An Enhanced NSGA-II Algorithm with Parameter Categorization for Computational-Efficient Multi-Objective Optimization of Active Glass Curtain Wall Shading Systems

Abstract

To address the limitations of the Non-Dominated Sorting Genetic Algorithm (NSGA-II) in optimizing active glass curtain wall shading systems—particularly its suboptimal convergence efficiency and high computational demands—this study proposes an improved NSGA-II algorithm incorporating parameter categorization. Shading system parameters (e.g., slat width, angle, separation, and blind-to-glass distance) are classified into distinct categories based on their character and optimized sequentially. This phased approach reduces the search space dimensionality, lowering computational complexity while maintaining optimization accuracy. The framework integrates user preferences and climatic adaptability to balance energy efficiency and glare mitigation. The louver parameters were optimized under the same experimental conditions, and the enhanced algorithm exhibits 49% lower energy consumption values and 5% smaller visual discomfort time duration compared to the baseline algorithm in the optimization outcomes.

1. Introduction

The energy consumption during the operation phase of buildings accounts for more than 20% of the total national energy consumption [1]. Glass curtain walls, as the most solar radiation-sensitive component of building envelopes, exhibit particularly high energy demands. While active shading systems effectively reduce cooling loads [2], conventional devices often compromise natural daylighting, increasing artificial lighting consumption [3]. To address this trade-off, intelligent shading systems dynamically adjust shading angles and positions based on parameters such as outdoor illuminance and solar altitude angles, maximizing daylight utilization while maintaining shading efficacy [4]. This dual-benefit approach not only preserves indoor visual comfort but also significantly reduces lighting energy usage, thereby optimizing overall building energy performance [5]. Furthermore, movable louvers demonstrate notable effectiveness in mitigating glare-related discomfort in indoor environments [6]. Generally, dynamic louvers have more performance advantages than static louvers [7], but under some specific conditions, the performance difference between movable louvers and fixed louvers is not significant [8], and even static louvers perform better [9]. These variations stem from diverse architectural configurations, climatic contexts, geographic locations, and user preferences, necessitating the context-specific optimization of movable louver systems.

In the last decade, researchers have begun to combine optimization algorithms with building simulation software to improve the performance of shading systems. For instance, Liu et al. [10] employed the NSGA-II algorithm for multi-objective optimization of fixed louver geometric parameters, achieving simultaneous reductions in energy consumption and improvements in thermal–visual comfort. Zhang et al. [11] reduced the energy consumption of the building and improved the visual comfort of the building through the BO-LGBM algorithm. Ziaee et al. [12] adopted the method of combining simulation modeling and multi-objective optimization to explore the optimal design parameters of a typical classroom light louver with the goal of improving classroom lighting performance. Jung et al. [13] optimized the angle of the slat through an optimization algorithm to improve the illumination inside the building. Fan et al. [14] optimized louver blade quantities across building facades to minimize daylight glare probability (DGP). In addition to the fixed parameters of louvers, some studies have attempted to optimize the variable parameters [15] and control methods [16] of louvers. However, current studies predominantly focus on isolated parameter optimization, lacking systematic integration of design and operational parameters. Chen et al. [17] used machine learning algorithms to optimize the thermal environment of buildings. Moreover, computational inefficiency in building performance simulations remains a critical barrier [18], highlighting the need for algorithmic enhancements to improve optimization feasibility.

A critical oversight in existing research lies in the neglect of occupant preferences during actual building operations [19]. Nezamdoost et al.’s research shows that the behavior of occupants through the manual control of lighting, shading facilities and air conditioning systems to create a better fit for their personal needs will have a great impact on the performance of buildings [20]. As Liu et al. [21] emphasize, user behavior patterns substantially affect shading system efficacy, while Natalia et al. [22] identify glare prevention as the primary motivator for manual shading adjustments. These findings underscore the imperative to incorporate user preferences into shading system design processes. In addition, the current architectural shading design framework based on artificial simulation and algorithmic optimization still exhibits notable limitations. The simulation-driven optimization process imposes substantial computational demands, particularly when addressing large-scale spatial configurations in building design. This technical constraint manifests as prolonged computational durations, which results in compromised optimization efficiency and thereby hinders their practical implementation in real-world engineering projects [23].

