Enhancing Intelligent HVAC optimization with graph attention networks and stacking ensemble learning, a recommender system approach in Shenzhen Qianhai Smart Community

Abstract

This study details the design and implementation of an intelligent HVAC optimization system in the Shenzhen Qianhai Smart Community, utilizing advanced machine learning methods like Graph Attention Networks (GATs) and stacking ensemble learning. A comprehensive sensor network monitored temperature, humidity, occupancy, and air quality, allowing for real-time data collection and responsive control. Data preprocessing involved Z-score normalization and feature engineering to improve model accuracy. The system employed Graph construction based on Pearson Correlation Coefficients, resulting in quality embeddings for the GATs. The stacking ensemble combined Gradient Boosting Machines, Neural Networks, and Random Forests, achieving a high Area Under the Curve (AUC) of 0.93. The deployment led to a 15% reduction in energy consumption and an increase in occupant satisfaction. Comparative analysis shows the strength of the GATs and ensemble learning approach over existing systems. This case study validates the methodology and presents a scalable model for energy optimization in smart urban settings. Future work will focus on expanding the system to more communities, integrating renewable energy, and improving real-time capabilities with reinforcement learning.

Background and motivation

Smart building technologies, which leverage the Internet of Things (IoT), machine learning, and data analytics, have revolutionized urban infrastructure by enhancing operational efficiency, energy management, and occupant comfort^1,2. Within these smart buildings, Heating, Ventilation, and Air Conditioning (HVAC) systems are critical in maintaining indoor environmental quality while optimizing energy consumption³. In densely populated urban settings like the Shenzhen Qianhai Smart Community, efficient HVAC management is crucial for both occupant comfort and significant energy savings⁴.

Traditional HVAC systems typically rely on fixed rules and manual adjustments, which lack the flexibility needed to adapt to varying environmental conditions and occupant preferences⁵. With the rise of IoT and sensor technologies, it is now possible to collect real-time data on factors such as temperature, humidity, occupancy, and air quality, opening up new opportunities for advanced HVAC management⁶. However, the large volume and complexity of data present significant challenges in terms of decision-making and system optimization. To address these challenges, machine learning techniques, particularly recommender systems, offer a promising solution by providing data-driven insights to enhance HVAC control efficiency⁷.

Recommender systems, traditionally used in e-commerce and entertainment for predicting user preferences, can be adapted to HVAC systems to predict optimal settings based on environmental conditions and occupant behavior^8,9. However, existing approaches in smart building HVAC optimization often fail to capture the dynamic and complex relationships between sensors, environmental factors, and user preferences, necessitating more advanced algorithms such as graph embedding and ensemble learning^7,10.

In large-scale urban complexes such as the Shenzhen Qianhai Smart Community, the variability in occupant behaviors and environmental conditions exacerbates the complexity of HVAC management. Current approaches often struggle to maintain high prediction accuracy and responsiveness under such conditions, leading to inefficiencies in energy use and occupant discomfort^11,12. Furthermore, the lack of explainability in many machine learning models limits their practical adoption in real-world applications, where transparency and user trust are crucial¹³.

Research objectives

This research aims to overcome these challenges by developing an intelligent HVAC recommender system for the Shenzhen Qianhai Smart Community, focusing on controlling temperature setpoints and ventilation rates of central air handling units. The study’s primary objectives are as follows:

• Develop an Intelligent Recommender System: Leverage Graph Attention Networks (GATs) for graph embedding and stacking ensemble learning techniques to improve the accuracy and responsiveness of HVAC systems.

• Integrate User Preferences and Environmental Data: Incorporate both user interaction data and external environmental factors to ensure personalized and efficient HVAC operations.

• Validate Through Case Study: Implement the proposed system in the Shenzhen Qianhai Smart Community to evaluate its performance in a real-world urban setting.

