Rapid Evaluation of University Classrooms Using an MLP Classification Model Based on Daylight–Thermal Performance

Abstract

Classrooms in severe cold regions face the dual challenge of ensuring high-quality daylighting while minimizing heating energy consumption. To address this challenge, this study develops a data-driven workflow that integrates building performance simulation, multi-objective optimization and a classification-based surrogate model, aiming to explore integrated improvements in daylighting and heating energy consumption in university classrooms. The results show that: (1) multi-objective optimization significantly enhances overall performance. Daylighting performance improves, with Spatial Daylight Autonomy (sDA) and Useful Daylight Illuminance (UDI) increasing by 0.15 and 10.67%, respectively, and Daylight Glare Probability (DGP) decreasing by 16.35%. Meanwhile, Heating Energy Consumption (E_h) is reduced by 6.20 kWh/m²; (2) SHAP analysis further identifies classroom depth, height, and glazing option as key design parameters influencing integrated daylight–thermal performance; (3) the MLP classification model achieves stable predictive accuracy, with accuracy, recall, and F1-score exceeding 0.95, demonstrating strong generalization ability. This study provides quantitative insights into the relationship between spatial parameters and daylight–thermal performance, offering researchers a method for rapidly evaluating design schemes at the early design stage.

1. Introduction

1.1. Background

Urbanization in China has made building energy consumption a significant issue [1]. Educational buildings account for approximately 8% of the nation’s total energy consumption. The per capita energy use of students is four times the national average for residential areas, and this student energy use has shown a continuous upward trend [2]. As high-occupancy spaces with intensive daily use, educational buildings are the primary contributors to energy consumption on university campuses, while also offering considerable potential for energy savings [3].

Classrooms are the primary spaces for educational activities in universities [4]. They require sufficient daylighting and a comfortable thermal environment [5,6], both of which directly influence students’ visual health, learning efficiency [7], and emotional well-being [8]. A well-designed indoor environment can alleviate visual fatigue, enhance attention and cognitive engagement, and foster a positive and healthy learning atmosphere, thereby indirectly contributing to improved learning motivation and psychological well-being [9].

However, during the early stages of building design, classroom spaces often involve complex interactions between various morphological parameters and performance objectives [10], making it difficult to achieve optimal daylight–thermal performance. In severe cold regions, design strategies often prioritize heat-loss mitigation, for example, by reducing the window-to-wall ratio to limit transmission losses [11]. While effective in improving thermal performance, these measures often compromise daylighting, thereby intensifying the trade-off between daylight demand and heating energy consumption [12]. Consequently, investigating the integrated optimization of daylighting and heating energy consumption in university classrooms in severe cold regions has become a critical and highly relevant research topic.

1.2. Related Work

1.2.1. Research on Classroom Daylighting and Energy Consumption

With advances in computing and machine learning, the quantitative assessment of building performance has become increasingly feasible, and performance-driven computational design methods are playing an increasingly important role in architectural practice. Ding et al. [13] investigated classrooms in universities located in Shenyang, a severe cold region in China, examining the influence of window geometry on carbon emissions associated with daylighting performance. Their results indicated that a west-facing drawing classroom in a 4.2 m-high frame-structure building achieved the lowest lighting-related carbon emissions when the window height, sill height, and window width were 2.55 m, 0.75 m, and 9.6 m, respectively. Moreover, increasing the window size within a reasonable range effectively reduced emissions. Shi et al. [14] conducted a two-month field study involving 89 architecture students in a cold region and found that the neutral operative temperature in design classrooms was 23.1 °C, with the Predicted Mean Vote (PMV) demonstrating good predictive accuracy for thermal comfort. Wang et al. [15] conducted a field study in university classrooms in Harbin to examine the influence of indoor and outdoor climatic conditions on human thermal adaptation. The results indicated that, in Harbin, winter indoor temperatures should be maintained near the lower limit of the thermal comfort range, with setpoints set lower in winter than in spring. This control strategy can reduce the energy consumption of district heating systems by approximately 9.6%. Jia et al. [16] assessed the impact of interior spatial geometry and window properties on daylighting performance using sDA and UDI as metrics. Their findings recommended a window-to-wall ratio (WWR) of 0.3–0.5 for south-facing classrooms and 0.6–0.7 for north-facing classrooms in severe cold regions. In addition, numerous studies have investigated daylighting and energy consumption in offices [17,18], libraries [19,20], hotels [21] and university campuses [22], providing design strategies for improving environmental performance across different building types.

In summary, existing research on classroom building performance still faces two main limitations. First, many studies on university educational buildings rely mainly on field measurements and questionnaire surveys, while the application of performance simulation tools remains limited. Consequently, such studies often address only individual performance indicators and do not adequately account for the complex interactions among multiple design objectives. Second, traditional manual simulation workflows are computationally inefficient, making them unsuitable for multi-variable and multi-objective optimization and thereby constraining comprehensive performance-driven design exploration.

1.2.2. Multi-Objective Optimization

The above-mentioned approaches show limited efficiency in evaluating classroom environmental performance. With the growing recognition that classroom design constitutes a complex decision-making problem involving multiple interacting performance objectives, an increasing number of researchers have adopted multi-objective optimization methods to support performance-driven design. Qin et al. [23] focused on summertime overheating in university classrooms located in China’s severe cold regions and investigated the overheating mitigation potential of an integrated shading system. The results showed that the integrated shading system can reduce overheating hours by up to 59.2%, while maintaining adequate daylighting performance and limiting increases in energy consumption. Zhai et al. [24] optimized window design parameters using a multi-objective approach and employed Pareto-based methods to identify optimal configurations. Their study demonstrated that the process provides designers with insights into the influence of window parameters on energy use and daylight–thermal performance, enabling the selection of configurations that reduce energy consumption while improving indoor daylight–thermal performance. Liu et al. [25] proposed optimized classroom design schemes for the five thermal design zones in China. Their results indicated that an orientation of 5° east or 5° west of due south is conducive to balancing energy consumption and daylighting performance in both small and large classrooms. Lakhdari et al. [26] conducted parameter optimization for secondary school classrooms in Algeria’s hot–dry climate, focusing on window-to-wall ratio (WWR), glazing type, wall materials, and shading devices. Their results demonstrated improvements in UDI, thermal comfort, and reductions in energy consumption. Khani et al. [27] optimized classroom design variables, including window-to-wall ratio and building orientation, for the hot and humid climate of Qeshm Island in Iran. The results indicated that the configuration combining open corridors with dual-sided daylighting yields the most significant performance improvement, with UDI increasing by 34.84% and PMV decreasing by 19.44%.

