to the inspection results, U-testing of the box figure was carried out to verify the impact of varied input-parameter resolution on the simulation performance. This included energy simulation errors, temperature simulation errors and computational load. In the end, the main findings were given based on the testing results.

Specifically, this innovative U-testing process was conducted using six input parameters for the building: external windows, internal windows, solar shading, occupancy schedule, thermal zoning, and HVAC system. These input parameters for the building had varying modelling resolutions, ranging from low to high (A-B-C). This sensitivity analysis controlled for the six groups using independent variables. Using boxplots, 54 models with A-level first components (external windows) and 54 models with B-level first components (internal windows) were compared, while the same method was utilised for the other five groups. Subsequently, the influence of input data at different resolutions on various modelling input parameters might be observed separately. The notation of resolution in this article is defined as follows: the first to sixth letters represent six model-input parameters, where “A”, “B”, and “C” represent three resolutions from high to low and “X” indicates one parameter that includes all resolutions.

3. Results and Discussion

Following the preceding procedure, 108 model IDF files were executed via the Windows command line while the computational load was tracked. First, Python scripts were used to generate a comparison of computational load versus CVRMSE for all model points, both global and local. The original computational-load data and simulation time for all 108 models can be found in the Appendix A at the end of the article.

Using the CVRMSE and R² values of energy consumption and zone mean air temperature, computational loads as cases, and 108 models as variables, a hierarchical clustering analysis plot was obtained using SPSS software (as shown in Figure 5). The six-digit letter notation for each subregion in the figure illustrates the type of model that dominated the subregion. This figure reflects that the last three input parameters showed a strong correlation and clustering distribution overall while the first three input parameters did not show significant distribution regularity.

Subsequently, error diagrams of the computational load against the CVRMSE for both energy consumption and zone mean air temperature were plotted. After possible erroneous data points were discarded, the preliminary results showed very obvious cluster distribution, regularly distributed in six regions (see Figure 6 and Figure 7).

To proceed further, additional sensitivity analysis and validation will be provided, with the final results reviewed and discussed in accordance with the data provided herein.

3.1. Sensitivity Analysis for Simulated Results of Model Input Parameters

3.1.1. Sensitivity Analysis of Energy Simulation Error

Figure 8 illustrates the data normality test for the entirety of the CVRMSE findings. This graph demonstrates that the data’s normality is low, which may be attributed to inadequate resolution of the six input parameters as a whole; hence, this set of data did not match the standards.

It was appropriate to analyse statistical associations using t-testing (Student testing). As a result, Mann–Whitney testing (U-testing) [74] was used to analyse the statistically greater than, less than, and equal relationships between each data group in the boxplot. This method was selected because the Mann–Whitney rank sum test permits unequal samples and does not need data normality. In this case, the confidence level was set to 95%; therefore, a p-value of less than 0.05 would indicate a statistically significant difference between the two variables being compared, and vice versa. The Figure 9 and

Table 5. The analysis of U-testing for the CVRMSE boxplot (energy consumption).

BEM Input Parameter	Resolution	Statistic	p-Value	Description
External Windows	AXXXXX to BXXXXX	1348.0	0.501	No significant difference found
Internal Windows	XAXXXX to XBXXXX	659.5	0.901	No significant difference found
	XAXXXX to XCXXXX	568.0	0.371	No significant difference found
	XBXXXX to XCXXXX	568.0	0.371	No significant difference found
Solar Shading	XXAXXX to XXBXXX	614.0	0.706	No significant difference found
	XXAXXX to XXCXXX	736.0	0.324	No significant difference found
	XXBXXX to XXCXXX	814.5	0.061	No difference found
Occupancy Schedule	XXXAXX to XXXBXX	2859.5	$7.41\times {10}^{-18}$	Significant difference: the accuracy of the XXXAXX type was considered to be lower than that of the XXXBXX type
Thermal Zoning	XXXXAX to XXXXBX	1894.0	$9.86\times {10}^{-5}$	Significant difference: the accuracy of the XXXXAX type was considered to be lower than that of the XXXXBX type
HVAC System	XXXXXA to XXXXXB	1665.0	0.011	Significant difference: the accuracy of the XXXXXA type was considered to be lower than that of the XXXXXB type

below shows the results of the examination for simulated energy consumption.

3.1.2. Sensitivity Analysis for Accuracy of Simulated Zone Mean Air Temperature

To validate the preliminary conclusion established via using energy consumption as an indicator, the hourly average room temperature of each model and the CVRSME of the corresponding temperature of the most refined model were computed. The Figure 10 and

Table 6. The analysis of U-testing for the CVRMSE boxplot (zone mean air temperature).