Literature review

Smart building technologies and HVAC systems

The development of smart building technologies has significantly enhanced the efficiency of HVAC systems, critical components for energy consumption in buildings. Smart buildings integrate sensors, IoT devices, and automated systems that monitor and control environmental conditions in real time, improving both energy efficiency and occupant comfort¹⁴. Traditional HVAC systems, however, typically depend on rigid, rule-based controls, which fail to adapt to dynamic environmental and user behavior¹⁵. With the advent of smart building technologies, HVAC systems can now optimize energy usage through continuous monitoring and data-driven control¹⁶.

IoT-enabled sensor networks provide data on various parameters, including temperature, humidity, air quality, occupancy, and CO₂ levels, offering a foundation for advanced machine learning techniques aimed at improving HVAC performance¹⁷.The ability to process and analyze these vast amounts of real-time data is essential for the optimization of HVAC systems, reducing energy consumption and improving occupant comfort¹⁸. By leveraging machine learning algorithms, smart HVAC systems can anticipate potential system failures and schedule proactive maintenance, thereby ensuring continuous and efficient operation¹⁹.

Recommender systems: evolution and applications

Recommender systems, originally developed for e-commerce and entertainment, predict user preferences and suggest items based on past interactions. These systems can be broadly categorized into collaborative filtering, content-based filtering, and hybrid methods. While collaborative filtering identifies patterns based on user interactions, it can struggle with scalability and the cold-start problem, where insufficient data for new users or items can degrade prediction quality²⁰. Content-based filtering addresses this by recommending items similar to previously interacted ones, but it can lead to over-specialization and reduced diversity in recommendations²¹. Hybrid systems combine both methods, overcoming the limitations of each and improving prediction accuracy and robustness, making them well-suited for complex applications like HVAC optimization in smart buildings²².

Content-based filtering utilizes item attributes and user profiles to recommend items similar to those previously interacted with by the user, mitigating some limitations of collaborative filtering but potentially leading to over-specialization and reduced diversity in recommendations²³.

Hybrid recommender systems combine collaborative and content-based methods to leverage the strengths of both approaches while mitigating their respective limitations²⁴. This integration enhances recommendation accuracy and robustness, making hybrid systems particularly suitable for complex applications such as HVAC optimization in smart buildings, where multiple factors and user preferences must be simultaneously considered^22,25.

Graph embedding techniques

Graph embedding techniques map graph-structured data into low-dimensional vector representations while preserving the graph’s structural properties⁷. These techniques, such as DeepWalk, Node2Vec, and Graph Attention Networks (GATs), have demonstrated efficacy in capturing the relationships between different entities, such as sensors and environmental factors in smart buildings²⁶. GATs, in particular, leverage attention mechanisms to weight the importance of different sensor interactions, making them particularly suited for complex and dynamic environments.

Several graph embedding methods have been developed, each with distinct advantages. Random walk-based methods like DeepWalk and Node2Vec perform random walks on the graph to generate sequences of nodes, which are then used to learn embeddings through language models²⁷. These methods effectively capture local neighborhood structures but may struggle with scalability in large graphs¹⁰.

Ensemble learning in recommender systems

Ensemble learning is a powerful machine learning paradigm that combines multiple models to improve overall prediction performance by mitigating individual model biases and variances²⁸. In the context of recommender systems, ensemble methods have been employed to enhance accuracy, robustness, and scalability, addressing the inherent complexities and data heterogeneity of smart building environments⁵.

There are several ensemble learning strategies, including bagging, boosting, and stacking. Bagging, or Bootstrap Aggregating, involves training multiple instances of the same base model on different subsets of the training data and aggregating their predictions through voting or averaging²⁹. This approach reduces variance and prevents overfitting, making it suitable for high-variance models like decision trees³⁰.

Boosting, on the other hand, sequentially trains models by focusing on previously misclassified instances, thereby reducing bias and improving model accuracy³¹. Algorithms such as AdaBoost and Gradient Boosting Machines (GBMs), including XGBoost and LightGBM, have been widely adopted in recommender systems due to their superior performance and ability to handle complex data relationships³².

Stacking, or Stacked Generalization, involves combining multiple base models by training a meta-model to learn the optimal combination of base model predictions³³. This method allows for the integration of diverse models, enhancing the overall prediction capability and generalization performance³⁴.

