Integrating BIM, Machine Learning, and PMBOK for Green Project Management in Saudi Arabia: A Framework for Energy Efficiency and Environmental Impact Reduction

Abstract

This study introduces a comprehensive framework combining building information modeling (BIM), project management body of knowledge (PMBOK), and machine learning (ML) to optimize energy efficiency and reduce environmental impacts in Riyadh’s construction sector. The suggested methodology utilizes BIM for dynamic energy simulations and design visualization, PMBOK for integrating sustainability into project-management processes, and ML for predictive modeling and real-time energy optimization. Implementing an integrated model that incorporates building-management strategies and machine learning for both commercial and residential structures can offer stakeholders a thorough solution for forecasting energy performance and environmental impact. This is particularly essential in arid climates owing to specific conditions and environmental limitations. Using a simulation-based methodology, the framework was evaluated based on two representative case studies: (i) a commercial complex and (ii) a residential building. The neural network (NN), reinforcement learning (RL), and decision tree (DT) were implemented to assess performance in energy prediction and optimization. Results demonstrated notable seasonal energy savings, particularly in spring (15% reduction for commercial buildings) and fall (13% reduction for residential buildings), driven by optimized heating, ventilation, and air conditioning (HVAC) systems, insulation strategies, and window configurations. ML models successfully predicted energy consumption and greenhouse gas (GHG) emissions, enabling targeted mitigation strategies. GHG emissions were reduced by up to 25% in commercial and 20% in residential settings. Among the models, NN achieved the highest predictive accuracy (R² = 0.95), while RL proved effective in adaptive operational control. This study highlights the synergistic potential of BIM, PMBOK, and ML in advancing green project management and sustainable construction.

1. Introduction

The increasing urgency for sustainable construction drives the adoption of more effective and environmentally conscious project-management practices [1,2,3]. The construction sector is a primary contributor to carbon emissions, resource depletion, and environmental degradation; therefore, it is imperative to enact rules that mitigate its ecological impact [4,5,6]. Emerging as a vital discipline in response to these challenges, green project management (GPM) links sustainability principles into project design, implementation, and lifecycle management [7,8]. However, conventional project-management methodologies may be deficient in the necessary tools for proactive decision-making, predictive energy analysis, and real-time environmental monitoring. This gap requires the incorporation of machine learning (ML), the project management body of knowledge (PMBOK), and building information modeling (BIM) to deliver a more holistic and data-driven methodology for sustainable construction [9,10].

Enhancing energy efficiency, enabling precise simulations of building performance [11], conducting lifecycle assessments (LCAs), and the early detection of inefficiencies (for all of which, BIM has demonstrated to be an excellent tool) are all contingent upon it [12,13]. BIM lacks a systematic project-management framework to ensure the coherent integration of sustainability principles, notwithstanding its facilitation of technical visualization and performance evaluation [14,15]. In contrast, PMBOK provides a clearly delineated array of project-management methodologies [16]; however, it inherently lacks the capacity for predictive analytics about sustainability or data-driven decision-making in general [17]. The advancement of ML techniques has created new potential for predicting environmental impacts, optimizing resources, and analyzing energy use patterns [18,19].

While BIM, ML, and PMBOK have individually contributed to improvements in energy performance, predictive analytics, and execution control, their integration remains limited and largely fragmented in the existing literature. Most previous studies have focused on isolated applications—such as BIM-based energy simulation, ML-driven energy demand forecasting, or project scheduling—without establishing an interoperable workflow that connects these components into a unified decision-making system. Moreover, few frameworks have addressed the contextualization of such integration in regions with extreme climatic conditions and distinct regulatory constraints. This study addresses this gap by proposing a novel, hybrid framework that combines BIM, PMBOK, and ML in a closed-loop structure, allowing for dynamic data exchange, automated energy performance optimization, and compliance with sustainability policies.

An integrated model that combines building-management techniques with machine learning, applicable to both commercial and residential buildings, offers stakeholders a thorough approach for forecasting energy performance and assessing the environmental impact of buildings. This is particularly essential in dry climates because of unique conditions and environmental limitations. Saudi Arabia is particularly important for the proposed model because of its stringent environmental policies, elevated energy usage, and extreme climatic circumstances, which necessitate innovative solutions. This study employs a simulation-based framework benefiting BIM, PMBOK, and ML to address this gap, enhancing energy efficiency and reducing environmental impacts in building projects. Meanwhile, in this study, the effectiveness of the proposed framework in reducing greenhouse gas (GHG) emissions was analyzed. It initially examines the potential simultaneous utilization of BIM and PMBOK to create a project-management system driven by sustainability. Secondly, it illustrates the significance of data-driven decision-making in construction by examining how ML algorithms can enhance energy efficiency and environmental impact predictions. Ultimately, it evaluates the efficacy and feasibility of the proposed framework through a validation approach.

2. Literature Review

2.1. Sustainable Construction and GPM

Sustainable building, accountable for a significant portion of global carbon emissions, resource depletion, and energy consumption, has emerged as a crucial remedy to the escalating environmental issues posed by the construction industry [20,21]. To mitigate environmental impacts and ensure enduring economic and social benefits, GPM combines sustainability principles into project planning, execution, and evaluation [22]. The adoption of GPM remains challenging because of insufficient knowledge among stakeholders, substantial initial costs, and the absence of standardization [23].

As part of its Vision 2030 sustainability initiatives, Saudi Arabia has established regulations and rules designed to promote eco-friendly construction and reduce carbon emissions. Compulsory standards for energy conservation, water efficiency, and sustainable materials in construction projects are delineated in the Saudi Green Building Code (SBC 1001–1006) and the Saudi Energy Efficiency Program. While these regulations provide a legal framework, enforcement and adherence differ among projects, with numerous developers prioritizing cost-efficiency over environmental considerations [24,25,26]. Highlighting the necessity for the enhanced integration of digital technologies to monitor sustainability compliance in real time, previous studies, including those by Madkhali et al. [27], have evaluated the efficacy of green construction policies in Saudi Arabia.

Although internationally recognized sustainability-assessment systems such as LEED, BREEAM, and SAP provide structured evaluation methodologies, their direct application in Saudi Arabia presents technical and regulatory limitations. For instance, LEED’s daylighting metrics or BREEAM’s seasonal energy factors are not optimized for extreme desert climates, where solar overexposure and dust accumulation affect both daylight quality. A location-specific sustainability model that integrates parameters aligned with Saudi building codes, such as minimum U-value thresholds (e.g., ≤0.57 W/m²·K for external walls in hot zones) and window-to-wall ratio (WWR) limits (≤25%), was developed based on recommendations from the Saudi Building Code (SBC-601 and SBC-602). However, the model is capable of updating local constraints for different regions and is not limited to Saudi Arabia.

2.2. BIM in Energy Efficiency and Sustainability

BIM, an acronym for the digital representation of a facility’s physical and functional characteristics, encompasses 3D models that include building geometry, spatial relationships, and construction elements, facilitating comprehensive simulations of energy consumption, structural integrity, and operational efficiency throughout a structure’s lifespan. BIM facilitates the real-time simulation of energy performance, lifecycle cost estimation, and material efficiency analysis, revolutionizing the design, analysis, and management of building projects [28,29]. Its ability to graph energy-consumption patterns and predict carbon emissions during the early design phase significantly enhances informed decision-making, marking a crucial contribution to sustainability [30]. Advanced energy modeling facilitated by tools like Autodesk Revit [22.0.2.392], EnergyPlus [9.6.0], and Green Building Studio optimizes heating, ventilation, and cooling (HVAC) systems, hence improving overall building efficiency [31].