To address these challenges, in this paper, an improved NSGA-II algorithm is proposed for multi-objective optimization of glass shading systems. By coordinating the optimization of fixed and variable louver parameters at different stages, we have developed a comprehensive design framework that balances building characteristics and user preferences. Key innovations include: (1) The NSGA-II algorithm is improved, and the phased optimization method is adopted to greatly reduce the total search space size, reduce the computing resources required in the algorithm optimization process, and improve the optimization efficiency; (2) Distributed parameter optimization for adaptive design solutions; (3) Integration of user preference factors to enhance practical applicability and provide decision-makers with diversified options.

2. Establishment of Building Model

The workflow commenced with establishing the corresponding building model based on the experimental site and objects. Subsequent experimental investigations were conducted to collect datasets, followed by numerical computations using EnergyPlus to analyze indoor thermal conditions. These computational outputs underwent rigorous comparative analysis with experimental measurements through a data-model validation framework to verify simulation accuracy. Finally, operational parameters were systematically configured in accordance with designated building typologies and regional design specifications, ensuring compliance with optimization algorithm requirements.

2.1. Experimental Object

The case room in this paper is an air-conditioned room, with an open surrounding environment without any buildings or landscape obstructions. The south side of the room features a glass curtain wall, with full-sized active shading louvers installed externally. The angle of the blind slats and the distance between the louvers and the curtain wall are both adjustable. Figure 1 shows the schematic diagram of the experimental room’s exterior.

The experimental house is located in Zhuzhou, Hunan, which belongs to the subtropical monsoon climate. Figure 2 shows the annual temperature changes in Zhuzhou over the past five years, and Figure 3 shows the annual solar radiation changes in the area. From the patterns of monthly maximum temperatures and solar radiation intensities, it can be seen that the cooling load in summer is very high, while the heating load in winter is relatively low, with significant differences in solar radiation levels between winter and summer. Therefore, it is necessary to formulate corresponding control strategies for the active shading system according to different seasons. Due to the short duration of the transitional seasons in the area, this paper categorizes different months of the transitional seasons into summer and winter control periods based on their radiation levels to simplify the control strategies of the louvers.

2.2. Validation Model

To validate the accuracy of the building model, this study designed and conducted systematic experimental tests with slat angles and blind-to-glass distance as key variables. These parameters were systematically manipulated to evaluate their influence on indoor thermal dynamics, with experimental data subjected to rigorous comparative analysis against simulation outputs. Eight experimental groups were configured, each representing distinct parameter combinations as detailed in

Table 1. Design variables parameters.

Experimental Group	Angle	Distance from Glass	Experimental Time
a	60°	5 cm	July
b	90°	15 cm	July
c	135°	5 cm	July
d	60°	15 cm	July
e	60°	15 cm	November
f	30°	10 cm	November
g	135°	5 cm	November
h	150°	10 cm	November

. Groups a–d correspond to summer experimental conditions, while groups e–f represent winter scenarios.

The parameter selection followed an empirically validated protocol derived from preliminary studies and engineering practices, ensuring representative sampling within conventional ranges: slat angles spanned 30–150°, and blind-to-glass distances varied between 5 cm and 15 cm. This stratified parameterization enabled comprehensive coverage of typical architectural configurations while maintaining experimental validity. Figure 4 shows the slat geometry of the louver parameters in this study.