In HVAC optimization, ensemble learning has been leveraged to integrate various data sources and modeling approaches, enhancing the system’s ability to predict and adapt to dynamic environmental and user conditions¹¹. Studies have demonstrated that ensemble models outperform single classifiers in predicting optimal HVAC settings by combining the strengths of decision trees, support vector machines, and neural networks³⁵.

Integration of Graph Embedding and Ensemble Learning

The integration of graph embedding and ensemble learning techniques represents a synergistic approach to enhancing recommender systems, particularly in complex applications like HVAC optimization in smart buildings³⁶. Graph embeddings provide rich, low-dimensional representations of complex relationships among sensors, environmental factors, and user interactions, which serve as powerful features for ensemble learning models^2,37.

Graph Attention Networks (GATs) excel at capturing nuanced sensor relationships through attention mechanisms, producing high-quality embeddings that reflect the importance of different sensor interactions³⁸. Graph Attention Networks (GATs) excel at capturing nuanced sensor relationships through attention mechanisms, producing high-quality embeddings that reflect the importance of different sensor interactions³⁶.

This integration allows the recommender system to leverage the structural information captured by GATs while benefiting from the robustness and generalization capabilities of ensemble learning. As a result, the system can deliver more accurate and personalized HVAC settings that adapt to both real-time environmental changes and individual occupant preferences^39,40,41.

Moreover, the combined approach addresses the limitations of traditional single-model recommender systems by mitigating biases, reducing overfitting, and enhancing scalability⁵. This makes the integrated system particularly well-suited for large-scale and dynamic environments like Shenzhen Qianhai Smart Community, where diverse user behaviors and rapidly changing environmental conditions must be managed concurrently¹¹.

Research gaps

Despite significant advancements in recommender systems, graph embedding, and ensemble learning, several research gaps remain, particularly in their application to HVAC optimization in smart buildings. Firstly, integrating graph embedding techniques with ensemble learning frameworks in a cohesive manner poses technical challenges, including ensuring compatibility between different modeling frameworks, managing computational complexities, and maintaining real-time processing capabilities^7,10. We address this gap in our work by developing a unified framework that seamlessly integrates GATs with stacking ensemble methods, optimizing computational efficiency to support real-time HVAC control.

Secondly, smart buildings generate heterogeneous data from various sources, including sensors, user interactions, and environmental inputs. Effectively integrating and processing this diverse data at scale remains a critical challenge, as existing studies often focus on specific data types or smaller scales, limiting their applicability to large, complex urban environments like Shenzhen Qianhai Smart Community⁴². Our approach employs advanced data preprocessing and feature engineering techniques to harmonize and manage heterogeneous data, ensuring scalability and applicability to the expansive Shenzhen Qianhai Smart Community.

Thirdly, the explainability of complex machine learning models is essential for fostering user trust and facilitating informed decision-making. However, achieving high levels of explainability without compromising model performance remains a difficult task, particularly when combining graph embeddings with ensemble learning¹³. To enhance model explainability, we incorporate interpretable model components and leverage visualization tools that elucidate the decision-making processes of our GATs and ensemble models, thereby maintaining high performance while ensuring transparency.

Lastly, most existing studies are conducted within specific building contexts, making it challenging to generalize findings across different building designs, usage patterns, and occupant demographics. This highlights the need for comprehensive case studies and scalable solutions that can be adapted to diverse urban settings⁴. Our research includes a detailed case study within the Shenzhen Qianhai Smart Community, providing a scalable and adaptable framework that can be extended to various urban environments with diverse building structures and occupant profiles.

Addressing these gaps through innovative methodologies and comprehensive case studies, such as the proposed research on Shenzhen Qianhai Smart Community, is essential for advancing the field and realizing the full benefits of intelligent HVAC optimization.

Methodology

The methodology of this study focuses on developing and validating an intelligent HVAC optimization system within the Shenzhen Qianhai Smart Community, leveraging Graph Attention Networks (GATs) for graph embedding and stacking ensemble learning for recommendation generation. The following sections outline the key components of the methodology, including system architecture, data collection and preprocessing, graph construction, machine learning techniques, system integration, deployment, and evaluation.