BEM Input Parameter	Resolution	Statistic	p-Value	Description
External Windows	AXXXXX to BXXXXX	1395.0	0.701	No significant difference found
Internal Windows	XAXXXX to XBXXXX	637.0	0.906	No significant difference found
	XAXXXX to XCXXXX	576.0	0.421	No significant difference found
	XBXXXX to XCXXXX	613.0	0.680	No significant difference found
Solar Shading	XXAXXX to XXBXXX	650.0	0.987	No significant difference found
	XXAXXX to XXCXXX	648.0	1.000	No significant difference found
	XXBXXX to XXCXXX	686.0	0.673	No significant difference found
Occupancy Schedule	XXXAXX to XXXBXX	1980.0	0.001	Significant difference: the accuracy of the XXXAXX type was considered to be lower than that of the XXXBXX type
Thermal Zoning	XXXXAX to XXXXBX	2592.0	$3.091\times {10}^{-17}$	Significant difference: the accuracy of the XXXXAX type was considered to be lower than that of the XXXXBX type
HVAC System	XXXXXA to XXXXXB	1853.0	0.0001	Significant difference: the accuracy of the XXXXXA type was considered to be lower than that of the XXXXXB type

below shows the results of the examination for simulated zone mean air temperature.

As depicted in Figure 10, the error of each component was less than 10%, as temperature change was more consistent throughout the year; consequently, the form of the boxplot differed somewhat from the energy-consumption map. Nevertheless, U-testing indicated that the statistical results were identical. This indicated that the occupancy schedule, the thermal zoning, and the HVAC system were the most crucial components for achieving optimal model fidelity.

3.1.3. Sensitivity Analysis for Computational Load

The Figure 11 and

Table 7. The analysis of U-testing for the computational-load boxplot.

BEM Input Parameter	Resolution	Statistic	p-Value	Description
External Windows	AXXXXX to BXXXXX	1247.0	0.196	No significant difference found
Internal Windows	XAXXXX to XBXXXX	307.0	0.0001	Significant difference: the computational load of the XAXXXX type was considered to be lower than that of the XBXXXX type
	XAXXXX to XCXXXX	345.0	0.0007	Significant difference: the computational load of the XAXXXX type was considered to be lower than that of the XCXXXX type
	XBXXXX to XCXXXX	969.0	0.0003	Significant difference: the computational load of the XAXXXX type was considered to be higher than that of the XCXXXX type
Solar Shading	XXAXXX to XXBXXX	612.0	0.689	No significant difference found
	XXAXXX to XXCXXX	527.0	0.174	No significant difference found
	XXBXXX to XXCXXX	568.0	0.371	No significant difference found
Occupancy Schedule	XXXAXX to XXXBXX	1403.0	0.738	No significant difference found
Thermal Zoning	XXXXAX to XXXXBX	1642.0	0.054	No difference found
HVAC System	XXXXXA to XXXXXB	2291.0	$2.94\times {10}^{-5}$	Significant difference: the computational load of the XXXXXA type was considered to be lower than that of the XXXXXB type

below shows the results of the examination for computational load.

3.2. Main Findings

Based on the observation and calculation results, the AAABBB type (No. 8) was determined to be the optimal model in theory. According to this finding, the sensitivity-analysis results for each model input parameter can also be given, as shown in

Table 8. Summary of sensitivity analysis of simulated energy and temperature.