Ma et al. [32] employed BIM-based LCA and energy modeling to assess the embodied carbon and environmental effects of high-rise buildings, revealing that concrete constituted 74% of the building’s lifecycle global warming potential and a significant portion of other environmental impacts. Likewise, Ibe [33] utilized BIM in LCA, illustrating that BIM-facilitated sustainability assessments contributed to a 25% reduction in material waste in infrastructure projects. Notwithstanding its benefits, BIM adoption encounters constraints including elevated implementation costs, a deficiency of experienced individuals, and interoperability challenges with conventional project management systems.

2.3. PMBOK and Sustainability

PMBOK constitutes a compilation of standardized rules and optimal practices for project management. It encompasses a thorough framework for overseeing project scope, schedule, budget, quality, human resources, communications, and risks. PMBOK is extensively utilized to guarantee that projects are finalized punctually, inside financial constraints, and adhere to the requisite quality benchmarks. PMBOK is a well-established framework that standardizes project management processes; nevertheless, its use in sustainable construction is yet inadequately developed [34]. Although PMBOK prioritizes time, money, quality, and risk management, it does not provide explicit means for integrating environmental sustainability into its project lifecycle stages [35]. Recent breakthroughs have brought sustainability-oriented project-management methodologies, such as PRiSM (projects integrating sustainable methods), which link PMBOK with environmentally-conscious project-management practices [36].

Piras et al. [37] discovered that the combination of PMBOK with digital tools like BIM and internet of things (IoT)-driven monitoring systems can markedly enhance sustainability compliance in large-scale projects. Nonetheless, the disjointed character of PMBOK guidelines and the hesitance of project managers to adopt sustainability-oriented approaches persist as considerable constraints. This gap indicates that PMBOK must adapt to include predictive analytics, automation, and simulation-based validation to facilitate sustainability-oriented decision-making.

2.4. ML in Energy Optimization and Environmental Impact Prediction

ML has become significant in sustainable building by facilitating predictive analytics for energy usage, anticipating carbon emissions, and optimizing resource allocation. ML models can evaluate previous energy-consumption trends, enhance HVAC systems, and recommend alternate materials to mitigate environmental effects [38].

Numerous studies have proven the efficacy of ML in improving sustainability metrics. Sundaram et al. [39] utilized deep learning algorithms to forecast real-time energy efficiency in commercial buildings, attaining an accuracy rate of 90%. Ntafalias et al. [40] combined ML with IoT sensors to enhance water and electricity efficiency in extensive residential developments, achieving a peak energy consumption reduction of up to 86% and an overall decrease of up to 60%. Notwithstanding its benefits, ML encounters many restrictions, including reliance on data, algorithmic intricacy, and the requirement for substantial computer resources. Moreover, ML models necessitate high-quality, organized datasets for optimal performance, a requirement frequently hindered by irregular data-gathering procedures in building projects. These limits underscore the necessity for a systematic application of ML with building information modeling and the PMOK to establish a more resilient, predictive sustainability framework.

2.5. Gap Identification and Contribution

Although BIM, PMBOK, and ML provide substantial benefits for sustainable construction, their amalgamation is currently constrained in existing studies. Most studies concentrate on one or two of these components in isolation rather than investigating their synergistic potential within a cohesive framework. Moreover, the lack of simulation-based validation models complicates the evaluation of the long-term sustainability advantages of these integrations.

Table 1. Comparative synthesis of studies on sustainable construction.

Reference	BIM Modeling	ML Application	PMBOK/Scheduling Integration	Sustainability Focus	Regional Adaptation	Key Limitations
Madkhali et al. [27]	-	-	-	Green policy analysis	Saudi regulations	Lacks digital integration
Ma et al. [32]	BIM-based LCA	-	-	Carbon footprint	-	No control feedback; static modeling
Ibe [33]	Material waste tracking	-	-	LCA and sustainability	-	Lacks energy simulation; no ML
Piras et al. [37]	BIM integration	-	PMBOK + IoT	Monitoring-focused	-	Non-predictive; lacks ML–BIM–PMBOK synergy
Sundaram et al. [39]	-	Deep learning	-	Commercial energy use	-	No BIM/project linkage
Ntafalias et al. [40]	-	IoT-enhanced ML	-	Water and energy reduction	-	Residential only; no BIM–PM interface
Current study	Parametric BIM + EnergyPlus	Optimization loop	PMBOK	Regulatory + design	Saudi SBC	Full integration and contextual relevance

presents a critical synthesis of selected recent studies that address various facets of sustainable construction, ranging from policy and regulatory frameworks to digital modeling and AI-driven analytics.

This study proposes a comprehensive framework that incorporates BIM, PMBOK, and ML inside a simulation-based environment to overcome these shortcomings. This approach employs predictive analytics and a sustainability metric assessment to enhance energy efficiency, resource usage, and carbon footprint reduction, in contrast to prior research that frequently lacks real-time validation. Simulation tools offer advanced functionalities for energy simulation and BIM-based modeling; however, they often operate in isolation and lack systematic integration with project-management methodologies and adaptive ML control systems. The suggested framework stands out by incorporating PMBOK’s sustainability-driven project workflows and ML-based predictive analytics into simulation processes, resulting in a holistic decision-support tool adapted to Saudi Arabia’s regulatory and climatic contexts. This research will utilize case studies (commercial complex and residential building) from Saudi Arabia to offer practical insights for surmounting legal and technological obstacles, while illustrating the scalability and viability of AI-enhanced sustainable project management.

3. Methodology

3.1. Framework Design

The design follows a modular, five-phase structure that facilitates seamless data exchange among simulation, optimization, and managerial components. Each phase plays a distinct role and feeds into subsequent steps via standardized data formats and application programming interfaces (Figure 1).

The proposed hybrid framework is operationalized through a structured, multi-platform data flow architecture. The interoperability begins with a BIM model created in Autodesk Revit, which provides geometrical and performance-related input data. The models are exported preserving thermal properties, material specifications, HVAC zoning, and occupancy profiles. The resulting data is then used to train predictive and optimization models, of which outputs are fed back into both the BIM environment (for updated design iterations) and a PMBOK-aligned control dashboard (for project-level decision-making).

The ML component trains predictive models to estimate energy consumption and CO₂ emissions. It generates optimized configurations by running multiple forward passes through the trained model, varying controllable design parameters, and selecting those that minimize energy consumption, while respecting defined constraints. The ML outputs are pushed to the PMBOK project control layer. Each predicted metric of energy consumption and CO₂ emission is benchmarked against target values. Deviations beyond tolerance thresholds trigger predefined control logic based on PMBOK’s Monitoring and Controlling and Risk Management knowledge areas.

Once a deviation is detected, the PMBOK layer initiates a rule-based response protocol. For example, if GHG emissions are predicted to exceed baseline limits, the PMBOK dashboard flags this issue and suggests a corrective action plan such as switching the façade material or altering the project schedule. This action is recorded in the project control log, and a feedback signal is generated to modify BIM parameters accordingly. Through this process, a data-governed feedback loop is established where quantitative ML outputs directly influence design updates and managerial decisions.

In the first phase, the building is modeled in Autodesk Revit, incorporating architectural elements, materials, glazing ratios, and HVAC components. The model is exported in a structured data format, typically gbXML, which serves as the foundation for energy simulation and feature extraction. In the second phase, EnergyPlus simulates baseline energy performance, yielding key performance indicators such as energy consumption and CO₂ emissions. These factors are used both as training targets for ML and as benchmark values for PMBOK-based monitoring. The simulation output is processed and formatted into feature vectors for input into ML models. The ML models are trained to predict energy performance under various scenarios (Phase 3). The PMBOK framework serves as the project control layer, where managerial processes interpret ML outputs in light of project KPIs. For instance, if predicted GHG emissions exceed acceptable thresholds, the Risk Management process is activated to reassess building envelope materials or revise construction schedules. PMBOK’s Monitoring and Controlling knowledge areas align with automated alerts and decision pathways established in the ML module (Phase 4). Any deviation in energy key performance indicators or sustainability indicators triggers a closed-loop feedback process that returns the control to the BIM modeling phase. Adjustments to geometry, materials, or occupancy schedules are implemented and resubmitted for simulation, retraining, and managerial review. This cyclic flow ensures that design decisions continuously evolve in response to predictive insights and management goals.