To ensure precise measurement of the dry bulb temperature within the occupied zone of the experimental room, six thermistor sensors were strategically deployed, as depicted in Figure 5. These sensors were positioned to comprehensively capture temperature fluctuations across the areas where occupants are active. The average of the temperature readings from these sensors was utilized to represent the room’s dry bulb temperature at any given time. Furthermore, a weather station was installed outdoors near the laboratory to record key meteorological parameters during the experiment, including temperature, humidity, wind speed, and solar radiation. The primary equipment employed in the study is illustrated in Figure 6, with their key parameters summarized in

Table 2. Main equipment and their function.

Equipment	Function	Device Model	Range/Sensitivity
Meteorological station	Record of meteorological conditions	JLG-QTF, JinZhouLiChen, JinZhou, China	Total solar radiation: 0~2000 w/m2, ±0.5 w/m2; Temperature: −40~120 °C ± 0.5 °C; humidity: 0~100 %RH, 1 %RH ± 3 %RH; wind rate: 0~70 m/s
Thermistor sensors	measure temperature	WRNT-010, HangZhouHongDa, HangZhou, China	±0.1 °C
Paperless recorder	Data collection and preservation	WPR50A-48XUSBVO, SuZhouXunPeng, SuZhou, China	24 V, ±0.2%

.

The case room in this paper is an air-conditioned room, as illustrated in Figure 1. The room dimensions are 3 m in length, 1.8 m in width, and 2.6 m in height, with its building type designated as an office building. The curtain wall is oriented southward, featuring a transparent facade composed of 3 mm thick glass. The glazed component measures 1.4 m in width and 2 m in height, with its optical parameters—Glass U-Factor [W/m²·K], Glass SHGC, and Glass Visible Transmittance specified as 5.894, 0.862, and 0.899, respectively. Key parameters of other primary building structural components are provided in

Table 3. Structure and thermal parameters of enclosure structure.

Construct	Description	Heat Transfer Coefficient (W/(m2∙k))
Roof	cinder concrete and reinforced concrete (90 mm thickness)	1.05
Wall	Extruded polybenzene board (30 mm thickness) + cement mortar (20 mm thickness)	0.72

.

The experimental meteorological data underwent temporal alignment processing in strict accordance with the EnergyPlus EPW file format specifications. Dynamic thermal simulations were executed using EnergyPlus v24.1; the results of the experiment and simulation are shown in Figure 7 and Figure 8, with model validation performed following the ASHRAE Standard 140-2020 [24] testing protocols. Data quality control adhered to the Standard for Human Behavior Data in Building Energy Simulation (T/CABEE 042-2023 [25]). A dual-metric evaluation system was implemented:

• Root Mean Square Error:

$$RMSE=\sqrt{{\displaystyle \frac{\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}{n}}}$$

(1)

2.

Normalized RMSE:

$$NRMSE={\displaystyle \frac{RMSE}{\overline{y}}}$$

(2)

Numerical results demonstrated:

• Summer condition: RMSE = 0.138 °C (NRMSE = 0.00359)
• Winter condition: RMSE = 0.0964 °C (NRMSE = 0.00403)

All NRMSE values remained below the threshold limit of 0.5%, satisfying the Class III accuracy requirements specified in the National Standard for Civil Building Energy Consumption (GB/T 51161-2016) [26]. The finalized model achieved a 95% prediction confidence interval (p < 0.05), qualifying as a reliable input source for genetic algorithm optimization while meeting the advanced precision criteria prescribed in the Standard Guidelines for Building Model Design and Optimization. This model will be used as the basic model in the following algorithm optimization.

2.3. The Building’s Internal Loads and the HVAC System

After establishing the building model, appropriate internal parameters were set based on office buildings. First, according to the “Building Lighting Design Standards”, the room lights are turned on when the illuminance is less than 300 lx, with a lighting power density set at 11 W/m². Second, the occupancy density is set at 4 m²/person, and according to ASHRAE 55 recommendations, the activity level of occupants is set at 126 W per person. Finally, the air conditioning cooling temperature is set at 24 °C, and the heating temperature is set at 18 °C. More input parameters are shown in

Table 4. Input parameters used in the EnergyPlus simulation.