In summary, multi-objective optimization has shown strong potential for addressing multiple building performance objectives in architectural design and has become an effective approach for classroom design in severe cold regions. However, simulation-based optimization still requires computationally intensive evaluations of thousands of design alternatives. Because such simulations involve complex calculations across multiple performance dimensions [28] and require repeated parameter adjustment, result verification, and data processing, the workflow remains time- and labor-intensive, thereby limiting the overall efficiency of performance-driven optimization [29].

1.2.3. Machine Learning-Based Approaches for Building Performance Prediction

With the rapid development of machine learning techniques, ML-based surrogate models have become a promising approach for addressing the limitations of conventional simulation workflows [30]. By learning from large datasets generated through prior simulations, surrogate models are capable of capturing the mapping relationships between design parameters and performance indicators. Once trained, these models can serve as an efficient alternative to traditional simulation procedures, generating high-accuracy performance predictions within a fraction of the time required for physics-based simulations. This substantially reduces computational burden and improves efficiency while maintaining predictive reliability [31]. As a result, ML models effectively mitigate the high computational cost associated with conventional simulation methods [32] and provide an efficient pathway for supporting design optimization.

Recent studies have demonstrated the strong potential of ML-based surrogate models in building performance prediction and design optimization. For example, Zheng et al. [33] developed a customized artificial neural network for data-driven architectural form generation and analysis. Chen et al. [34] integrated transfer learning with machine learning algorithms to optimize photovoltaic façades for prefabricated educational buildings, significantly improving prediction efficiency and multi-objective optimization performance. Zhang et al. [35] proposed an improved neural network model for predicting the performance of integrated shading systems, achieving better predictive accuracy than several conventional ML models. In addition, Shi et al. [36] combined a GAN and a DRL model to optimize natural ventilation and daylighting in dormitory buildings, demonstrating strong predictive performance and robust stability. Ge et al. [37] further integrated a GA-BP neural network with NSGA-II for rooftop solar optimization in railway station buildings, substantially improving radiation prediction accuracy.

These studies confirm the effectiveness of ML-based surrogate models in reducing computational cost and supporting performance-driven optimization. However, their application to integrated daylight–thermal performance evaluation and optimization of university classrooms in severe cold regions remains limited, with most existing studies focusing primarily on regression-based prediction rather than classification-oriented rapid screening.

1.3. Research Gap and Contributions

In summary, although existing studies have made valuable contributions, several limitations remain. First, some studies focus on a single performance dimension, either daylighting or thermal performance, and lack a systematic investigation of integrated daylight–thermal performance. Second, many existing technical workflows still rely on inefficient conventional simulation processes, and only a limited number of studies have integrated multi-objective optimization algorithms with ML-based surrogate models. As a result, a closed-loop “simulation–optimization–prediction” framework for multi-performance evaluation has not yet been fully established. Third, most existing studies have concentrated on regression-based models, while classification-based approaches for the rapid identification of high-performance design alternatives have received relatively limited attention. To address these gaps, this study integrates the NSGA-II algorithm with machine learning techniques to establish an optimization and rapid evaluation pathway for the integrated daylight–thermal performance of university classrooms in severe cold regions. The main contributions are summarized as follows:

• A rapid evaluation method was developed for classroom design in severe cold regions to support early-stage assessment of integrated daylight–thermal performance.
• Analysis of the Pareto-optimal scheme set identified optimal ranges of key design parameters, and SHAP analysis further revealed the major spatial factors influencing daylighting and heating energy performance.
• A classification model was introduced to provide a more intuitive evaluation of design scheme performance, extending beyond the predominant regression-based approaches used in previous studies.

2. Methodology

2.1. Overall Research Framework

This study aims to provide architects with an efficient early-stage decision-support method for selecting high-performance design alternatives, thereby enhancing the scientific rigor and efficiency of classroom design in severe cold regions. The overall research framework is illustrated in Figure 1 and consists of three main stages. In the first stage, a systematic field investigation was conducted to identify the key design factors influencing classroom performance, based on which a parametric prototype model of the classroom was established. In the second stage, multi-objective optimization was performed using daylight–thermal performance metrics, including Spatial Daylight Autonomy (sDA), Useful Daylight Illuminance (UDI), Dynamic Glare Probability (DGP), and heating energy consumption (E_h). Based on the Pareto-optimal schemes, favorable ranges of key design parameters were identified. In addition, 2500 historical schemes were labeled through a Pareto-filtering mechanism to construct a classification dataset for model training. In the third stage, an MLP classification surrogate model was developed using the classification dataset, and SHAP analysis was conducted to reveal the key factors influencing daylighting and heating energy performance.

2.2. Simulation of Daylighting and Heating Energy Performance

This study focuses on Daqing, a representative city in the severe cold region of northern China. Daqing experiences long and extremely cold winters [38], short daylight duration, and low solar radiation levels [39]. As a result, heating energy consumption is substantially higher than cooling demand. In this context, achieving an effective balance between daylighting and heating energy consumption is particularly important, and the following analysis explores the integrated optimization relationship between these two performance dimensions.

2.2.1. Daylight–Thermal Performance Metrics

• Daylighting Performance Metrics

Daylighting performance indicators can be classified into static and dynamic metrics. Given the limited daylight availability in Daqing, dynamic daylighting metrics provide a more accurate representation of annual daylight availability and temporal variations. Therefore, this study adopts three dynamic annual metrics—sDA, UDI, and DGP—as the evaluation metrics for daylighting performance.

sDA, introduced by the Illuminating Engineering Society (IES), is an annual daylighting metric that describes the percentage of floor area that receives sufficient daylight illumination throughout the year. Previous studies suggest that when at least 50% of a space receives illuminance levels greater than or equal to 300 lx, visual comfort and occupant satisfaction are considerably improved [40,41]. Accordingly, this study employs the sDA_300lx/50% method to evaluate daylighting performance from a spatial perspective. The calculation equation is expressed as follows:

$$\mathrm{s}\mathrm{D}{\mathrm{A}}_{300\mathrm{l}\mathrm{x}/50\%}={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}\mathrm{ST}\left(\mathrm{i}\right)}{\mathrm{N}}}$$

(1)

$$\mathrm{f}\left(\mathrm{x}\right)=\left\{\begin{array}{c}1, \mathrm{s}{\mathrm{t}}_{\mathrm{i}}\ge \mathsf{\omega }{\mathrm{T}}_{\mathrm{i}}\\ 0, \mathrm{s}{\mathrm{t}}_{\mathrm{i}}\le \mathsf{\omega }{\mathrm{T}}_{\mathrm{i}}\end{array}\right.$$

where ST(i) represents the number of hours during which the illuminance at the i-th sensor point exceeds the sDA threshold of 300 lx; Ti denotes the total number of annual time steps (h); and ω is the threshold of the required time percentage (50%).