3.1.1. BIM Integration and Energy Modeling

BIM served as the primary instrument for developing digital twins of construction projects. These models were employed to mimic energy dynamics within structures and throughout large metropolitan environments. Comprehensive energy modeling was conducted within BIM, facilitating an in-depth investigation of energy consumption and thermal performance under many environmental circumstances pertinent to Riyadh’s climate. BIM facilitated the ongoing optimization of energy-consumption patterns in building design and material choices through real-time updates and simulations reflecting changing conditions [41,42].

The capacity of BIM to monitor energy efficiency throughout the full lifecycle of a project (from design to construction and post-occupancy) was essential for the simulation. The energy-consumption data obtained from BIM models were amalgamated with the simulation to furnish comprehensive insights into the effects of construction techniques on energy usage and emissions over time. This integration facilitated the creation of sustainable architectural solutions that comply with Riyadh’s energy efficiency criteria while reducing GHG emissions [43,44,45].

Simulations were employed to evaluate the energy consumption of various building designs, including differences in insulation, HVAC systems, and window positions. The data from these simulations facilitated the optimization of thermal comfort in buildings while minimizing energy consumption during operation. The real-time visualization capabilities of BIM facilitated the identification of potential concerns prior to the commencement of building, hence improving energy-efficient results.

3.1.2. ML Combination, Predictive Analytics for Energy Optimization, and Environmental Impact

ML was integrated into the framework to improve the system’s predictive capabilities, facilitating the more precise modeling of energy usage and environmental impacts throughout a project’s lifecycle. Supervised learning algorithms were employed to examine historical data regarding energy consumption, meteorological patterns, and building efficiency.

Within the simulation-based framework, ML was employed to discern patterns in energy consumption, facilitating the optimization of building systems for peak efficiency. Clustering techniques were utilized to categorize buildings with analogous energy-consumption trends, facilitating more precise treatments. Regression models were employed to forecast carbon emissions based on energy consumption, facilitating the evaluation of the environmental consequences of various construction decisions.

3.1.3. PMBOK Combination, Sustainable Project Management Practices

PMBOK offered a systematic methodology for project management, guaranteeing the integration of sustainability concepts throughout the project’s lifecycle. The PMBOK framework for managing scope, time, money, and quality was modified to incorporate sustainability objectives [46]. The PMBOK methodology was utilized to delineate explicit roles and responsibilities for the integration of sustainability within the project team.

Furthermore, PMBOK’s risk-management procedures were employed to address potential risks related to energy inefficiency and environmental consequences. The system enabled ongoing monitoring and modification, guaranteeing that sustainability objectives were achieved at each phase of the project. During the design phase, risk evaluations revealed possible energy inefficiencies that might be addressed through material selection or design modifications [47,48]. The PMBOK focuses on four major knowledge areas that are critical for sustainability-driven project delivery, scope management, schedule management, risk management, and monitoring and controlling. Each of these is mapped to specific processes within the BIM and ML modules, forming an active control loop.

Table 2. PMBOK to operational tasks.

PMBOK Knowledge Area	Inputs	Output	Execution/Work Task
Scope Management	Project charter; BIM LOD data	Model refinement; scope redefinition	Define BIM model LOD levels based on target energy KPIs; adjust spatial zoning to match envelope constraints
Schedule Management	BIM schedule; ML iteration time	Activity rescheduling; buffer reallocation	Align simulation cycles with construction milestones; allocate buffers for iterative recalibration of energy models
Risk Management	ML stochastic sensitivity; performance variance	Preventive design/material changes	Trigger re-evaluation of HVAC system or material selection if ML output deviates > ±15% from benchmark efficiency
Cost Management	Lifecycle cost output from EnergyPlus; ML-based utility forecast	Budget adjustment; cost-risk balancing	Perform trade-off analysis between alternative energy systems under varying tariff scenarios

presents a refined and actionable mapping of PMBOK knowledge areas to specific digital inputs, control outputs, and practical work tasks within the proposed energy-management framework.

3.2. Case Study: Riyadh, Saudi Arabia

Riyadh, the capital of Saudi Arabia, is a rapidly growing urban hub with a unique set of environmental challenges and opportunities. It is essential to the nation’s Vision 2030 environmental measures, which seek to diversify the economy and diminish reliance on oil. The city’s severe climate, elevated energy usage, and aggressive growth initiatives render it an exemplary case study for examining the amalgamation of BIM, PMBOK, and ML to enhance energy efficiency and mitigate environmental effects in building.

Table 3. Key characteristics of Riyadh relevant to the current study.

Characteristic	Description	Relevance to Study
Population Size	Riyadh’s population is estimated at approximately 7.5 million	A large population correlates with higher energy demand, emphasizing the importance of energy optimization and sustainability in urban development.
Urban Growth Rate	Riyadh is expanding at a growth rate of around 4% per year, with urban sprawl extending to surrounding areas	Rapid urbanization demands sustainable construction practices to manage energy usage, making it a prime location for applying BIM and ML in real-world scenarios.
Energy Consumption	Riyadh consumes approximately 40% of the country’s total electricity supply, with residential cooling accounting for 60% of total energy usage	Energy demand for cooling is a major challenge, making energy efficiency optimization using BIM and ML crucial for sustainability.
Construction Projects	Over 60 major development projects are currently underway in Riyadh, including the King Salman Park, NEOM, and King Abdullah Financial District	These large-scale projects provide an ideal environment for testing the integration of BIM, PMBOK, and ML for energy optimization and green construction.
Energy Efficiency Goals	Saudi Arabia has set a target to reduce energy consumption by 20% by 2030, with Riyadh at the forefront of these initiatives	Aligns with the study’s focus on developing a framework that helps achieve national energy efficiency targets using digital tools like BIM and ML.
Green Building Initiatives	Riyadh is investing in 30+ green building projects in alignment with the Saudi Green Building Code	Provides a regulatory foundation for applying BIM and PMBOK in sustainable building practices.
Smart City Developments	Riyadh’s smart city initiatives include projects like the Riyadh Metro and Riyadh’s Smart Infrastructure Program to improve energy efficiency	These projects provide the infrastructure for integrating ML for real-time energy monitoring and optimization.

presents a comprehensive examination of the principal characteristics of Riyadh, which are crucial for this study.

Riyadh’s rapid urban growth, high energy consumption, green building initiatives, and smart city projects make it an ideal case study for applying this simulation-based framework. The city’s ongoing commitment to achieving Vision 2030 sustainability goals provides a fertile ground for integrating BIM, PMBOK, and ML. Through this case study, Riyadh could benefit from optimized energy usage, reduced carbon emissions, and enhanced sustainability metrics. By aligning the current study with Riyadh’s development plans and sustainability targets, this research has the potential to provide critical insights into the scalability and effectiveness of such integrated frameworks for cities across the region.

Given Riyadh’s rapid urban growth, high energy consumption, and ambitious sustainability goals, conducting real-world case studies to validate the proposed framework may pose significant logistical and time-related challenges. Real-world construction projects, especially large-scale ones, often involve complex and long-term planning cycles, making it difficult to monitor energy efficiency and environmental impact in real time. Furthermore, the integration of BIM, PMBOK, and ML in real-world projects could be constrained by the availability of accurate data, the cost of implementing such technologies across ongoing projects, and the unpredictability of construction schedules.