Location	Zhuzhou
Latitude, Longitude, Height	113.15, 27.84, 100.00
Orientation	South
Equipment heat gain	Light: 11 W/m2; 0:00–24:00
Occupation	4 m2/person
People radiant fraction	0.3
Activity level	126 w/person
Heating thermostat setpoint	18 °C
Cooling thermostat setpoint	24 °C

.

In addition, louver control settings were established according to different times of day. Based on the variation patterns of solar altitude angles, the day is divided into three periods: 8:00 to 11:00 is Period A, 11:00 to 14:00 is Period B, and 14:00 to 18:00 is Period C. In Periods A and C, the solar altitude angle is lower, while in Period B, it is higher. Different slat angles are set for each time period. During nocturnal operational phases, the dynamic louver system was configured to achieve a near-complete closure state (95% occlusion ratio) to optimize thermal buffering performance.

Table 5. The daily schedule for adjusting the slat angle.

Work Schedule Period	Time	Value	Unit
A	8:00–11:00	A1	°
B	11:00–14:00	A2	°
C	14:00–18:00	A1	°
N	17:00–24:00, 0:00–8:00	5	°

shows the daily schedule for adjusting the slat angle. During the optimization process, the algorithm can adjust the values of louver variable parameters for different time periods.

3. Methodology

Shading blinds offer dual functions of improving indoor light and thermal environment while reducing building energy consumption. By utilizing the blinds, natural lighting in buildings can be improved to reduce the use of artificial lighting, thereby decreasing electricity consumption [2]. On the other hand, controlling sunlight entering the room can also reduce glare [6], which stimulates the user’s vision. Achieving a comfortable light environment and low building energy consumption clearly pose a conflict and cannot simultaneously reach a theoretically optimal state. This is a multi-objective optimization problem, where optimization goals are in conflict with each other. Improving one goal may lead to the degradation of one or more other goals.

3.1. Multi-Objective Optimization

Deb [27] and others proposed an improved NSGA-II genetic algorithm based on NSGA, which has been widely applied in multi-objective optimization problems. The advantages of this algorithm include:

• Reducing Computational Complexity: NSGA-II introduces a fast non-dominated sorting method, reducing the algorithm’s computational complexity from O(MN³) to O(MN²), where M is the number of objectives and N is the population size.
• Maintaining Population Diversity: NSGA-II uses a crowding comparison operator to estimate the crowding degree among individuals. Within the same non-dominated level, individuals with higher crowding degrees are more likely to be selected, preventing the algorithm from prematurely converging to local optimal solutions.
• Introducing the Elite Retention Mechanism: The offspring generated by individuals selected for reproduction compete with their parent individuals to form the next generation population.

In this paper, the software JEPlus+EA (v2.2) [28] was used for multi-objective optimization. Lara et al. [29] calibrated and validated the JEPlus+EA software by optimizing 72,000 EnergyPlus models, providing a robust framework for coupling optimization algorithms with building energy simulations. Firstly, the JEPlus (v2) program was used to input meteorological parameters and simulation data to obtain objective functions. Then, the algorithm model was generated by setting constraints and other parameters. Finally, JEPlus+EA utilized the NSGA-II genetic algorithm for optimization, obtaining the Pareto front. Through multiple adjustments of the algorithm, the population size, maximum number of generations, crossover rate, and mutation rate were set to 50, 80, 100%, and 20%, respectively.