UDI is a dynamic daylighting metric calculated based on illuminance values on the working plane (0.75 m above the floor). It is used to evaluate daylight availability within an effective illuminance range. UDI represents the proportion of time during which the daylight illuminance at given sensor points falls within an acceptable range throughout the year [42]. Following relevant literature [43] and national standards, this study defines the effective UDI range as 450–2000 lx. The calculation equation is expressed as follows:

$$\mathrm{UD}{\mathrm{I}}_{450–2000\mathrm{l}\mathrm{x}} = {\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}\mathsf{\alpha }\left(\mathrm{i}\right)\cdot \mathrm{t}\left(\mathrm{i}\right)}{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}\mathrm{t}\left(\mathrm{i}\right)}}$$

(2)

$$\mathrm{f}\left(\mathrm{x}\right)=\left\{\begin{array}{c}1, 450 \mathrm{l}\mathrm{x}\le {\mathrm{I}}_{\mathrm{i}}\le 2000 \mathrm{l}\mathrm{x}\hfill \\ 0, {\mathrm{I}}_{\mathrm{i}}<450 \mathrm{l}\mathrm{x}\cup {\mathrm{I}}_{\mathrm{i}}>2000 \mathrm{l}\mathrm{x}\hfill \end{array}\right.$$

where I_i represents the illuminance at the i-th indoor sensor point (lx); t(i) denotes the time step (h); and α(i) is a weighting coefficient whose value is determined by I_i.

DGP is a primary metric used to quantify the probability of glare occurrence within the visual field. Its output is expressed as a percentage, which aligns well with perceived visual experience and provides an intuitive representation of glare severity. The calculation equation is given as follows:

$$\mathrm{DGP} = 5.87 \times {10}^{-5}\cdot {\mathrm{E}}_{\mathrm{V}} + 9.18 \times {10}^{-2}\cdot \log \left(1+\sum _{\mathrm{i}}{\displaystyle \frac{{\mathrm{I}}_{\mathrm{s},\mathrm{i}}\cdot {\mathrm{O}}_{\mathrm{s},\mathrm{i}}}{{\mathrm{E}}_{\mathrm{V}}^{1.87}\cdot {\mathrm{P}}_{\mathrm{i}}^2}}\right) + 0.16$$

(3)

where Ev denotes the vertical eye illuminance produced by the light source (lx); I_s,i represents the luminance of the light source (cd/m²); O_s,i is the solid angle subtended by the light source within the observer’s field of view (rad); and Pi is the position index, which reflects the angular displacement (azimuth and elevation) between the observer’s line of sight and the glare source.

2.

Energy Performance Metric

Heating, cooling, and lighting energy consumption constitute the major components of building energy consumption, accounting for approximately 50–70% of the building energy consumption [44]. In severe cold regions, heating energy consumption accounts for the most significant portion of building energy use [45]. For example, in Daqing, the official heating season generally extends from October 20 to April 20 of the following year, lasting approximately six months. In addition, a previous study on university teaching-office buildings reported that heating energy consumption in Harbin accounts for 94.6% [46] of the total cooling and heating energy consumption, further indicating that building operation in severe cold regions is dominated by heating energy consumption. Therefore, cooling and lighting energy consumption were not considered in this study, and heating energy intensity (kWh/m²), denoted as Eh, was adopted as the primary metric for evaluating energy performance.

To ensure the accurate simulation of performance metrics, this study uses the Rhino–Grasshopper parametric platform, which integrates three core modules: geometric modeling, performance simulation, and automated optimization [47]. The platform incorporates tools such as Ladybug and Honeybee to support simulation workflows. Among the metrics, sDA, UDI, and DGP are simulated using the Radiance engine, while E_h is computed using the EnergyPlus engine, providing reliable simulation outputs for subsequent analysis.

2.2.2. Parameter Settings for the Building Envelope, Occupants, and Equipment

The building envelope has a significant influence on both daylighting performance and heating energy consumption of classrooms. In accordance with relevant building codes and standards [48,49], the material and thermal parameters of the envelope components are specified as follows (

Table 1. Material parameter settings of the building envelope.

	Enclosure Structure	Set Value
U-value W/(m2·K)	Exterior Wall	0.2
	Floor	0.2
	Roof	0.2
	Interior Wall	0.5
Material Reflectance	Ceiling	0.75
	Interior Walls	0.75
	Floor	0.50

):

Occupancy density and equipment-related parameters have a significant impact on daylighting performance and heating energy consumption. According to relevant literature [50] and guidelines [48,51,52], the settings for occupant activities and equipment loads are summarized in

Table 2. Parameter settings for occupants and equipment.

Parameter	Setting
Occupant density (m2/person)	1.5
Room summer setpoint temperature (°C)	26
Room winter setpoint temperature (°C)	18
Fresh air volume [m3/(h·person)]	30
COP	2.3
Test grid size (m)	2 m × 2 m
Illuminance sensor height (m)	0.75
Intensity of infiltration	0.0003 m3/(s·m2)

.

2.3. Multi-Objective Optimization Research

This study defines four optimization objectives, aiming to identify high-performance design schemes by minimizing or maximizing the corresponding objective functions. The mathematical expressions of the four objective functions are given as follows:

$$\left\{\begin{array}{c} \mathrm{F}\left(\mathrm{sDA}\right)=\left\{\mathrm{f}\left({\mathrm{x}}_1\right),\mathrm{f}\left({\mathrm{x}}_2\right),\mathrm{f}\left({\mathrm{x}}_3\right),\dots ,\mathrm{f}\left({\mathrm{x}}_{2500}\right)\right\}\\ \mathrm{F}\left(\mathrm{UDI}\right)=\left\{\mathrm{f}\left({\mathrm{x}}_1\right),\mathrm{f}\left({\mathrm{x}}_2\right),\mathrm{f}\left({\mathrm{x}}_3\right),\dots ,\mathrm{f}\left({\mathrm{x}}_{2500}\right)\right\}\\ \mathrm{F}\left(\mathrm{DGP}\right)=\left\{\mathrm{f}\left({\mathrm{x}}_1\right),\mathrm{f}\left({\mathrm{x}}_2\right),\mathrm{f}\left({\mathrm{x}}_3\right),\dots ,\mathrm{f}\left({\mathrm{x}}_{2500}\right)\right\}\\ \mathrm{F}\left({\mathrm{E}}_{\mathrm{h}}\right)=\left\{\mathrm{f}\left({\mathrm{x}}_1\right),\mathrm{f}\left({\mathrm{x}}_2\right),\mathrm{f}\left({\mathrm{x}}_3\right),\dots ,\mathrm{f}\left({\mathrm{x}}_{2500}\right)\right\}\\ \mathrm{F}{\left(\mathrm{x}\right)}_{\min }=\left\{{\displaystyle \frac{1}{\mathrm{F}\left(\mathrm{sDA}\right)}}+{\displaystyle \frac{1}{\mathrm{F}\left(\mathrm{UDI}\right)}}+\mathrm{F}\left(\mathrm{DGP}\right)+\mathrm{F}\left({\mathrm{E}}_{\mathrm{h}}\right)\right\}\\ \end{array}\right.$$

(4)

This study conducts multi-objective optimization on the Grasshopper platform using the Wallacei plugin, which implements the NSGA-II algorithm. The genetic algorithm parameters, including population size, number of generations, crossover probability, mutation probability, and random seed, are listed in

Table 3. Parameter settings of the genetic algorithm.

Generation Size	Generation Count	Crossover Probability	Mutation Probability	Random Seed
50	50	0.9	0.05	1

. After parameter configuration, Wallacei was run to obtain the Pareto-optimal schemes, together with their geometric configurations, optimization parameters, and performance metrics. All optimization runs were carried out on a workstation equipped with 32 GB of RAM and an NVIDIA GeForce RTX 4060 Ti, NVIDIA Corporation, Santa Clara, CA, USA.

2.4. Classification Predictive Model

2.4.1. Dataset Construction

Binary labels were assigned to the 2500 samples generated through multi-objective optimization based on their membership in the Pareto-optimal set. These labeled samples were then used to develop and evaluate the classification model. To ensure reliable model training and performance assessment, the dataset was divided into a training set and a test set at a ratio of 8:2. The training set, comprising 2000 samples (80%), was used for model training, while the test set, containing 500 samples (20%), was used to evaluate the model’s predictive performance and generalization ability on unseen data.

2.4.2. Model Training

Figure 2 illustrates the working principles of Linear Regression (LR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Multilayer Perceptron (MLP). The parameter settings for these algorithms were based on relevant studies and further refined through iterative cross-validation and hyperparameter tuning.

Linear Regression (LR): A classical statistical learning method that estimates the target variable as a linear combination of input features. It is suitable for datasets exhibiting linear or approximately linear separability [53].

Random Forest (RF): An ensemble learning method based on bootstrap resampling. Multiple subsets of the original dataset are sampled to construct decision trees, and the final prediction is obtained by aggregating the outputs of all trees through a voting mechanism (for classification) or averaging (for regression) [53].

Extreme Gradient Boosting (XGBoost): An efficient gradient boosting framework that iteratively builds weak learners (decision trees) and integrates their predictions to enhance model performance. Its parallelized architecture substantially improves training efficiency, making it well-suited for large-scale datasets [53].

Light Gradient Boosting Machine (LightGBM): A high-performance gradient boosting algorithm that uses histogram-based feature binning and a leaf-wise tree growth strategy. LightGBM reduces computational cost and accelerates training speed while maintaining strong predictive accuracy [54].

Multilayer Perceptron (MLP): A typical feedforward neural network composed of an input layer, one or more hidden layers, and an output layer. The MLP maps multi-dimensional input features through nonlinear activation functions and updates its weights using the backpropagation algorithm, enabling the effective modeling of complex nonlinear relationships [55].

2.4.3. Model Evaluation

To evaluate the performance of the classification model, this study employs precision, recall, and F1-score as evaluation metrics. Precision reflects the reliability of positive predictions, with higher values indicating fewer false positives. Recall measures the model’s ability to correctly identify positive samples, and higher recall indicates fewer false negatives [56]. F1-score, defined as the harmonic mean of precision and recall, provides a balanced evaluation when both false positives and false negatives are important [57]. These three metrics capture different aspects of classification performance and offer a robust assessment of the model. Their mathematical definitions are given as follows:

$$\mathrm{Precision} = {\displaystyle \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP}}}$$

(5)

$$\mathrm{Recall}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}$$

(6)

$$\text{F1-score}=2 \times {\displaystyle \frac{\mathrm{Precision} \times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}$$

(7)

In these equations, TP (True Positive) denotes the number of positive samples correctly classified by the model; FP (False Positive) denotes the number of negative samples incorrectly classified as positive; and FN (False Negative) denotes the number of positive samples incorrectly classified as negative.

3. Measurement and Validation

3.1. Case Background and Research Object

From October 2024 to June 2025, field investigations were conducted in Daqing and Harbin, two representative cities in China’s severe cold region. A total of 50 non-lecture classroom units across 10 university campuses were surveyed to obtain geometric and dimensional data, providing the basis for constructing the subsequent parametric baseline model. To ensure sample comparability, the surveyed classrooms were selected from conventional teaching buildings, while stepped lecture halls, laboratories, and other specialized spaces were excluded. During the field survey, on-site observation, dimensional measurement, and photographic documentation were conducted, and the key geometric information of each classroom was systematically recorded. The collected data included nine morphological parameters, such as classroom orientation, width, depth, floor height, total window width, and window-sill height. These data were then summarized and statistically analyzed to identify the typical spatial characteristics of university classrooms in severe cold regions and to determine the parameter ranges for subsequent parametric modeling. The survey process is illustrated in Figure 3.

The field survey results indicate that medium-sized classrooms (61–100 occupants) account for the largest share, at 56%, followed by small classrooms (25–60 occupants) at 28%, while large classrooms (101–150 occupants) account for only 16%. Based on this distribution and considering that medium-sized classrooms are the most representative type in the surveyed educational buildings, the subsequent modeling and analysis focus primarily on medium-sized classroom units.

3.2. Field Measurements of the Daylighting and Thermal Environment in a Medium-Sized Classroom

From 6 April to 6 June 2025, field measurements were conducted in a medium-sized classroom at Northeast Petroleum University to assess its daylight–thermal performance. The technical specifications of the measuring instruments are summarized in

Table 4. Technical specifications of the measuring instruments.