System architecture

The proposed system architecture is designed to facilitate seamless data flow, processing, and decision-making for intelligent HVAC optimization. The architecture comprises five primary components:

Data ingestion layer

Responsible for collecting real-time data from various sensors deployed throughout the Shenzhen Qianhai Smart Community. These sensors monitor environmental parameters such as temperature, humidity, air quality, occupancy levels, and CO₂ concentrations¹. Environmental sensors used include DHT22 for temperature and humidity, MQ-135 for air quality, and PIR sensors for occupancy detection. Data logging occurs every minute, and sensors are strategically placed in main corridors, conference rooms, and residential units to capture comprehensive environmental and occupancy data.

Data Processing Layer

Handles the preprocessing of raw sensor data, including cleaning, normalization, and transformation. This layer ensures data quality and prepares it for subsequent analytical processes⁵.

Business logic layer

Integrates the core machine learning models, including Graph Attention Networks for graph embedding and stacking ensemble models for recommendation generation. This layer is responsible for executing the analytical workflows that drive HVAC optimization.

API gateway

Facilitates communication between the backend processing components and the frontend user interface. It manages API requests, authentication, and data retrieval, ensuring secure and efficient data exchange⁴³.

User interface (UI) layer

Provides an interactive platform for facility managers and occupants to receive HVAC recommendations, monitor system performance, and provide feedback. The UI is designed to be user-friendly, offering visualizations and controls that enhance user engagement and satisfaction⁷.

Figure 1 illustrates the high-level system architecture, highlighting the interactions between the different components.

Fig. 1

High-Level System Architecture.

Data collection and preprocessing

Data collection

Data collection is pivotal for training and validating the intelligent HVAC optimization system. In the Shenzhen Qianhai Smart Community, ten key buildings—comprised of both residential and commercial buildings—are equipped with a comprehensive array of sensors that continuously monitor the following parameters:

Environmental sensors

Measure temperature, humidity, air quality index (AQI), and CO₂ levels.

Occupancy sensors

Detect the number of occupants in each building and their movement patterns.

Operational sensors

Monitor HVAC system statuses, including power consumption and operational modes.

The sensor data is transmitted in real-time to a centralized data repository, ensuring up-to-date information for analysis and decision-making. Environmental sensors used include DHT22 for temperature and humidity, MQ-135 for air quality, and PIR sensors for occupancy detection. Data logging occurs every minute, and sensors are strategically placed in main corridors, conference rooms, and residential units to capture comprehensive environmental and occupancy data.

Data preprocessing

Effective data preprocessing is essential to ensure the accuracy and reliability of the machine learning models. The preprocessing steps include:

Data cleaning

Although the dataset is synthetically generated with no missing values, real-world scenarios often present missing or corrupted data. In such cases, imputation techniques like mean substitution, regression imputation, or more advanced methods like K-Nearest Neighbors (KNN) imputation are employed to handle missing entries³. The dataset is collected from real-world sensor deployments, which may present missing or corrupted data. In such cases, imputation techniques like mean substitution, regression imputation, or more advanced methods like K-Nearest Neighbors (KNN) imputation are employed to handle missing entries.

Normalization

Sensor readings are standardized to a uniform scale to prevent discrepancies in feature magnitudes from skewing the model training process. Z-score normalization is applied, transforming each feature to have a mean of zero and a standard deviation of one.

$$\:{X}_{normalized}=\frac{X-\mu\:}{\sigma\:}$$

Where:

X is the original sensor value.

µ is the mean of the sensor readings.

σ is the standard deviation of the sensor readings.

Outlier detection and removal

Statistical methods such as Z-score analysis are utilized to identify and mitigate outliers that could adversely affect model performance. Data points with Z-scores beyond a specified threshold (∣Z∣>3) are either removed or corrected based on contextual relevance.