3.2. Visual Discomfort Time

Sunlight entering through external windows during the day may benefit the lighting environment inside the building, but too much sunlight can also cause discomfort to the occupants. The Discomfort Glare Index (DGI) is used by EnergyPlus to determine inhabitants’ visual comfort [30]. The following Equation (3) is the formula that defines DGI [31]:

$$\mathrm{D}\mathrm{G}\mathrm{I}=10\log 10\left[0.478{\sum }_{\mathrm{i}=0}^{\mathrm{n}}\left({\displaystyle \frac{{\mathrm{L}}_{\mathrm{s}\mathrm{i}}^{1.6}{\Omega }_{\mathrm{s}\mathrm{i}}^{0.8}}{{\mathrm{L}}_{\mathrm{b}}+0.07{\Omega }_{\mathrm{s}\mathrm{i}}{\mathrm{L}}_{\mathrm{w}\mathrm{i}\mathrm{n}}{\mathrm{P}}_{\mathrm{i}}^{1.6}}}\right)\right]$$

(3)

where the definitions of L_b, L_si, and L_win are the background luminance, the glare source, and the window luminance, and their units are cd/m². Ω_si represents the solid angle encompassing the glare source from the occupants’ perspective, adjusted by Guth’s position index. Additionally, n can be defined as the number of glare sources [31]. For office buildings, the maximum permissible DGI threshold is set at 22 [30]. Detailed DGI threshold values are provided in

Table 6. DGI threshold values.

Glare Characterization	DGI
Intolerable	>28
Only intolerable	28
Uncomfortable	26
Only uncomfortable	24
Acceptable	22
Only acceptable	20
Perceptible	18
Only perceptible	≦16

. In this study, the duration during which DGI exceeded the set threshold was defined as visual discomfort time (unit: hr).

3.3. Building Energy Consumption

The building energy consumption is divided into different parts. Electricity for artificial lighting electricity and air conditioning system electricity used for cooling and heating office space are included. Equation (4) shows the building energy consumption.

$$\mathrm{E}\mathrm{n}\mathrm{e}\mathrm{r}\mathrm{g}\mathrm{y} \mathrm{c}\mathrm{o}\mathrm{n}\mathrm{s}\mathrm{u}\mathrm{m}\mathrm{p}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}=\mathrm{l}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}\mathrm{i}\mathrm{n}\mathrm{g} \mathrm{e}\mathrm{l}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}+\mathrm{A}\mathrm{i}\mathrm{r} \mathrm{c}\mathrm{o}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{i}\mathrm{n}\mathrm{g} \mathrm{s}\mathrm{y}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{m} \mathrm{e}\mathrm{l}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}$$

(4)

3.4. Objective Function

In this paper, the optimization objectives are set as the lowest building energy consumption and the lowest visual discomfort time. The objective function can be expressed as:

$$\mathrm{F}=(\mathrm{m}\mathrm{i}\mathrm{n}{f}_1,\mathrm{m}\mathrm{i}\mathrm{n}{f}_2)$$

(5)

where the definitions of f₁ and f₂ are the building energy consumption and the visual discomfort time. f₁ and f₂ are derived by simulating buildings using the EnergyPlus software, which is driven by JEPlus.

3.5. Bound Variable

To ensure that the results calculated by the algorithm are realistic, constraints have been set for various variables in the model:

• Slat Width Constraint: 0.01 m ≤ B ≤ 0.2 m;
• Slat Angle Constraint: 20° ≤ A ≤ 160°;
• Blind to glass distance Constraint: 0.01 m ≤ D ≤ 0.25 m;
• Slat Separation Constraint: 0.01 m ≤ S ≤ 0.2 m.

Where the definitions of f₁ and f₂ are the building energy consumption and the visual discomfort time. f₁ and f₂ are derived by simulating buildings using the EnergyPlus software, which is driven by JEPlus.

3.6. Improved Optimization Algorithm

The inherent complexity of building system optimization necessitates rigorous computational methodologies to ensure solution fidelity. While simulation platforms (e.g., EnergyPlus, TRNSYS) enable robust objective function formulation, their significant computational demands often raise concerns about optimization efficiency. Furthermore, multi-variable optimization problems with interacting parameters pose inherent challenges, which can be exacerbated by reduced population sizes in evolutionary algorithms. To address these limitations, systematic design of objective functions, algorithm selection tailored to problem characteristics, and diversity-preserving strategies that balance exploration–exploitation trade-offs are critical for improving convergence rates and optimization outcomes.