Input Parameters for the Building	Sensitivity-Analysis Results
	For Simulated Energy and Temperature	For Computational Load
External Windows	U-testing results indicated that there was no statistically significant difference between the two datasets. According to the hourly CVRMSE analysis, the exterior windows’ resolution had a minor effect on the ultimate energy consumption.	Boosting the resolution of the external windows raised the computational load marginally. The U-testing, on the other hand, revealed no statistically significant difference between the two sets of data.
Internal Windows	Based on the hourly CVRMSE study, the modelling resolution, as with the exterior windows, had a minor effect on the annular electrical consumption. However, when the model with the highest resolution was used, the CVRMSE became unstable, which might have been the result of the complex internal windows interacting with the other input parameters. In addition, the U-test demonstrated that there was no statistically significant distinction between the three datasets. This phenomenon proved that the internal windows were not a substantial contributor to the building’s energy use.	In addition, the U-testing demonstrated that there was a statistically significant difference between the three datasets. Consequently, the resolution of the internal windows displayed an intriguing phenomenon. The intermediate resolution (XBXXXX), the default option for the internal windows, resulted in a noticeable increase in computational load. In addition, U-testing demonstrated that this dataset was, statistically, significantly larger than the other two datasets. This issue could have been the result of the default configuration increasing the number of windows, which could have increased the interaction of heat flow between adjacent zones. In addition, the high resolution (XCXXXX), based on the real scenario, increased the computational load by a small amount.
Solar Shading	Generally identical to the exterior windows. However, the model with the maximum resolution marginally improved the result’s stability.	The U-testing results indicated that there was no statistically significant difference between the three datasets. However, the computing load increased slightly as the model complexity grew. Typically, the computing amount of the low resolution (XXAXXX) decreased significantly in the absence of shading, indicating that the model without shading could reduce interaction between solar lighting and the internal zone to reduce the computational load.
Occupancy Schedule	In contrast to the previous component, the U-testing demonstrated a statistically significant difference between the two sets of data. According to the hourly CVRMSE analysis, the higher the accuracy of the occupancy schedule, the higher the pertinence to the annual electricity consumption.	The U-testing results indicated that there was no statistically significant difference between the two datasets. Therefore, it is difficult to discern the difference in computational load between the low-occupancy and high-occupancy schedules. Surprisingly, the computational load was slightly lower at the B-level input resolution than at the A-level. This could have been due to the increased speed of reading the CSV file directly from the computer’s local directory compared to that of retrieving the information from an IDF file.
Thermal Zoning	Similarly to the occupancy schedule, better resolution of the thermal zoning revealed an obvious impact on annual electrical consumption.	The U-testing revealed that there was no statistically significant difference between the two sets of data. More sophisticated thermal-zoning approaches had much higher computational loads in general, but this trend appeared to decrease when other input factors were equally complex.
HVAC System	In general, the U-testing found a statistically significant difference between the two datasets. Consequently, the effect of HVAC system resolution on energy consumption was undetectable. However, lower resolution is more volatile than higher resolution.	The U-testing indicated that the difference between the two datasets was statistically significant. Consequently, the computational quantity of the high-resolution (XXXXXB) custom HVAC value was much less than that of the low-resolution value (XXXXXA). The autosized setting could have used unique mathematical methods to determine these parameters, but the custom-value setting entered these parameters directly, resulting in less work.

, which follows.

4. Limitations and Future Work

This study analysed in depth how resolution of input parameters affects accuracy and computational load of energy modelling for case buildings. The findings of this investigation have revealed significant knowledge gaps. Reflection on these findings could guide necessary adjustments to modelling procedures, which would advance the subject. The methodology proposed in this study was limited in terms of input-data resolution and its quantification, making expansion to other building types or larger modelling scales challenging. In addition, a closer link must be formed between the accuracy required for the application and the modelling technique, as methodology is typically detached from its application, making it questionable whether the theoretically optimal resolution attained was accurate enough. Therefore, future research must aim to address these modelling-technique issues.

Based on modeller skills and experience, the quantitative indicators for the six energy-modelling input parameters were initially categorised into two or three groups. Despite the difficulty of achieving quantitative indications at the percentage level, a more granular division is still possible. This research’s authors were unable to access hourly data for the full year due to the case building’s early construction year and the absence of a smart-metering system. Instead of using actual energy-consumption statistics, model No. 144, which was the theoretically most accurate model, was chosen as the baseline model. A comprehensive retrofitted smart-metering program would have been challenging to implement due to the number and complexity of the building’s electrical systems. In this regard, it is proposed that older buildings, such as the building in this case study, conduct precise smart metering at the room scale, generate data, and cluster them into room-scale energy models, which could eventually be expanded to larger scales via superposition. To realise further extension of this method to other types of buildings, cluster analysis of architectural prototypes based on different levels of different types of buildings is a necessary prerequisite. Based on the results of these analyses, modellers can be guided to divide resolution of input parameters and generalise most building stock via applying this method to typical buildings.

In practice, however, it would be essential to consider the amount of effort expended by a modeller in addition to the computational load on the system. Some intricate component details, for example, can be generated rapidly via compiling IDF files with DesignBuilder or Python scripts, whilst others require frequent manual modifications by the modeller. In general, it is regarded that the former model is more complex than the latter. Due to varying objective boundaries, it is difficult to quantify manual work. In this scenario, the purpose of this study is to determine the level of rigor of modelling components in terms of input-data resolution and to reduce the effort and computational load of the modeller. Future work will involve ongoing production of automated scripts to construct and run data for BEM and UBEM, thereby reducing the amount of manual labour required.