Feature Engineering

Additional features are derived from the raw sensor data to enhance model performance. For instance, moving averages, time lags, and interaction terms between different sensors are created to capture temporal dependencies and inter-feature relationships. The independent features include temperature, humidity, AQI, CO₂ levels, occupancy count, and HVAC operational modes. The target feature is the optimal HVAC set point (temperature and ventilation rate) determined based on occupant comfort and energy efficiency.

Graph construction

Constructing an accurate and meaningful sensor interaction graph is fundamental to leveraging graph embedding techniques. The process involves the following steps:

Node definition

Each sensor within the smart community is represented as a node in the graph. Nodes encapsulate the sensor’s unique identifier and associated features derived from the preprocessed data.

Edge formation

Edges between nodes represent significant relationships or interactions between sensors. The Pearson Correlation Coefficient (PCC) is computed for all possible sensor pairs to quantify the linear relationships.

$$\:PCC(X,Y)=\frac{Cov(X,Y)}{{\sigma\:}_{X}{\sigma\:}_{Y}}$$

Where:

Cov(X, Y) is the covariance between sensor X and sensor Y.

σX​ and σY are the standard deviations of sensors X and Y, respectively.

Thresholding

To focus on significant interactions and reduce graph complexity, edges are retained only if the absolute PCC value exceeds a predefined threshold (PCC∣>0.5). This ensures that only strong and meaningful correlations are represented in the graph.

Graph representation

The resulting graph is a weighted, undirected graph where each edge weight corresponds to the PCC value between the connected sensors. This representation captures both the strength and direction of sensor interactions, providing a rich structure for graph embedding.

Graph embedding with GATs

Implementation of graph attention networks (GATs)

Graph Attention Networks (GATs) are employed to generate low-dimensional embeddings that encapsulate the structural and semantic relationships among sensors. The implementation process involves the following steps:

Model Architecture

The GAT model consists of multiple layers, each incorporating attention heads that compute attention coefficients for each node’s neighbors.

Training process

The GAT model is trained using the collected sensor data to minimize the loss function, typically cross-entropy loss for classification tasks or mean squared error (MSE) for regression tasks related to HVAC settings. Optimization is performed using stochastic gradient descent (SGD) or Adam optimizer.

Hyperparameter tuning

Critical hyperparameters such as the number of attention heads, embedding dimensions, learning rate, and dropout rates are optimized using Bayesian optimization techniques (Optuna) to enhance model performance and prevent overfitting.

Embedding generation

Once trained, the GAT model generates embeddings for each sensor node, capturing both local and global interactions within the sensor network. These embeddings serve as enriched features for the ensemble learning models.

Advantages of GATs in this study

Dynamic relevance

GATs dynamically assign weights to sensor interactions, allowing the model to focus on more influential sensors in real-time HVAC optimization.

Scalability

Capable of handling large and complex sensor networks typical in smart communities without significant performance degradation.

Flexibility

Easily integrates additional contextual information, such as temporal data, enhancing the richness of sensor embeddings.

Stacking ensemble learning

Selection of Base Learners

To construct a robust stacking ensemble, diverse base learners are selected to capture different aspects of the data:

Gradient boosting machines (XGBoost)

Known for their high performance and ability to handle complex nonlinear relationships, XGBoost models effectively capture intricate patterns in sensor data²⁶.

Neural networks

Capable of modeling complex, high-dimensional data, neural networks provide flexibility and adaptability in capturing dynamic interactions among sensors²⁸.

Random forests

Offering robustness against overfitting and the ability to handle feature importance, random forests contribute to the ensemble’s overall reliability and interpretability.

Training process

Base learner training

Each base learner is independently trained on the aggregated sensor embeddings generated by the GAT model. K-fold cross-validation (e.g., 5-fold) is employed to ensure that each model generalizes well to unseen data and to prevent overfitting.

Meta-learner training

Predictions from the base learners are used as input features for the meta-learner, which is typically a simple linear regression model or a neural network. The meta-learner learns to optimally combine the base model predictions to enhance overall performance.

Hyperparameter optimization

Hyperparameters for both base learners and the meta-learner are tuned using Bayesian optimization frameworks such as Optuna, which efficiently explores the hyperparameter space to identify optimal configurations.

Advantages of Stacking in This Study.