Simulation-based validation offers several advantages in this context. First, it allows for the controlled testing of various energy optimization scenarios and environmental impact assessments under different conditions, without the risk of disrupting actual construction projects. Second, simulations provide the opportunity to model a variety of potential future scenarios (such as different urban development strategies, changes in energy-consumption patterns, or policy interventions) thus offering a broader understanding of how the integrated framework can contribute to sustainable urban development in Riyadh. Lastly, simulation-based methods enable the study to incorporate historical and predictive data, including weather patterns, energy demand fluctuations, and the effects of specific building designs, which may not be fully available in real-world case studies.

3.3. Building Type Selection

To assess the adaptability and efficacy of the integrated BIM–PMBOK–ML framework in practical scenarios, two different building types were chosen: (i) a residential villa, and (ii) a large-scale commercial structure. The choice of these building types was influenced by their differing operational characteristics, energy-consumption patterns, and significance to Saudi Arabia’s Vision 2030 sustainability initiative. The residential sector constitutes a significant share of energy consumption due to extensive dependence on air conditioning, whereas commercial buildings, especially big institutional facilities, are defined by centralized HVAC systems, prolonged operational hours, and elevated interior loads. The study sought to encompass the comprehensive range of energy-consumption habits in Riyadh and to illustrate the versatility of the proposed framework across various building typologies.

The commercial case study involved a large, multi-functional office and service building located in Riyadh. The building was evaluated as a commercial facility due to its large scale, office-type occupancy, centralized HVAC systems, and operational energy patterns, all of which align with the characteristics of typical commercial structures such as office buildings and business centers. In contrast, the residential case study, examined a standard two-story detached villa in Riyadh, emblematic of middle-income housing in Saudi Arabia. Both edifices are located in Climate Zone 1, distinguished by exceptionally elevated cooling requirements in the summer. Their principal specifications are encapsulated in

Table 4. Building characteristics.

Parameter	Commercial Building	Residential Building
Total Floor Area	89,970 m2	403 m2
Number of Floors	Multiple	2
Construction Type	Prefabricated concrete panels with insulation	Masonry (Concrete blocks) without insulation
Roof U-value	0.34 W/m2·K	2.13 W/m2·K
Wall U-value	0.29 W/m2·K	2.15 W/m2·K
Window Type	Double-glazed low-emissivity	Single-pane glazing (U = 5.78 W/m2·K)
HVAC System	Central Chillers, FCUs, AHUs	Packaged DX split AC units (CoP = 2.5)
Annual Energy Consumption	13,859 MWh (≈154 kWh/m2/year)	46,143 kWh (≈114.5 kWh/m2/year)
Occupancy	≈1000 persons	6 persons

.

The utilization of both residential and commercial case studies provided a solid foundation for analyzing the advantages of the suggested framework across various situations. The commercial structure, characterized by its consistent working timetable, substantial internal energy gains, and centralized mechanical systems, served as an optimal environment for assessing large-scale energy-management solutions. The prefabricated insulated envelope facilitated detailed simulations to assess the effects of envelope performance and system optimization. In contrast, the residential building presented constraints associated with tenant behavior, inconsistent thermal loads, and inadequate envelope insulation—issues prevalent in Saudi housing stock.

3.4. ML Model Development

In the current study, a combination of neural networks (NNs), decision trees (DTs), and reinforcement learning (RL) algorithms was applied to optimize energy efficiency and reduce environmental impacts in construction.

3.4.1. Neural Networks (NNs)

The NN model was chosen due to its ability to model non-linear relationships and handle large volumes of complex data [49,50]. These models were trained using BIM-based energy performance datasets, which allowed for capturing intricate patterns between energy use, building design parameters, and external environmental factors (e.g., temperature, humidity, occupancy).

The forward propagation process in an NN involves calculating the input to each layer and applying an activation function. For each layer i, the activation is computed as (Equation (1)) [51].

$${\mathrm{z}}^{\left(i\right)}={\mathrm{W}}^{\left(i\right)}\cdot {\mathrm{a}}^{(i-1)}+{\mathrm{b}}^{\left(i\right)},$$

(1)

where, z⁽ⁱ⁾ is the input to the activation function of layer i. W⁽ⁱ⁾ is the weight matrix for layer i. a⁽ⁱ⁻¹⁾ is the activation from the previous layer. b⁽ⁱ⁾ is the bias term.

The activation function a⁽ⁱ⁾ for Rectified Linear Unit (ReLU) is computed as (Equation (2)).

$${\mathrm{a}}^{\left(i\right)}=max\left(0, {\mathrm{z}}^{\left(i\right)}\right).$$

(2)

The mean squared error (MSE) is the loss function used to optimize the weights during the training process. It quantifies the difference between the actual energy consumption y_i and predicted values ${\widehat{y}}_i$ across all data points (Equation (3)).

$$L_{\mathrm{M}\mathrm{S}\mathrm{E}}={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N {\left(y_i-{\widehat{y}}_i\right)}^2,$$

(3)

where N is the number of data points. y_i is the actual value. ${\widehat{y}}_i$ is the predicted value.

3.4.2. Decision Trees (DTs)

DTs were incorporated to gain transparency in the model and interpretability of how certain building factors contributed to energy efficiency and carbon emissions. DTs were used to identify critical thresholds where energy usage would significantly increase or decrease, helping make practical decisions on resource utilization [52].

Gini impurity is the most commonly used criterion for splitting nodes in decision trees. It measures the impurity of a node and is calculated as (Equation (4)).

$$\mathrm{Gini} =1-{\sum }_{i=1}^C p_i^2,$$

(4)

where p_i is the probability of an instance being classified into class i. C is the number of classes.

The information gain is used to determine the best feature to split on in a decision tree. It is calculated as (Equation (5)) [51].

$$\mathrm{Information} \mathrm{Gain} = \mathrm{Entropy} (\mathrm{Parent})-\sum _i {\displaystyle \frac{N_i}{N}}\cdot \mathrm{Entropy} \left({\mathrm{Child} }_i\right)$$

(5)

where, Entropy (Parent) is the entropy of the parent node. N_i is the number of instances in child node i, and N is the total number of instances in the parent node.

3.4.3. Reinforcement Learning (RL)

The Q-learning algorithm, a core component of RL, is used to update the value function Q(s,a), which represents the expected reward for taking action a in state s. The Q-value is updated iteratively as follows (Equation (6)) [53].

$$R\left(s_t,a_t\right)=Q\left(s_t,a_t\right)+\alpha \left(r_{t+1}+\gamma \underset{a^{\prime }}{max} Q\left(s_{t+1},a^{\prime }\right)-Q\left(s_t,a_t\right)\right),$$

(6)

where Q(s_t,a_t) is the Q-value for state s_t and action a_t. α is the learning rate. γ is the discount factor. r_t+1 is the reward obtained after taking action a_t in state s_t. max_a′Q(s_t+1,a′) is the maximum Q-value for the next state s_t+1.

This update rule enables the RL agent to make real-time adjustments to building operations, such as optimizing HVAC systems, lighting, and insulation, based on environmental feedback. By continually adjusting the Q-values, the RL model can optimize energy consumption over the course of the building’s lifecycle.

The data for training the ML models were collected from BIM-based energy performance datasets. These datasets contained detailed building parameters, such as floor plans, materials, HVAC system types, and energy-consumption histories. The data were complemented by environmental factors, including local temperature patterns, solar radiation, and wind speed, collected from Riyadh’s weather stations. The BIM models provided the structural data, which were combined with real-time energy usage records from buildings to create a comprehensive dataset suitable for ML analysis. The data were preprocessed to remove outliers and ensure consistency across time periods. Missing values were handled using interpolation methods based on historical data, and categorical data (e.g., building materials, insulation types) were encoded using one-hot encoding techniques to prepare the data for ML algorithms.