The current study indicated that BEM is especially susceptible to weather [75], which is mostly unknown and constantly changing. The weather input for this study was based on a centralised weather data file (8760 hourly values), an EPW file received directly from the EnergyPlus website, and a typical meteorological year (TMY) selected from the weather station’s historical weather data [76]. In addition, because the official website does not publish local meteorological documents for Ningbo, the EPW document was for the 50 km-away Dinghai District. Clearly, relying on a single TMY weather file for the entire city is impractical and would have led to errors in urban building-energy models [77]. The inaccuracy of these approaches in depicting the actual states of past meteorological years may have an effect on the microclimates and boundary conditions of cities in UBEM [78].

This study used a high-resolution profile developed by Chen et al. to determine presence and behaviour of occupancy. Although it could be argued that it is closer to a realistic usage scenario than the predetermined DesignBuilder timeline, there is still scope for improvement. Davis and Nutter [79] compiled occupancy profiles for eight university buildings using information from numerous sources (security cameras, doorway count sensors, classroom scheduling data, and human observations). Similar datamining frameworks can achieve a high degree of accuracy [80], but integrating them with energy simulation software remains difficult. Occupancy models based on Markov chains [81,82], stochastic simulations [83], cluster analysis [84], and agent-based modelling (ABM) [85,86,87] have been developed in recent years. Their results are significantly different from those of standard methods, although they are more accurate [88].

In addition, only the fundamental input parameters of building-energy modelling were addressed in this research, despite the existence of numerous additional variables that routinely replace the DB default values. For example, window-opening behaviour was not considered because it is uncommon in most centrally air-conditioned buildings. Similarly, window-shading devices such as curtains and blinds were excluded from the building simulation despite having an impact on the building’s thermal and cooling loads. Future research should focus on classification and quantification of these input properties, as well as development of a more thorough standardisation of building-energy models. In general, the energy-modelling resolution research strategy proposed in this study, as well as the initial findings, serves as an ideal benchmark for low-rise office buildings. Future major work will be required to continue testing the proposed method in the same type of building, make necessary modifications and optimizations, and eventually extend this method to various types of building to aid researchers in determining the appropriate level of abstraction for building-energy modelling.

5. Conclusions

In recent decades, the need to mitigate the effects of energy and climate challenges has increased. In this context, research interest in BEM as a crucial component of the UBEM field has increased. This study investigated the connection between building-energy modelling (BEM) accuracy and computational load. The primary goal of this research was to assess the degree to which external windows, internal windows, solar shading, occupancy scheduling, thermal zoning, and HVAC design affect the accuracy of final simulation results. For this reason, a total of 108 IDF files were constructed, and the computational load incurred during the execution of each simulation was meticulously measured.

The results demonstrated that the performance of the BEM changed as the input parameters’ resolutions varied. As a result, drawing conclusions necessitated a careful study of the combined computed loads, as it is not a single defining feature that determines accuracy but rather the common deficiency of all attributes. The errors in the building for external windows, internal windows, and shading were typically minor, and they were within acceptable limits even at medium- and low-input-parameter resolutions. However, at medium or high resolution, they generated up to 2× more computational load, which was beyond the tolerance threshold. In contrast, for the occupancy schedule, thermal zoning, and HVAC configurations, the medium- and low-input-parameter resolutions resulted in larger errors, with minimal decreases in computational load compared to those of the high-resolution input and individual models being even more computationally intensive. In addition, these results intuitively demonstrated the specific effects of key input-parameter resolution on the final performance of BEM. The practical significance was to deepen modeller understanding of the BEM input parameters so that they can better allocate the limited manual and computational costs when collecting BEM input parameters.

Overall, the lack of defined quantitative procedures for BEM input-parameter resolution and computational-load findings could have a significant effect on the accuracy of BEM. There are neither any consistent reporting practices for error measurement in reporting of other relevant input-parameter resolutions nor accurate measurements of computational load, which made it difficult to compare studies. Similarly, rather than merely evaluating the textbook data presented in studies, it is advised that researchers in the BEM and UBEM industries propose and investigate new error-measuring methodologies. Capturing the dynamic behaviour associated with building energy, such as peak hour consumption and daily maximum temperature, is also recommended, as it is essential for emerging energy system integration and climate resilience applications. As a result of these efforts, BEM and UBEM will eventually become tools for creation of healthy, energy-efficient, and low-carbon communities.