Illuminance meter (TA8123)	Measurement range	200,000 lx
	Resolution	0.1 lx
	Accuracy	±3%
Temperature and humidity meter (TA622B)	Temperature range	−10 °C to 50 °C
	Temperature accuracy	±1.5 °C
	Temperature resolution	0.1 °C
	Humidity range	5.0%RH to 98%RH
	Humidity accuracy	±4%RH (41–80%RH)
	Humidity resolution	0.1%RH
Laser distance meter (SW-DA100)	Distance measurement accuracy	1.5 mm

. To ensure measurement reliability, the selected classroom was located on the fourth floor of a south-facing building, minimizing the influence of shading from surrounding buildings and vegetation. The classroom measures 9.3 × 11.4 m in plan and 3.9 m in height. With a capacity of 88 occupants, it is highly representative of medium-sized classrooms in university buildings located in severe cold regions and was therefore selected as the reference classroom for subsequent baseline model development.

Daylighting measurements were conducted on 19 April 2025, between 9:00 and 18:00. Indoor and outdoor illuminance were recorded simultaneously at 30 min intervals. The measurement points were arranged at a height of 0.75 m above the floor. A rectangular grid with 2 m spacing was applied to the primary occupied area, with the center of each grid cell defined as an illuminance measurement point [58].

Thermal environment measurements were conducted on 1 June 2025, between 9:00 and 18:00. Indoor and outdoor air temperatures were recorded simultaneously at 1 h intervals. The measurement point was positioned at the midpoint of the classroom’s diagonal axis. The detailed measurement layout is presented in Figure 4.

3.3. Establishment of the Prototype Model

In accordance with China’s building energy-saving design standards [52] and considering the functional requirements of university classrooms, this study summarizes the geometric characteristics of medium-sized classrooms obtained from the field survey. To ensure that the selected variables and their ranges are representative rather than arbitrarily assigned, the surveyed morphological parameters were further subjected to statistical analysis to identify their distribution characteristics and representative value ranges (Figure 5). A set of design parameters was defined for constructing the prototype model (

Table 5. Design parameters for the medium-sized classroom.

Design Parameter	Range (m)	Step (m)
Bay (B1)	8.0–14.5	0.5
Depth (D1)	6.5–12.0	0.5
Height (H1)	3.6–5.4	0.3
Orientation (θ)	0–180°	45°
Total window width (B2)	4.5–9.0	0.5
Number of windows (n)	3–5 window	1
Window-sill height (H2)	0.9–1.2	0.1
Window height (H3)	1.8–2.7	0.3
Glazing setback from wall (D2)	0.40–0.60	0.05
Side-wall distance to the window edge (B3)	0.9–1.8	0.3
Glazing option (x)	0–3 types	1

). A total of 11 parameters were selected: bay width (B₁), room depth (D₁), floor height (H₁), building orientation (θ), total window width (B₂), number of window modules (n), window-sill height (H₂), window height (H₃), setback distance of glazing from the interior wall surface (D₂), side-wall distance to the window edge (B₃), and glazing option (x), which includes four types of glass: clear glass, heat-absorbing glass, Low-E glass, and heat-reflective glass (

Table 6. Layer configurations and key thermal–optical properties of different glazing option.

Glazing Option	Construction	Visible Transmittance	SHGC	U-Value [W/(m2·K)]
Clear glass	6 mm clear glass + 12 mm air layer + 6 mm clear glass	0.81	0.75	2.59
Heat-absorbing glass	6 mm heat-absorbing glass + 12 mm air layer + 6 mm clear glass	0.68	0.49	2.60
Low-E glass	6 mm Low-E glass + 12 mm air gap + 6 mm clear glass	0.68	0.46	1.72
Heat-reflective glass	6 mm heat-reflective glass + 12 mm air gap + 6 mm clear glass	0.43	0.42	2.45

). Among them, D₂ was included to represent the recessed window configuration commonly found in teaching buildings in severe cold regions, which can partially capture the passive self-shading effect associated with thick exterior walls. By adjusting the values of these parameters, the overall classroom geometry can be systematically controlled.

3.4. Verification of Simulation Accuracy

Before constructing the multi-objective optimization platform, the accuracy of the simulation tools was first validated. A parametric model of the measured classroom was developed, using on-site measurements of illuminance, temperature, and humidity as reference data. Radiance and Honeybee were then employed to simulate daylighting and thermal performance under the same conditions and time period as the field measurements. Finally, the web-based SPSSPRO platform was used to perform correlation analysis and significance testing between the simulated and measured values, thereby quantifying the accuracy of the simulation results.

Daylighting simulation: The simulation period was set from 9:00 a.m. to 6:00 p.m. on 19 April 2025. A CIE standard overcast sky model consistent with the weather conditions during the field test was adopted. To match the measurement layout, indoor illuminance points were arranged using a 2 m × 2 m grid with a measurement height of 0.75 m, resulting in a total of 20 measurement points. After linking sky conditions, material reflectance parameters, and measurement point coordinates, the Honeybee module was executed to generate illuminance results, which were then fed back to the Grasshopper platform. The analysis shows that the Mean Absolute Percentage Error (MAPE) of illuminance was 8.4%. The temporal fitting between simulation and measurement was strong, with a coefficient of determination R² of 0.997. Correlation analysis further confirmed a significant relationship between the two datasets.

Thermal environment simulation: The simulation period was set from 9:00 to 18:00 on June 1. Similar to the field test, thermal simulations were conducted at 1 h intervals, with the simulation point located at the midpoint of the classroom’s diagonal axis. The simulation results for temperature and humidity showed MAPE values of approximately 0.89% and 1.01%, and R² values exceeding 0.95, demonstrating a strong correlation and high consistency with the measured data (Figure 6 and

Table 7. R² and MAPE values for illuminance, temperature, and humidity.

	R2	MAPE
Illuminance	0.997	8.4%
Temperature	0.979	0.89%
Humidity	0.983	1.01%

).

Overall, these validation results confirm that Honeybee and Radiance provide reliable and accurate simulation outputs for the parametric classroom model, supporting their use as the computational foundation for the multi-objective optimization platform.

4. Result

4.1. Analysis of Multi-Objective Optimization Results

4.1.1. Analysis of the Pareto Historical Scheme Set

As shown in Figure 7, the Pareto front illustrates the trade-offs among the four performance indicators under different design variables. Following relevant literature [30,32], the optimization targets were defined as follows: sDA within 0.5–1, UDI within 50–100%, DGP within 0–35%, and E_h within 30–40 kWh/m².