The performance of the ML models was evaluated using various metrics relevant to energy efficiency and sustainability goals. These metrics include root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), normalized root mean square error (NRMSE), and the coefficient of determination (R²).

RMSE is calculated as follows (Equation (7)) [54].

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{N}}{\sum }_{i=1}^N {\left(y_i-{\widehat{y}}_i\right)}^2}.$$

(7)

where N is the number of data points. y_i is the actual value. ${\widehat{y}}_i$ is the predicted value.

MAE measures the average magnitude of errors in a set of predictions, without considering their direction. It provides a straightforward measure of average model error and is less sensitive to outliers than RMSE (Equation (8)):

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N \left|y_i-{\widehat{y}}_i\right|.$$

(8)

R² (coefficient of determination) is used to assess how well the model explains the variance in GHG emissions (Equation (9)):

$${\mathrm{R}}^2=1-{\displaystyle \frac{\sum _{i=1}^N {\left(y_i-{\widehat{y}}_i\right)}^2}{\sum _{i=1}^N {\left(y_i-\overline{y}\right)}^2}}.$$

(9)

where $\overline{y}$ is the mean of the actual values.

MAPE provides a scale-independent percentage-based evaluation of prediction accuracy, making it easier to interpret errors across different building types or time periods (Equation (10)) [55].

$$\mathrm{M}\mathrm{A}\mathrm{P}\mathrm{E}={\displaystyle \frac{1}{n}}\sum _{i=1}^n{\displaystyle \frac{\left|y_i-\widehat{y_i}\right|}{y_i}}\times 100.$$

(10)

NRMSE is used to normalize the RMSE by the mean of the actual values, making it easier to compare across datasets of different scales (Equation (11)):

$$\mathrm{N}\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}={\displaystyle \frac{\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}{\overline{y}}}\times 100.$$

(11)

Also, the normalized mean bias error (NMBE) was used to validate the simulation outputs against real-world data to ensure their accuracy and reliability. NMBE is a statistical metric commonly used in model validation, especially in building energy simulation, to quantify how much a model systematically overestimates or underestimates the measured values (Equation (12)).

$$\mathrm{N}\mathrm{M}\mathrm{B}\mathrm{E}={\displaystyle \frac{\sum _{i=1}^n\left(y_i-\widehat{y_i}\right)}{n·\overline{y}}}\times 100.$$

(12)

An NMBE > 0 shows that the model underestimates on average, while an NMBE < 0 indicates that the model overestimates on average. Also, an NMBE = 0 shows no bias (perfect prediction).

NN models were tuned using both grid search and random search methods to optimize their architecture and learning parameters. The hyperparameters adjusted included the number of hidden layers (ranging from 2 to 5), the number of neurons per layer (between 32 and 256), and the learning rate (from 0.001 to 0.01). An ReLU activation function was applied in all hidden layers, and the Adam optimizer was used to minimize the MSE loss function. The dataset was split into 80% training and 20% testing subsets. The validation set was used to prevent overfitting and to identify the optimal configuration based on early stopping criteria.

For DT models, the CART (classification and regression tree) algorithm was used. The key parameters tuned included the maximum tree depth, which was varied between 5 and 20, and the minimum number of samples per leaf, which was set between 2 and 10. The best model was selected based on its performance on the validation set using gini impurity as the splitting criterion. Pruning techniques were also applied to minimize overfitting and ensure generalization.

In the RL setup, Q-learning was implemented to guide sequential decision-making for dynamic energy optimization. The Q-values were updated using a learning rate (α) of 0.1 and a discount factor (γ) of 0.9. The exploration–exploitation balance was managed using an epsilon-greedy strategy, where the epsilon value decayed from 0.9 to 0.1 over training episodes. The state space included the indoor temperature, occupancy levels, and external weather data, while actions involved adjusting the HVAC setpoints, lighting intensity, and operational schedules.

All models were trained using BIM-based energy performance datasets, enriched with real-time environmental data collected from building sensors and weather stations in Riyadh. The final outputs generated by each model included daily energy-consumption forecasts and corresponding GHG emission estimates, which were validated using statistical performance metrics such as RMSE, MAE, MAPE, NRMSE, and R².

3.5. Implementation of BIM and PMBOK Principles

In this study, the BIM framework was integral to the simulation of energy performance and sustainability optimization. Specifically, Autodesk Revit and EnergyPlus were employed to model the energy behavior of buildings and to simulate their environmental performance under different operational conditions.

Autodesk Revit was used as the primary BIM tool for designing the architectural and structural elements of the buildings in the study. The software allowed for the creation of detailed 3D models that represent not only the building’s layout but also the properties of materials, insulation, and windows, which are critical factors influencing energy consumption. The BIM model developed in Revit included detailed data on the building geometry, heating, cooling, lighting, and ventilation systems, which were necessary for accurate energy analysis.

Once the building design was finalized in Revit, the energy simulation was transferred to EnergyPlus, a powerful energy modeling software that is widely used for evaluating building energy performance. EnergyPlus was chosen due to its capability to simulate complex energy systems and its detailed models for building heating, cooling, and lighting demands. Through the integration of Revit with EnergyPlus, the energy-consumption patterns of the building were predicted across various scenarios, taking into account local climate conditions, operational schedules, and energy efficiency measures.

EnergyPlus allowed for the simulation of building thermal dynamics, HVAC systems, and solar heat gain based on real-time environmental data and building parameters. The tool was crucial in modeling how different building design strategies, such as window orientation or insulation material selection, could impact energy demand and consumption. These simulations provided insights into the most energy-efficient configurations for reducing the building’s carbon footprint, aligning with the overall goal of the study to promote sustainable building practices.

The integration of BIM tools in the simulation allowed for dynamic energy assessments, which helped to identify energy inefficiencies early in the design process. It also facilitated the exploration of multiple design alternatives and their corresponding environmental impacts, offering a comprehensive approach to energy optimization.

To ensure that the simulation results aligned with sustainable construction practices, the study also incorporated principles from the PMBOK, particularly focusing on its sustainability-oriented project-management practices. PMBOK provides a structured approach to managing projects and integrates sustainability into its framework through the emphasis on stakeholder engagement, resource optimization, and risk management.

The PMBOK process groups (initiating, planning, executing, monitoring and controlling, and closing) were applied at each stage of the project. The initiating phase involved identifying key sustainability goals, such as energy efficiency targets and GHG reduction objectives, which were aligned with Saudi Arabia’s Vision 2030 sustainability goals. During the planning phase, specific energy-efficient measures and sustainability actions were developed, such as the incorporation of green building standards (e.g., Saudi Green Building Code) and energy-saving strategies informed by the results of the BIM-based energy simulations.

In the executing phase, the project team used the data generated by the BIM tools to inform decisions about building materials, system installations, and energy-saving techniques. For instance, construction teams were guided by the energy performance data provided by EnergyPlus, ensuring that the chosen materials and systems were in line with the sustainability objectives.

The monitoring and controlling phase involved tracking the project’s sustainability metrics in real-time. PMBOK emphasizes the importance of monitoring and controlling throughout the project lifecycle, which was facilitated through BIM’s real-time monitoring capabilities and the predictive analytics provided by ML models. For example, the project management team could track energy-consumption patterns and make adjustments as needed to maintain alignment with sustainability goals. Additionally, the risk-management processes outlined in PMBOK were applied to identify potential risks related to energy consumption and environmental impacts and to develop mitigation strategies.

Finally, during the closing phase, the sustainability outcomes of the project were assessed. PMBOK encourages the documentation of project outcomes, which in this case included the final energy performance of the building, GHG emissions reduction, and the achievement of other sustainability targets. This phase also ensured that any lessons learned from the project regarding energy management and sustainability practices were captured for future projects.