The optimization process adopts a two-stage parametric approach, and the flow chart of its work is shown in Figure 9. Phase I optimizes static louvers using four geometric parameters (width, angle, spacing, and glass distance) to establish Pareto front solutions. Through systematic data processing, optimal fixed parameters are subsequently selected from the phase I Pareto frontier. Phase II integrates selected fixed parameters into the model, optimizing variable parameters (angle and distance) under operational constraints. Finally, the control strategy for dynamic louver operation is formulated by analyzing the variable parameter combinations derived from the second-stage Pareto front solutions. By improving the algorithm, the theoretical maximum population size is greatly reduced, thus decreasing computational complexity. Simultaneously, after completing the fixed parameter optimization, a fixed parameter more aligned with the user’s optimization preferences will be more advantageous in obtaining a solution that satisfies the user’s optimization objectives.

4. Results

4.1. The Result of Optimization of Fixed Parameters

Figure 10 presents the Pareto front for the multi-objective optimization of fixed louver parameters. As shown, the optimization algorithm provides multiple solutions, all of which are non-dominated solutions, meaning that there is no absolute superiority or inferiority in their values. To find the optimal solution within the Pareto front, this paper combines the weighted sum method with the comparison of the adaptability of different fixed parameters to select the best fixed parameters.

The weighted sum method is frequently used to determine the final result from the obtained Pareto solution set [32,33,34], and it is based on Equation (6):

$$f_{ws}\left(x\right)={\sum }_{i=1}^n{a}_i{\displaystyle \raisebox{1ex}{$f_i\left(x\right)-{\mathrm{m}\mathrm{i}\mathrm{n}(f}_i\left(x\right))$}\!\left/ \!\raisebox{-1ex}{${\mathrm{m}\mathrm{a}\mathrm{x}(f}_i\left(x\right))-{\mathrm{m}\mathrm{i}\mathrm{n}(f}_i(x\left)\right)$}\right.}$$

(6)

The objective function values are denoted as f_i(x), with their maximum and minimum values represented by max(f_i(x)) and min(f_i(x)), respectively. Using weighting coefficients a_i, two distinct weighting scenarios were examined:

• Energy consumption weight (a₁) = 0.4 vs. daylight discomfort duration weight (a₂) = 0.6;
• Energy consumption weight (a₁) = 0.6 vs. daylight discomfort duration weight (a₂) = 0.4

Equation (2) was applied to calculate weighted values for all Pareto-optimal solutions, followed by hierarchical sorting. The minimal weighted solutions yielded:

• Scenario G: Louver width = 0.01 m, louver spacing = 0.19 m;
• Scenario E: Louver width = 0.12 m, louver spacing = 0.09 m.

This dual-weight analysis simulates practical design priorities: Scenario G emphasizes daylight environment optimization; Scenario E prioritizes energy efficiency as the primary objective.

4.2. The Result of Variable Parameter Optimization

Following the completion of Phase I optimization, the fixed parameters from Scenario E (energy-optimized configuration) and Scenario G (visual comfort-optimized configuration) were independently integrated into the computational model for Phase II optimization. In this phase, the optimization variables transitioned to:

• Angle θ₁: Louver orientation during period B;
• Angle θ₂: Shared orientation for periods A and C;
• Blind to glass distance.

The temporal configuration rationale stems from solar geometry symmetry: periods A (morning) and C (afternoon) exhibit mirroring solar altitude patterns; horizontal louver configuration enables operational consolidation without compromising optimization effectiveness. This strategic merging reduces computational complexity while maintaining solution integrity. Figure 11 and Figure 12 illustrate the optimization outcomes of variable parameters for the selected fixed louver configurations under two distinct preference scenarios.