Analysis of the Pareto scheme set for the medium-sized classroom model indicates that the schemes are more densely distributed near the X–Z plane. This distribution pattern suggests that most design configurations are capable of maintaining relatively high UDI values while simultaneously achieving lower E_h values. However, the schemes exhibit substantial dispersion in DGP performance. Specifically, DGP values range from 33.70% to 49.77%, reflecting instability in the optimization outcomes for this indicator, with most schemes failing to meet the predefined target range. This relatively limited improvement in DGP is mainly due to the trade-off between glare control and the other optimization objectives, as well as the limited exploration of more direct and commonly used glare-control strategies within the current parameter scope. In contrast, the optimization performance for sDA is notably superior to that of the other metrics, with 87% of the samples reaching the target value of 1, indicating a high level of daylight utilization. Overall, these findings underscore the importance of prioritizing the control of DGP in the design optimization of medium-sized classrooms.

4.1.2. Optimal Ranges of Design Parameters

After completing the performance optimization, this study extracted the design parameters from the top 312 Pareto-optimal schemes to derive the optimal parameter ranges for classroom design, as shown in Figure 8. For the spatial configuration, the optimal ranges are identified as B₁ = 12.0–12.5 m, D₁ = 6.5–7.0 m, and H₁ = 3.6 m, while θ, aligned with the south–north axis is found to be the most favorable orientation. For window design, the optimal ranges are as follows: B₂ = 8.0–8.5 m, H₂ = 1.2 m, and D₂ = 0.55–0.60 m. The number of window modules is preferably set to 4 or 5. The glazing option is recommended as heat-absorbing glass or Low-E glass, and a B₃ value of 1.8 m is suggested.

4.1.3. Performance Improvement and Comparison with the Baseline Model

To evaluate the optimization effect, a baseline model was established based on field measurements of a medium-sized classroom. The model preserved the envelope and window parameters obtained from the field survey. To reflect the actual layout of the teaching building, it was mirrored about the corridor centerline, with the 2 m corridor width measured on site treated as a fixed geometric condition.

After identifying the optimal design parameter ranges, this study further compared the optimized schemes with the baseline model to assess the performance improvements. Figure 9 shows the performance distribution characteristics of the optimized medium-sized classroom dataset. Compared with the baseline model, the optimized schemes demonstrate significant improvements in all performance metrics: average sDA and UDI increased, while average DGP and E_h decreased. Specifically, sDA ranges from 0.5 to 1.0, with an average of 0.97, representing an increase of 0.12 compared with the baseline model. Notably, a large number of schemes reached the maximum value of 1.0, and none fell below 0.5, indicating sufficient indoor daylight utilization. UDI values range from 40% to 72%, with an average of 66.65%, which is a 17.9% increase compared with the baseline model. Approximately 94% of the values fall within the range of 54–74%, meeting the design requirements. DGP values range from 32% to 48%, with an average of 38.34%, showing a 27.5% decrease from the baseline model. Although most schemes fall between 34% and 40%, indicating a substantial reduction in glare probability, the predefined target range was not fully achieved. E_h values range from 29.60 to 40 kWh/m², with an average of 35.92 kWh/m², showing a 13.5% reduction compared with the baseline model. Most values are concentrated between 34 and 38 kWh/m², with a relatively even distribution. Overall, the optimization achieved significant improvements in sDA, UDI, and E_h, whereas DGP remained the most challenging metric to control.

4.1.4. Comparison Between Extreme Schemes and the Overall Optimal Scheme

Based on the overall optimization results, four extreme schemes were selected to represent the best performance in individual metrics, namely the schemes with the maximum sDA, maximum UDI, minimum DGP, and minimum E_h. These schemes were then compared with the overall optimal scheme. The calculation method for the overall optimal scheme is given in Equation (8). Their performance metrics and design parameters, together with those of the baseline model, are summarized in

Table 8. Objective values for the classroom Pareto front schemes.

	sDA-Max Scheme	UDI-Max Scheme	DGP-Min Scheme	Eh-Min Scheme	Overall Optimal Scheme	Baseline Model
B1 (m)	13.0	12.5	12.0	13.0	13.0	11.0
D1 (m)	6.0	6.0	12.0	12.0	7.0	9.5
H1 (m)	3.6	3.6	3.6	3.6	3.6	3.9
θ (°)	0	0	0	0	0	0
B2 (m)	8.0	8.5	7.5	8.0	8.0	9.0
n (−)	5	5	5	3	5	4
H2 (m)	1.2	1.2	1.1	1.2	1.2	1.1
H3 (m)	1.8	2.1	1.8	1.8	2.1	2.4
D2 (m)	0.5	0.6	0.6	0.4	0.5	0.5
B3 (m)	1.8	1.8	1.2	1.8	1.8	1.8
X (−)	2	1	2	0	2	1
sDA (−)	1	1	0.48	0.73	1	0.85
UDI (%)	71.52	74.04	33.89	47.34	67.17	56.50
DGP (%)	37.54	39.52	33.65	48.45	36.54	52.89
Eh (kWh/m2)	36.45	37.61	31.88	29.60	35.33	41.53

.

As shown in Figure 10, the six design schemes exhibit clear geometric differences in classroom spatial proportions and window-opening layout. Specifically, the DGP-min scheme and Eh-min scheme are characterized by greater classroom depth, whereas the sDA-max scheme, UDI-max scheme, and the overall optimal scheme exhibit relatively shallower spatial forms. These geometric differences are consistent with the parameter variations presented in Table 8 and further illustrate the trade-off relationship among daylight availability, glare control, and heating energy consumption.

$$\left\{\begin{array}{c}\mathrm{S}={\displaystyle \frac{100}{{\mathrm{sDA}}_{\max }-{\mathrm{sDA}}_{\min }}}\\ \mathrm{U}={\displaystyle \frac{100}{{\mathrm{UDI}}_{\max }-{\mathrm{UDI}}_{\min }}}\\ \mathrm{D}={\displaystyle \frac{100}{{\mathrm{DGP}}_{\max }-{\mathrm{DGP}}_{\min }}}\\ \mathrm{E}={\displaystyle \frac{100}{{\mathrm{E}}_{{\mathrm{h}}_{\max }}-{\mathrm{E}}_{{\mathrm{h}}_{\min }}}}\\ \mathrm{O}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{l}\mathrm{l} \mathrm{B}\mathrm{e}\mathrm{s}\mathrm{t}=\left({\mathrm{sDA}}_{\mathrm{i}}-{\mathrm{sDA}}_{\min }\right)\mathrm{S}+\left({\mathrm{UDI}}_{\mathrm{i}}-{\mathrm{UDI}}_{\min }\right)\mathrm{U}+\left({\mathrm{DGP}}_{\max }-{\mathrm{DGP}}_{\mathrm{i}}\right)\mathrm{D}+\left({\mathrm{E}}_{{\mathrm{h}}_{\max }}-{\mathrm{E}}_{{\mathrm{h}}_{\mathrm{i}}}\right)\mathrm{E}\end{array}\right.$$