4. Results and Discussion

4.1. Evaluation of Simulation Accuracy Using Electricity Bill Data

As this study relied on simulation modeling, it was essential to validate the simulation outputs against real-world data to ensure their accuracy and reliability. Actual monthly electricity bills were collected from both the residential and commercial buildings under investigation. The collected utility data were cross-checked with local climate profiles and occupancy schedules to validate their consistency and relevance.

Table 5. Mean monthly energy consumption for residential and commercial buildings.

Month	Residential (Measured, kWh)	Residential (Simulated, kWh)	Commercial (Measured, kWh)	Commercial (Simulated, kWh)
January	1040	1045	667,200	667,104
February	934	947	708,835	70,875
March	1587	1560	934,440	934,496
April	3041	3050	1,158,004	1,158,100
May	5186	5180	1,424,167	1,424,100
June	7067	7070	1,571,746	1,571,800
July	7470	7455	1,643,926	1,644,010
August	7700	7688	1,528,126	1,528,165
September	6469	6478	1,290,020	1,290,048
October	4532	4549	1,261,379	1,261,124
November	2185	2193	1,044,581	1,044,565
December	1060	1048	845,196	845,103

compares the measured and simulation-based mean monthly energy consumption (in kWh) for the residential and commercial buildings. These values were used to assess the reliability of the simulation results.

To quantify the similarity between simulated and measured values,

Table 6. Statistical comparison between measured and simulated monthly energy consumption.

Metric	Residential	Commercial
RMSE (kWh)	12.89	18,416.21
MAE (kWh)	11.33	5323.00
MAPE	0.53	7.50
NRMSE	0.32	15.69
NMBE	−0.01	−4.53

presents five statistical evaluation metrics, including RMSE, MAE, MAPE, NRMSE, and NMBE. These metrics confirm the credibility of the measured data as representative inputs and validate the robustness of the simulation process.

The residential model performs very well and is scientifically credible for use in this simulation-based study. Although not as strong as the residential model, the commercial model still shows sufficient accuracy for simulation validation. This issue confirms that the proposed BIM–PMBOK–ML framework is capable of generating accurate and reliable energy performance simulations.

4.2. ML-Based Prediction Results Analysis

This section presents the results of ML model predictions for energy consumption and GHG emissions across both residential and commercial buildings. Three ML algorithms—NN, DT, and RL—were trained using BIM-based simulation data and evaluated against simulation outputs using time-series trends, scatter plots, and statistical metrics. The aim was to assess each model’s predictive performance, generalization capacity, and suitability for real-time or large-scale deployment in sustainable building management.

4.2.1. Time Series Comparison of ML Predictions vs. Simulation Outputs

Daily energy consumption and GHG emission profiles were predicted using the trained models and compared with simulation baselines to evaluate how accurately each ML model could replicate simulation results over time. This comparison allows assessments of each model’s ability to capture seasonal fluctuations, peak demands, and transitional periods (Figure 2, Figure 3, Figure 4 and Figure 5).

The NN model (Figure 2a) closely tracks the simulation output throughout the entire year, particularly during the transition into peak summer months (May to September). The NN prediction nearly overlaps the simulation curve, demonstrating the model’s ability to capture both seasonal trends and short-term fluctuations. The model shows a very low deviation during low-consumption months (winter) and only minor overestimations during peak summer demand, confirming its high accuracy and generalization ability. The RL model (Figure 2b) also follows the simulation trend effectively, especially in the shoulder seasons (spring and fall). Its strength lies in maintaining responsive adjustments during fluctuating periods, reflecting its adaptive learning capability. However, it exhibits slightly more variance around peak summer (July and August), likely due to overfitting to extreme operational scenarios, which RL is designed to adapt to but not always generalize from as cleanly as NN. In contrast, the DT model (Figure 2c) presents the most notable fluctuations from the baseline, particularly evident in overpredictions during high-consumption periods and underpredictions during winter months. The sharp transitions and jaggedness reflect the model’s stepwise structure, which tends to oversimplify complex energy dynamics. Although DT models are valuable for interpreting threshold behaviors, their lack of smooth approximation limits their precision in capturing daily load variability. Across all three models, the overall seasonal trend of increasing energy use during summer and reduced consumption during winter is consistently captured. However, model fidelity varies: (i) NN offers the most precise temporal alignment, (ii) RL provides adaptive performance, and (iii) DT reveals threshold-driven estimation behavior that is less suitable for fine-grained prediction tasks.

In Figure 3a, the NN model demonstrated high alignment with the simulation throughout the year, effectively capturing the progressive rise in energy consumption from March to August, which corresponds to the transition into the summer season. The model also reflected the slight decline beginning in September and the stabilization observed during the cooler months of November through February. Minor overestimations were observed in the peak summer period, but the overall trend was accurately followed, indicating strong generalization and consistency in high-load scenarios. The RL model shown in Figure 3b also performed well, particularly during transitional months such as April, May, and October. The model exhibited adaptive responses to short-term variations in energy demand, a behavior matched with its dynamic learning framework. However, in the core summer months of June through August, the predictions showed more noticeable fluctuations around the simulated values, suggesting some sensitivity to sharp peaks in cooling demand. The performance of the DT model in Figure 3c was characterized by more pronounced irregularities, especially during summer. While the model was generally successful in tracking the overall trend, it tended to exaggerate variations in high-consumption months and slightly underestimate usage in the early months of the year. This behavior is indicative of the model’s piecewise learning structure, which can limit its ability to represent gradual or non-linear changes in energy patterns.

In Figure 4a, the GHG emission profile forecasted by the NN model closely aligns with the simulated trend over the course of the year. The model effectively documented the incremental increase in emissions commencing in March and reaching its zenith throughout the peak consumption summer months of June, July, and August. The progression from September to the year’s conclusion is similarly perfectly synchronized. Minor underestimations are evident in late spring, whereas modest overshoots transpire throughout summertime; however, the errors stay within an acceptable range. This signifies that the NN model is exceptionally proficient in representing non-linear correlations among input variables, including the temperature, energy consumption, and emission rates. Figure 4b illustrates the predictions generated by the RL model. This model demonstrates strong concordance with the simulation, especially during seasonal transitions in April and October. While the overall trend is accurately represented, certain fluctuations are evident in the warmer months, during which the model marginally overestimates emissions amid swift rises in the cooling demand. Nonetheless, RL sustains a stable baseline year-round and demonstrates its capacity to adjust to fluctuating energy-consumption patterns in residential environments. In Figure 4c, the DT model provides a reasonable approximation of the GHG emissions trend, correctly capturing the seasonal rise and fall in emissions. However, the model exhibits more noise compared to the other two, especially from May to August, where it tends to overestimate peak emissions. The lack of smooth transitions is characteristic of the model’s discrete structure, which can lead to sharp jumps in prediction when facing continuous or gradually changing input conditions. This reduces its suitability for accurate day-by-day emission forecasting, though it can still provide useful threshold-based insights.

NN indicated remarkable fidelity to the simulation, precisely following the gradual increase in GHG emissions from late winter into spring, culminating in peak levels throughout the summer months of June, July, and August. The trend of emission reductions from early October to December was accurately replicated. Intermittent overestimations transpired in July; yet, the model consistently upheld its integrity and effectively maintained seasonal dynamics. This outcome validates the NN model’s efficacy in elucidating intricate linkages between energy demand and environmental variables in extensive commercial contexts (Figure 5a). The RL model effectively adjusted throughout transitional seasons like May and October, maintaining the shape and amplitude of the summer peak. In contrast to NN, more pronounced deviations are observed around July and September, demonstrating the model’s sensitivity to real-time fluctuations in energy use while also highlighting heightened variability under swiftly changing load conditions (Figure 5b). The DT model accurately reflected the overall emission trend but exhibited increased noise, especially during the high-demand period from June to September. The model frequently overestimated GHG emissions at various times over the year, particularly in July and early autumn. The anomalies stem from the model’s inclination to discretize input–output linkages, which is less adept at simulating continuous emission behaviors in dynamic operational contexts. Notwithstanding this constraint, the DT model continued to discern overarching trends and significant alterations in emission levels (Figure 5c).