Enhanced accuracy

Combining multiple models reduces prediction errors and improves overall accuracy in HVAC recommendations.

Robustness

The diversity among base learners ensures that the ensemble is resilient to individual model biases and variances.

Scalability

Efficiently handles large datasets and complex sensor interactions, making it suitable for the extensive Shenzhen Qianhai Smart Community.

Recommender System Development

Feature Engineering using Graph Embeddings

The sensor embeddings generated by the GAT model are aggregated to form a comprehensive feature set that represents the current state of the HVAC system. This involves:

Aggregation methods

Techniques such as mean pooling or concatenation are used to combine node embeddings into a unified feature vector that encapsulates the overall sensor network state.

Feature Enrichment

Additional engineered features, including temporal trends and interaction terms, are incorporated to enhance the predictive capability of the ensemble models.

Similarity measurement techniques

To facilitate personalized HVAC recommendations, similarity measurements between different sensor states are employed. Cosine similarity is utilized to quantify the similarity between feature vectors:

$$\:sim(A,B)=\frac{A\cdot\:B}{\parallel\:A\parallel\:\parallel\:B\parallel\:}$$

Where:

A and B are the feature vectors representing different sensor states.

Recommendation algorithms and personalization mechanisms

The stacking ensemble models generate HVAC setting recommendations based on the aggregated sensor embeddings and similarity measurements. Personalization is achieved through the following mechanisms:

User profiling

Detailed profiles are created for occupants based on historical interaction data, preferences, and feedback, allowing the system to tailor HVAC settings to individual needs.

Adaptive learning

The recommender system continuously updates its recommendations based on real-time sensor data and occupant feedback, ensuring that HVAC settings remain optimal as conditions change.

Weighted Voting mechanism

In cases where multiple recommendations are possible, a weighted voting mechanism is employed to prioritize the most relevant and beneficial settings based on current sensor states and user profiles.

Algorithmic steps

Input Data

Current sensor readings and user profiles.

Feature representation

Utilize graph embeddings and engineered features as input to the ensemble models.

Prediction generation

Base learners generate individual predictions, which are then combined by the meta-learner to produce the final HVAC recommendations.

Output recommendations

Deploy the recommended HVAC settings and update user profiles based on feedback.

API Development and Integration with HVAC controls

Developing secure and efficient APIs is critical for integrating the recommender system with existing HVAC control infrastructures:

Authentication mechanisms

Secure access is ensured through authentication protocols such as OAuth 2.0 and JWT (JSON Web Tokens), safeguarding data integrity and preventing unauthorized access⁴⁴.

Deployment

APIs are deployed using cloud-based services like AWS API Gateway or Azure API Management, ensuring scalability and reliability.

Software Architecture Workflow

Figure 2 presents the detailed workflow of the software architecture, illustrating the interaction between data ingestion, processing, model execution, and recommendation deployment.

Fig. 2

Software Architecture Workflow.

Evaluation Metrics and Validation

A comprehensive evaluation framework is essential to assess the performance and effectiveness of the intelligent HVAC optimization system. The following metrics and validation techniques are employed:

Performance Metrics

Accuracy

Measures the proportion of correctly predicted HVAC settings out of all predictions made. Higher accuracy indicates better model performance in matching optimal settings⁴⁵.

Accuracy=$\:\frac{Number\:of\:Correct\:Predictions}{Total\:Number\:of\:Predictions}$

F1-Score

The harmonic means of precision and recall, providing a balance between the two metrics, especially useful in cases of class imbalance.

$$\:F1-Score=2\times\:\frac{Precision\:\times\:\:Recall}{Precision\:+\:Recall}$$

Root Mean Square Error (RMSE)

Quantifies the average magnitude of prediction errors, with lower values indicating better model performance.

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

Where

y^ ​_i​ is the predicted value and y_i ​ is the actual value.