(8)

Quantitative comparison shows that the optimization produced clear improvements across all four metrics. Compared with the baseline scheme, each extreme scheme shows improvement in specific performance dimensions. Specifically, the scheme with the maximum sDA achieves a 15% increase in sDA, the scheme with the maximum UDI results in a 17.54% increase in UDI, the scheme with the minimum DGP reduces glare probability by 19.24%, and the scheme with the minimum E_h reduces heating energy consumption by 11.93 kWh/m². While the overall optimal scheme does not outperform the corresponding single-objective optimal schemes in the three individual performance metrics, it achieved a more balanced improvement across all four objectives relative to the baseline scheme. In particular, sDA and UDI increase by 0.15 and 10.67%, respectively, while DGP and E_h decrease by 16.35% and 6.20 kWh/m².

Analysis of the extreme scheme design parameters reveals the inherent trade-offs between performance objectives. In the case of DGP and E_h, despite achieving extreme values of 33.65% and 29.60 kWh/m² (a 36.4% and 28.7% decrease compared with the baseline model values of 52.89% and 41.53 kWh/m²), this performance improvement comes at the cost of sDA and UDI. This trend suggests that stricter control of glare and energy consumption may require some reduction in daylight availability. A significant feature of the optimal schemes for DGP and E_h is the use of larger B₁, D₁, and B₂, combined with smaller H₁ and H₃. Conversely, for sDA and UDI, the optimal design schemes feature larger B₁ and B₂, but smaller D₁, H₁, and H₃, reflecting that D₁ may be a key factor affecting classroom performance, though further verification is needed. Notably, the design parameters for the scheme with the maximum sDA and the overall optimal scheme are highly similar.

4.2. Analysis of the Results of the Performance Classification Model for Medium-Sized Classrooms

4.2.1. Model Accuracy Evaluation

To improve prediction accuracy and provide efficient support for design decision-making and scheme evaluation, this study compared multiple classification models to identify the most suitable surrogate model. Five machine learning algorithms were adopted for model development and comparison. The dataset obtained from the multi-objective optimization process comprised 2500 samples, including 312 Pareto-optimal schemes and 2188 non-Pareto-optimal schemes. This indicates a clear class imbalance, with positive samples accounting for only 12.48% of the total dataset. The performance of each classification model on the test set is summarized in

Table 9. Comparison of five models for the medium-sized classroom.

Evaluation Indicator	Medium-Sized Classroom (Train)			Medium-Sized Classroom (Test)
	Precision	Recall	F1-Score	Precision	Recall	F1-Score
LR	0.95	0.86	0.87	0.95	0.76	0.82
XGBoost	0.98	0.93	0.95	0.95	0.85	0.89
RF	0.97	0.89	0.93	0.96	0.85	0.89
LightGBM	0.98	0.94	0.95	0.95	0.84	0.88
MLP	0.95	0.95	0.95	0.95	0.95	0.95

.

As shown in Table 9, the MLP model exhibits more balanced overall performance than the ensemble tree models, including XGBoost, RF, and LightGBM. The training and test loss curves of the MLP model decrease rapidly in the early stage and gradually stabilize as the number of iterations increases, indicating good convergence during training. Meanwhile, the training and test accuracy curves rise steadily and remain at a high level after approximately 100 iterations, with only slight fluctuations thereafter. The overall trends of the training and test curves are closely aligned, suggesting that the model maintains stable predictive performance without obvious overfitting, as illustrated in Figure 11. Specifically, the MLP model maintains stable accuracy, recall, and F1-score exceeding 0.95 on both the training and testing sets, with no significant fluctuations among the metrics. This indicates not only reliable overall classification performance but also a strong capability in identifying positive samples, thereby satisfying the study’s requirement for balanced model performance. In contrast, although the ensemble tree models achieve test accuracy values of approximately 0.95, their recall and F1-score are markedly lower, ranging from 0.84 to 0.91, indicating that high overall accuracy is accompanied by relatively weak identification of positive samples. Based on the comprehensive evaluation of accuracy, recall, and F1-score, the MLP algorithm was ultimately selected as the surrogate model for subsequent classification analysis.

4.2.2. Model Interpretability Analysis

To further explore the complex nonlinear relationships between the 11 influencing factors and the target variable, while enhancing model interpretability [59,60], this study introduces SHAP (SHapley Additive exPlanations) for interpretability analysis. The LR model was not included in Figure 12 because its predictive performance was inferior to that of the other models, while its linear structure offers stronger intrinsic interpretability. Accordingly, the SHAP analysis mainly focuses on the MLP and other representative nonlinear models. The SHAP value distributions for each feature under different models are illustrated in Figure 12. Each point in the figure represents the SHAP value of a specific feature for an individual sample in the test set, with high feature values shown in red and low feature values shown in blue. The features are ranked according to their mean absolute SHAP values, which reflect their global importance. Features ranked higher are interpreted as having a stronger influence on the target variable, namely whether a scheme belongs to the optimal class. It should be noted that feature importance rankings vary across different machine learning models. This observation is consistent with Li et al. [60], who concluded that the assessment of feature importance may differ depending on the machine learning model used.

Despite the differences among models, D₁, H₁, and glazing option were consistently identified as key factors affecting whether a scheme belonged to the Pareto-optimal class, with their global importance rankings remaining within the top five across all tested models. In addition, although H₃ and B₂ did not consistently rank among the most important features across all models, they exhibited relatively high contributions and strong explanatory power in several models. As such, these parameters can be regarded as secondary yet influential factors affecting overall classroom performance.

Further analysis indicates a high degree of consistency among the four models in their interpretation of feature impacts. Specifically, for D₁, H₁, and glazing option, lower feature values are consistently associated with positive SHAP values, whereas higher values correspond to negative SHAP values. In contrast, for H₃ and B₂, higher feature values are linked to positive SHAP values, while lower values are associated with negative SHAP values. Accordingly, potential strategies for achieving Pareto-optimal schemes include reducing D₁ and H₁, selecting heat-absorbing glass or Low-E glass for glazing option, and increasing both H₃ and B₂. These findings are consistent with the earlier conclusion that D₁ plays a critical role in determining classroom performance.