Furthermore, this study conducted a comparative analysis of the original NSGA-II algorithm’s performance against the enhanced version. Benchmark evaluations were performed under identical algorithmic parameters: population size = 50, maximum generations = 160, crossover rate = 100%, and mutation rate = 20%. This parametric consistency ensures a rigorous comparison of computational efficiency and solution quality between the baseline and modified algorithms while eliminating confounding variables associated with parameter selection. Figure 13 illustrates the Pareto front derived from optimization using the unmodified NSGA-II algorithm.

The results demonstrate that optimizing variable parameters in the second phase significantly enhances the overall performance of fixed-parameter louver systems. For instance, in Scenario E, applying Equation (6) with weighting coefficients of 0.4 for building energy consumption and 0.6 for visual discomfort time duration yielded an optimal louver control strategy through the weighted optimization approach. The comparative analysis included processing Pareto front solutions from the unmodified NSGA-II algorithm under identical conditions. Figure 14 presents a tripartite comparison of: (1) post-first-phase optimization, (2) post-second-phase optimization, and (3) results from the unmodified algorithm. The variable parameter optimization achieved substantial performance improvements, reducing building energy consumption by 49% and visual discomfort time by 5%. The values of the variables bound to the three results are shown in

Table 7. Variables Associated with Divergent Outcomes.

	B (m)	S (m)	A1 (°)	A2 (°)	D (m)	Energy Consumption (kWh)	Visual Discomfort Time (hr)
First-phase	0.12	0.09	40	40	0.13	1462.80	1830.17
Second-phase	0.12	0.09	125	35	0.25	743.99	1741.67
Unmodified algorithm	0.16	0.12	55	40	0.24	1460.11	1832.26

. Notably, the unmodified algorithm’s optimization efficacy remained comparable to Phase I results, demonstrating inferior performance relative to the enhanced algorithm’s final output. This demonstrates that achieving the theoretical Pareto front with the unmodified algorithm requires exponentially increased computational effort (e.g., doubling iteration counts), and the enhanced algorithm attains comparable solutions with markedly improved computational efficiency.

Figure 15 analyzes the distribution of the optimization results of variable parameters with different fixed parameters. The results show that the energy consumption of Scenario E is lower, but the frequency of uncomfortable lighting is higher, while Scenario G is the opposite. This indicates that in the optimization with fixed parameters, through biased selection, the algorithm can better meet the subjective needs of the designer. This demonstrates that the improved algorithm can more accurately meet the optimization requirements of the user.

5. Conclusions

This paper proposes an improved NSGA-II algorithm based on reducing computational complexity to solve the multi-objective optimization problem of the active shading system for glass curtain walls. By conducting step-by-step optimization of different types of parameters of the active shading louvers, it aims to reduce building energy consumption while optimizing the building’s light and thermal environments. For this purpose, a certain air-conditioned experimental room in Zhuzhou is taken as a research case, and the multi-objective optimization of its active shading system is carried out using software such as EnergyPlus and JEPlus+EA.

Under the same conditions, the optimized results of the improved algorithm are 49% lower than the unmodified algorithm in building energy consumption and 5% lower in light discomfort time. The improved algorithm demonstrates the capability to derive high-performance design parameters for active shading systems, achieving a balance between solution quality and computational efficiency. Optimizing the active shading system with the improved algorithm can significantly reduce building energy consumption, improve the light environment inside the building, and make the obtained results more accurately meet the optimization needs of users. This method provides an effective approach for the design of active shading systems for glass curtain walls and ordinary exterior windows and has practical significance.

Future work could explore the algorithm’s adaptability to diverse climatic conditions, such as tropical or arid regions, where shading strategies differ significantly. This aligns with the call by Hashemi (2014) for climate-specific optimization frameworks in automated shading systems [4].