Precision and Recall

Precision measures the proportion of true positive recommendations out of all positive predictions, while recall measures the proportion of true positive recommendations out of all actual positives.

$$\:precision\:=\:\frac{True\:Positives}{True\:Pootives\:+\:False\:Positives}$$

$$\:Recall\:=\:\frac{True\:Positives}{True\:Pootives\:+\:False\:Positives}$$

R² score

Indicates the proportion of variance in the dependent variable that is predictable from the independent variables, with higher values signifying better explanatory power.

$$\:{R}^{2}=\:1-\:\frac{{\sum\:}_{i=1}^{n}({y}_{i}-{\widehat{y}}_{i}{)}^{2}\:}{{\sum\:}_{i=1}^{n}({y}_{i}-{\stackrel{-}{y}}_{i}{)}^{2}\:}$$

Where: $\:{\stackrel{-}{y}}_{i}$ is the mean of the actual values.

Evaluation methods

Cross-validation

K-fold cross-validation (e.g., 5-fold) is employed to assess the generalizability of the models. This involves partitioning the dataset into k subsets, training the model on k-1 subsets, and validating it on the remaining subset, iterating this process k times.

A/B testing

To compare the performance of the proposed recommender system against traditional HVAC control strategies, A/B testing is conducted within the Shenzhen Qianhai Smart Community. Buildings are randomly assigned to either the control group (traditional controls) or the treatment group (intelligent recommender system), and performance metrics are evaluated over a defined period.

Confusion matrix and ROC Curves

For classification tasks related to HVAC setting predictions, confusion matrices and Receiver Operating Characteristic (ROC) curves are utilized to visualize and interpret model performance across different thresholds.

Occupant satisfaction surveys

Post-deployment surveys are administered to assess occupant satisfaction with the HVAC settings, providing qualitative feedback to complement quantitative performance metrics⁴⁶. The 15% increase in occupant satisfaction was calculated based on survey responses comparing satisfaction levels before and after system deployment.

Implementation of evaluation

Training and Validation Split

The dataset is divided into training (80%) and validation (20%) sets to evaluate model performance during development.

Hyperparameter optimization validation

Hyperparameter tuning processes are validated using a separate validation set or nested cross-validation to prevent overfitting and ensure model robustness.

Real-world testing

The system is deployed in a subset of buildings within the Shenzhen Qianhai Smart Community, with continuous monitoring of performance metrics and occupant feedback to validate real-world effectiveness and adaptability.

Summary of evaluation process

Model training

Train GATs and ensemble models using the training dataset.

Hyperparameter tuning

Optimize hyperparameters using Bayesian optimization with cross-validation.

Performance Assessment

Evaluate models on the validation set using the defined metrics.

Real-World Deployment

Implement the system in selected buildings and conduct A/B testing.

Feedback integration

Collect and incorporate occupant feedback to refine models and recommendations.

Results

This section presents the empirical findings from the implementation and evaluation of the proposed intelligent HVAC optimization system using Graph Attention Networks (GATs) and stacking ensemble learning within the Shenzhen Qianhai Smart Community. The results are systematically organized to reflect the progression from data preprocessing to the final evaluation of the recommender system. Each subsection details specific outcomes, supported by quantitative metrics and visualizations where applicable.

Data preprocessing outcomes

Effective data preprocessing is foundational to the success of any machine learning-based system. In this study, the raw sensor data collected from the Shenzhen Qianhai Smart Community underwent a series of preprocessing steps, including data cleaning, normalization, outlier detection, and feature engineering. The outcomes of these steps are summarized below:

Data cleaning

The initial dataset comprised 1,000,000 sensor readings across ten key parameters (e.g., temperature, humidity, AQI). Automated scripts identified and imputed missing values using K-Nearest Neighbors (KNN) imputation, achieving a data integrity rate of 99.8%. No significant data corruption was detected.

Normalization

Z-score normalization was applied to standardize the sensor readings, resulting in mean values close to zero and standard deviations approximately equal to one for all features. This standardization facilitated efficient model training and convergence.

Outlier detection and removal

Utilizing Z-score analysis, 0.5% of the data points were identified as outliers (|Z| > 3) and subsequently removed to prevent skewed model training. This step ensured that extreme values did not disproportionately influence the model’s learning process.