Conversely, θ and H₂ were consistently ranked as the least influential features across all four models, indicating that under the specific environmental conditions of Daqing, these two variables exert a limited influence on whether a classroom design can achieve a Pareto-optimal scheme.

5. Discussion

Unlike conventional studies that mainly rely on regression models to predict single performance metrics, this study proposes an MLP-based classification model built on multi-objective optimization results. This model can directly evaluate design schemes, thereby improving decision-making efficiency. By transforming complex performance evaluation into a rapid classification process, this approach enables designers to quickly select high-performance schemes during the early design phase, thereby meeting the needs of rapid architectural design iterations.

5.1. Multi-Objective Optimization Results

This study used the NSGA-II algorithm to perform integrated daylight–thermal optimization for university classrooms in severe cold regions. The optimization framework demonstrated a strong ability to balance multiple performance objectives. Compared with the baseline model, the overall optimal scheme achieved balanced improvements across all four key metrics: DGP and E_h decreased by 16.35% and 6.20 kWh/m², respectively, while sDA and UDI remained at high levels of 100% and 67.1%. These results indicate that the proposed framework can effectively mitigate the trade-off between daylighting performance and heating energy consumption in severe cold regions, achieving lower heating demand while maintaining satisfactory daylighting performance.

5.2. MLP Model Prediction Results

Among the five classification algorithms, the MLP was ultimately selected as the final classification model. The MLP model achieved stable accuracy, recall, and F1-score exceeding 0.95 on the independent test set, demonstrating its high predictive accuracy and generalization ability. In contrast, although the tree ensemble models showed higher precision on the training set, their recall and F1-score were significantly lower than those of the MLP. This discrepancy is attributed to the class imbalance in the dataset and the nonlinear nature of the classification task. Under such conditions, tree ensemble models may be more prone to bias toward the majority negative class, leading to weaker identification of critical positive samples. By contrast, the MLP showed greater robustness in this low-dimensional nonlinear classification task. This finding is consistent with Kristiansen et al. [61], who reported high efficiency and stability when using ANNs to predict building performance. Overall, the results suggest that MLP is a more suitable surrogate model than gradient boosting tree models for this type of building design classification task.

5.3. SHAP Analysis

This study employed SHAP for interpretability analysis, thereby improving the transparency of the MLP model and quantifying the global contributions of the 11 design parameters to whether a scheme belongs to the Pareto-optimal class. In terms of global key features and optimization directions, the SHAP global importance analysis confirmed that D₁, H₁ and glazing option were the top five key parameters shared across models. At the same time, θ and H₂ were ranked as the least influential features, indicating that they provide greater design flexibility in the context of classroom design in severe cold regions. In addition, although H₃ and B₂ were not consistently ranked among the top features across all models, they still contributed strongly in several models and can therefore be regarded as secondary but influential factors. Combining the optimal design parameter ranges from Section 4.1.2, the refined optimization strategy suggests that, to enhance classroom performance in severe cold regions, D₁ should be controlled within the range of 6.5–7.0 m, H₁ should be set to 3.6 m, glazing option should preferably be heat-absorbing glass or Low-E glass, and H₃ should be controlled within the range of 1.8–2.1 m and B₂ should be controlled within the range of 8.0–8.5 m.

6. Conclusions

This study aims to address the trade-off between heating energy consumption and daylighting in the design of university classrooms in severe cold regions of China. To this end, a classification framework integrating NSGA-II-based multi-objective optimization and an MLP classification model was developed, enabling a complete workflow from complex performance simulation to rapid design decision-making.

6.1. Key Findings and Contributions

Innovation in the technical approach: This study established a “simulation–optimization–prediction” framework, providing a fast and efficient decision-support method for the early design stage in severe cold regions. This framework can further quantify feasible parameter ranges under coupled performance constraints, providing more practical support for early-stage design decision-making. This approach significantly improves the efficiency and reliability of performance-driven building design.

Multi-objective optimization of classroom daylighting and heating in severe cold regions: Using the NSGA-II algorithm, the trade-offs between design parameters and daylight–thermal performance metrics were quantified, and a Pareto-optimal scheme set was generated. Compared with the baseline model, the overall optimal scheme achieved balanced improvements across all four key metrics: sDA increased by 0.15, UDI by 10.67%, while DGP and E_h decreased by 16.35% and 6.20 kWh/m², respectively. More importantly, the optimization results provided directly applicable parameter suggestions for design practice, rather than only confirming general intuitive judgments.

Development of a high-performance MLP classification model: After comparison of five classification algorithms, the MLP model achieved accuracy, recall, and F1-score values exceeding 0.95 on the test set, demonstrating more balanced performance than the tree ensemble models. It also showed strong robustness under class-imbalanced conditions and in this low-dimensional nonlinear classification task. Unlike conventional regression-based approaches, the proposed classification model can directly identify high-performance schemes and support rapid scheme screening in the early design stage.

SHAP analysis: This study identified depth, height, and glazing option as the most influential factors affecting classroom performance, while H₃ and B₂ were recognized as secondary but still important variables. In contrast, orientation and window-sill height showed relatively limited influence, indicating greater flexibility in design adjustment. These results suggest that the proposed framework can further distinguish relatively flexible variables from those that are secondary but still important under multi-objective conditions. Based on these findings, the refined optimization strategy suggests that depth should be controlled within 6.5–7.0 m, height should be set at 3.6 m, glazing option should preferably adopt heat-absorbing glass or Low-E glass, and window height and total window width should be controlled within 1.8–2.1 m and 8–8.5 m, respectively.

6.2. Limitations and Future Work

Although the proposed framework proved effective, several limitations remain. First, the empirical analysis focused primarily on medium-sized classrooms in Daqing, a representative city in China’s severe cold region. Future research should extend the framework to other climatic regions and classroom types, such as small and large classrooms, to enhance its generalizability and adaptability. Second, the optimization objectives did not include a detailed assessment of thermal comfort. Future studies should incorporate dynamic thermal comfort indicators, such as PMV and PPD, to support a more comprehensive integrated optimization of thermal comfort and energy performance. In addition, the current parameter scope did not include more direct shading-related design variables, such as blinds or overhangs, which may have limited the optimization effect on DGP. Future studies should further incorporate shading-related variables to improve the analysis of glare control. Finally, the cross-domain generalization ability of the MLP model is constrained by the distribution of the training data. Future work may explore ensemble learning methods, such as stacking and boosting, as well as transfer learning approaches, to further improve model generalization and practical applicability.