4.2.2. Analysis of Predicted vs. Simulated Values

To further validate the precision of the models, scatter plots were generated comparing predicted energy values against simulation outputs. A strong alignment along the diagonal line reflects the high predictive accuracy. These plots are critical for identifying bias, overfitting, and underprediction patterns (Figure 6 and Figure 7).

The NN model (Figure 6a) achieved a highly concentrated distribution of points near the diagonal, indicating excellent agreement with the simulation. Predictions were consistent across the full range of energy demand values, from low consumption days in winter to peak loads during summer. This performance highlights the model’s capacity to generalize from historical patterns and handle non-linear dependencies inherent in residential usage. In Figure 6b, the RL model also demonstrated solid alignment with simulation data, particularly in the middle range of consumption (10–18 kWh/m²). While slightly more scattered than the NN case, the majority of points still remained close to the reference line. This result reflects the RL model’s ability to adaptively learn energy patterns over time, although minor discrepancies emerged during periods of extreme demand. Figure 6c shows the DT model, which presented the widest dispersion among the three. While it captured the general increasing trend, the spread of points, especially for mid-to-high demand days, reveals the model’s reduced precision. These deviations are consistent with its rule-based nature, which may oversimplify transitions and underperform in capturing continuous variations in energy dynamics.

The commercial building predictions show similar trends but with slightly higher overall accuracy, particularly in the NN and RL models. Figure 6d indicates that the NN model performed exceptionally well, with most points lying close to the diagonal and covering the entire spectrum of operational loads. The more structured and consistent energy usage typical of commercial facilities likely contributed to this enhanced performance. Figure 6e, representing the RL model, also shows a tight clustering of data points, though with marginally higher variance than the NN results. The predictions maintained a strong correlation with the simulation, supporting the model’s ability to handle complex, real-time learning scenarios in structured environments like commercial buildings. The DT model (Figure 6f) again showed greater scatter compared to the other two models. While the overall trend was aligned, noticeable overestimations and underestimations were observed, particularly at the upper end of the demand range. This reinforces the limitations of decision trees in continuous, high-resolution energy-forecasting tasks, especially when non-linear behavior and interdependencies between variables are pronounced.

The NN model shows a strong agreement with the simulation values. The data points are tightly clustered around the diagonal, especially in the mid-to-high emission range (4–12 kgCO₂/m²). This indicates that the model successfully captured seasonal fluctuations in GHG emissions related to energy demand and external environmental conditions. Minor dispersions at the lower emission levels likely reflect the reduced predictability during milder weather months, where HVAC usage becomes more variable (Figure 7a). Figure 7b presents the RL model, which also maintains a high level of predictive accuracy, though with slightly more dispersion than NN. The model performs particularly well in the middle emission range (6–10 kgCO₂/m²), while a few predictions in the lower range slightly underestimate the actual values. Despite this, the RL model retained a generally strong correlation with the simulation, suggesting its capability to adapt to fluctuating energy-driven emissions. The DT model captures the overall trend of the simulation but demonstrates a broader spread of points, especially between 5 and 10 kgCO₂/m². The model’s rule-based structure led to more approximation in emission forecasting, and while it captured the emission pattern well, its performance lagged behind that of NN and RL in terms of accuracy. Still, the DT model’s consistency across the range supports its utility in offering interpretable, if slightly less precise, predictions (Figure 7c).

In Figure 7d, the NN model again exhibits high fidelity to the simulation data, with minimal deviation across all levels of GHG emissions. The commercial dataset, characterized by more regular operational loads, enabled the neural network to make highly accurate predictions, particularly between 8 and 13 kgCO₂/m². This strengthens the case for using neural networks in emissions forecasting for large-scale, systematically operated structures. Figure 7e illustrates that RL model’s results for commercial buildings. Although the points are generally well aligned with the diagonal, a moderate underestimation trend can be observed in the lower emission ranges (around 5–7 kgCO₂/m²). Nevertheless, the model maintains a strong overall correlation and effectively tracks the dominant emission trends driven by HVAC and lighting systems, especially during peak operational periods. Finally, Figure 7f presents the DT model, which shows acceptable performance in tracking GHG emissions. While the spread is greater than that observed in the NN and RL models, the overall trajectory aligns with simulation values. The broader scatter, especially above 12 kgCO₂/m², reveals the model’s limitations in accurately capturing sudden variations in emission outputs due to design or operation changes. Still, the DT model successfully captured the dominant patterns, reinforcing its interpretability advantage in scenarios requiring explainable results.

4.2.3. Comparison of ML Model Performance Metrics

Statistical performance metrics were compared using radar charts for both training and test sets to provide a comprehensive evaluation of the models for the energy consumption and GHG emission. Metrics include RMSE, NRMSE, MAE, MAPE, and R² (Figure 8 and Figure 9).

In the training phase for residential buildings, the NN model outperformed both RL and DT across all evaluation metrics. It attained the highest R² value, approaching 0.94, indicating a strong ability to explain variance in the training dataset. Additionally, NN recorded the lowest RMSE, MAE, and NRMSE values, reflecting high accuracy and the consistent learning of energy-consumption patterns. While the RL model followed closely, its slightly elevated MAE and NRMSE scores suggest a modestly reduced precision compared to NN. In contrast, DT exhibited noticeably inferior performance, with a lower R² of around 0.83 and comparatively higher error metrics, particularly in MAPE, underscoring its limited capability in capturing the non-linear dynamics of residential energy consumption (Figure 8a). During the residential testing phase, the NN model retained its superior performance, demonstrating minimal degradation in accuracy. The R² remained above 0.90, and error values remained consistently low, highlighting the model’s robustness and strong generalization ability. RL continued to perform competitively, maintaining a balanced profile across all metrics, though it showed slightly higher NRMSE and MAPE values than NN. DT again lagged behind, with a further reduction in R² and a pronounced increase in RMSE and MAE, confirming its weaker predictive reliability when applied to unseen residential data (Figure 8b).

For commercial building training, all three models achieved better overall performance compared to the residential case, likely due to the structured and predictable energy usage typical of commercial operations. NN once again led with the highest R², nearing 0.95, and the lowest RMSE, MAE, and NRMSE, validating its ability to accurately learn complex patterns in large-scale energy data. RL remained competitive, showing slightly higher error levels but retaining a solid R² of approximately 0.92. DT’s performance, while improved relative to its residential counterpart, still trailed in accuracy, particularly in MAPE and NRMSE, reinforcing its limited effectiveness for continuous energy-prediction tasks (Figure 8c). In the commercial testing phase, the NN model demonstrated continued dominance, maintaining a high R² above 0.90 and stable, low error values across all metrics. This minimal performance drop from training to testing highlights the model’s robustness and low overfitting tendency. RL followed closely, delivering consistent results and particularly strong performance in NRMSE. DT, however, exhibited the largest performance gap, with elevated MAE and MAPE values and a lower R² around 0.80. These results affirm that NN is best suited for accurate and generalizable energy-consumption forecasting in both residential and commercial contexts, while RL provides a reliable alternative for adaptive energy management. DT, despite its interpretability, is less effective for high-resolution, non-linear prediction tasks in dynamic building environments (Figure 8d).

The radar plots in Figure 9a–d also illustrate the comparative performance of NN, RL, and DT in predicting GHG emissions for both residential and commercial buildings during training and testing phases. In the residential training phase (Figure 9a), NN demonstrated a clear advantage, with a sharply extended R² spike nearing the upper bound of the radar chart. The compressed shapes of its RMSE and MAE zones also indicate minimal prediction errors. RL closely followed, displaying a similar performance footprint but with slightly broader margins, particularly in MAPE and NRMSE. However, DT showed the most compact area on the chart, reflecting comparatively higher error values and lower explanatory power. Its limited performance underscores its challenge in modeling the nuanced emission behaviors in residential settings. For the residential testing phase (Figure 9b), NN maintained its strong performance, retaining a high R² value and low deviation across all error metrics. This suggests that the model did not overfit and handled unseen data effectively. RL continued to perform reliably, though the radar shape slightly expanded in the error regions, hinting at greater variability when exposed to new conditions. DT’s spread widened further, particularly in MAPE, highlighting its struggle with generalization and finer emission estimation under test conditions.

Turning to commercial buildings, the training phase (Figure 9c) showed tighter and more favorable distributions across all models, with NN once again taking the lead. The nearly maximal R² and compressed error bands across RMSE and MAE emphasize its suitability for structured and steady operational profiles. RL trailed slightly but still exhibited commendable predictive consistency. Notably, DT fared better here than in residential training, likely due to the regular energy-use patterns in commercial environments, yet it still fell short of the performance levels achieved by the other two models. In the commercial test phase (Figure 9d), NN preserved its robustness, with minimal divergence from its training performance. The slight bulging in the error regions compared to training indicates a marginal dip in precision, but its overall accuracy remained high. RL kept pace, showing only moderate increases in RMSE and MAPE, suggesting good adaptability to operational dynamics. DT, however, saw an evident drop in prediction fidelity, with greater dispersion across all error indicators. This reinforces the notion that while DT can track basic trends, it lacks the granularity needed for accurate emission forecasting in dynamic, real-world commercial settings.

4.3. Comparative Validation of the Proposed Framework

To evaluate the effectiveness and practical superiority of the proposed framework, a comparison was conducted using EnergyPlus software results (

Table 7. Comparison of model performance between the EnergyPlus simulation and the proposed framework for commercial and residential building in Najran, Saudi Arabia.

Variable	EnergyPlus		Proposed Framework
	Commercial	Residential	Commercial	Residential
Energy consumption (kWh/m2/year)	4782	3684	4691	3602
CO2 emissions (kg/m2/year)	3326	2021	3308	1984
Iteration time per cycle (min)	39.5	45	16.7	19.2

). Each model was evaluated based on energy, annual CO₂ emissions, and the iteration time per optimization loop. The results clearly demonstrate that the proposed framework significantly improves energy prediction accuracy and CO₂ emission.

4.4. Impact of Hybrid BIM–PMBOK–ML Framework on Building Efficiency and Sustainability

This subsection evaluates the influence of the hybrid BIM–PMBOK–ML framework on energy savings and GHG emissions reductions across residential and commercial buildings, based on simulation outputs (Figure 10 and Figure 11).

Figure 10 illustrates seasonal energy savings (%) achieved through optimized design strategies, including high-efficiency insulation, advanced HVAC systems, and strategic window placement. These configurations were generated and refined through BIM simulations and enhanced via ML-driven predictive adjustments.

The highest energy savings were observed in spring for commercial buildings (15%) and fall for residential complexes (13%), highlighting the framework’s ability to adapt design and operational strategies to seasonal demands. In summer, energy savings reached 12% in commercial and 10% in residential buildings, emphasizing the importance of cooling system efficiency in Riyadh’s hot climate. Winter savings were slightly lower but still meaningful, with 10% for commercial and 8% for residential buildings.

Figure 11 presents GHG emissions reductions attributed to each ML model integrated into the BIM–PMBOK framework. The NN model achieved the highest reduction—15% for commercial buildings and 12% for residential complexes—due to its superior ability to model complex, non-linear relationships between building features and energy use. The RL model also performed well, contributing to 11% and 9% reductions in commercial and residential contexts, respectively, by optimizing HVAC operations in real-time. The DT model resulted in lower reductions (7% commercial, 5% residential), reflecting its limited capability in handling complex dynamic inputs.

The results confirm that ML models (especially NN and RL) enhanced the BIM-based simulations, enabling precise emissions forecasting and control. The role of PMBOK was pivotal in translating these technical enhancements into measurable sustainability KPIs, including GHG emission reductions (as shown in Figure 11), ensuring performance tracking and alignment with project goals throughout the building lifecycle.

4.5. Sensitivity Analysis and Key Parameters

This subsection evaluates the influence of key design and operational parameters on energy consumption and GHG emissions, emphasizing their role in the sustainability performance of buildings under Riyadh’s extreme climate conditions. The sensitivity analysis (Figure 12) identifies the relative impact of four primary parameters: HVAC system optimization, building orientation, insulation type, and window placement.

The results reveal that HVAC systems have the most substantial influence, contributing to over 60% of the combined impact, with approximately 55% attributed to GHG reduction and 7–8% to energy savings. This dominant effect is due to the combination of advanced ML-driven predictive control strategies that dynamically regulate heating, cooling, and ventilation based on real-time occupancy and external conditions. The suggested framework utilizes the hybrid BIM–PMBOK–ML to facilitate automated, high-efficiency operations. Building orientation ranks second, accounting for nearly 28% of the total impact (approximately 24% GHG and 4% energy). The strategic placement of buildings within the simulation environment facilitated passive solar gain during winter and minimized overheating in summer, thereby reducing reliance on mechanical systems. The insulation type and window placement contributed to a lesser extent—13% and 9%, respectively—though were still meaningful.

This analysis validates that the hybrid BIM–PMBOK–ML framework successfully prioritizes and optimizes high-impact parameters, particularly HVAC system performance.

5. Conclusions

This study proposed and validated a simulation-driven framework that combines BIM, PMBOK, and ML to enhance energy efficiency and reduce GHG emissions in the construction sector of Riyadh, Saudi Arabia. Following the Saudi Arabia’s Vision 2030 sustainability agenda, the framework demonstrates the synergistic potential of digital design tools, project-management methodologies, and data-driven analytics in delivering environmentally responsible buildings.

The use of two representative case study (commercial and residential building) simulation results confirmed the framework’s ability to adapt to building typologies and climatic variations. The main findings are summarized as follows:

• Energy consumption was reduced by up to 15% in the commercial building (spring) and 13% in the residential building (fall), reflecting the framework’s seasonal adaptability.
• GHG emissions decreased by 25% in the commercial case and 20% in the residential case, largely due to accurate ML-based energy prediction and optimized system design.
• NN achieved the highest prediction accuracy (R² = 0.95), effectively capturing non-linear interactions between building parameters and external conditions.
• RL demonstrated robust real-time control by dynamically adjusting HVAC operations, achieving up to 15% operational energy savings during active phases.

Despite these promising outcomes, some limitations were identified. Discrepancies between simulated and actual data, due to behavioral variability and unmodeled environmental factors, remain a challenge. Additionally, the limited availability of real-world data restricts broader generalization. Future research should address the following:

• Integrating high-resolution real-time data (e.g., occupancy, sensor feedback) to enhance model accuracy;
• Expanding validation to diverse building types and climatic regions to assess scalability;
• Developing adaptive models to account for environmental uncertainties, such as extreme weather;
• Incorporating regulatory and economic constraints to improve real-world applicability.

While the present framework concentrates on the BIM, PMBOK, and machine learning for predictive energy modeling and sustainable project management, it does not incorporate real-time data streams from physical sensors or IoT-based infrastructure. IoT can provide a time feedback loop that enhances both the accuracy and responsiveness of the proposed framework. For future research, it is suggested to use IoT capabilities in real-time monitoring.