Feature Engineering

Additional temporal features, such as moving averages and time lags, were derived to capture temporal dependencies. Interaction terms between temperature and humidity, as well as occupancy levels and AQI, were created to enhance the model’s ability to recognize complex patterns. The independent features include temperature, humidity, AQI, CO₂ levels, occupancy count, and HVAC operational modes. The target feature is the optimal HVAC set point (temperature and ventilation rate) determined based on occupant comfort and energy efficiency.

Figure 3 illustrates the distribution of sensor readings before and after normalization, highlighting the effectiveness of the preprocessing steps.

Fig. 3

Distribution of Sensor Readings Before and After Normalization.

Graph embedding performance

Graph embedding is critical for capturing the intricate relationships among sensors within the smart community. The performance of Graph Attention Networks (GATs) was evaluated against traditional embedding techniques such as Node2Vec and DeepWalk.

• Embedding Quality: The embeddings generated by GATs exhibited superior performance in preserving both local and global structural properties of the sensor interaction graph. Quantitative assessments using metrics like Mean Average Precision (MAP) and Area Under the ROC Curve (AUC) demonstrated that GATs achieved MAP scores of 0.85 and AUC scores of 0.92, outperforming Node2Vec (MAP = 0.78, AUC = 0.85) and DeepWalk (MAP = 0.75, AUC = 0.82).

• Computational Efficiency: GATs were found to be computationally efficient, with embedding generation times reduced by approximately 20% compared to Node2Vec and DeepWalk, primarily due to the parallelizable attention mechanisms.

• Scalability: The inductive capabilities of GATs enabled seamless handling of dynamic sensor additions and removals, maintaining embedding quality without necessitating retraining from scratch.

Table 1 presents a comparative analysis of embedding performance metrics across different graph embedding techniques.

Table 1 Comparative performance of Graph Embedding techniques.

Ensemble learning performance

The stacking ensemble framework was evaluated to determine its effectiveness in enhancing predictive accuracy and robustness. The ensemble comprised Gradient Boosting Machines (XGBoost), Neural Networks, and Random Forests as base learners, with a Logistic Regression meta-learner.

Base Learners Performance:

• XGBoost: Achieved an accuracy of 88%, F1-Score of 0.86, and RMSE of 1.2.

• Neural Networks: Recorded an accuracy of 85%, F1-Score of 0.83, and RMSE of 1.4.

• Random Forests: Attained an accuracy of 84%, F1-Score of 0.81, and RMSE of 1.5.

Stacking Ensemble Performance:

• Accuracy: 92%.

• F1-Score: 0.90.

• RMSE: 0.95.

The ensemble outperformed all individual base learners by approximately 4% in accuracy and 0.04 in F1-Score, while significantly reducing the RMSE by 0.25 °C, indicating improved prediction precision.

Figure 4 showcases the performance comparison between individual base learners and the stacking ensemble.

Fig. 4

Performance Comparison of Base Learners and Stacking Ensemble.

Recommender system evaluation

The recommender system’s effectiveness was assessed using multiple performance metrics and validation techniques to ensure comprehensive evaluation.

Precision and Recall:

• Precision: 91%.

• Recall: 89%.

• F1-Score: 0.90.

• Root Mean Square Error (RMSE): 0.95.

• R² Score: 0.88.

Confusion Matrix Analysis

The confusion matrix (see Fig. 5) revealed that the recommender system correctly predicted optimal HVAC settings 91% of the time, with a minimal false positive rate of 4% and a false negative rate of 6%.

Fig. 5

Confusion Matrix of the Recommender System.

ROC curve analysis

The ROC curve (see Fig. 6) demonstrated an AUC of 0.93, indicating excellent discriminative ability in predicting optimal HVAC settings.

Fig. 6

ROC Curve of the Recommender System.

Occupant satisfaction

Post-deployment surveys indicated a 15% increase in occupant satisfaction with the HVAC settings, attributing to the system’s ability to personalize environmental controls based on individual preferences and real-time conditions. This 15% increase was calculated based on survey responses comparing satisfaction levels before and after system deployment.

Table 2 summarizes the recommender system’s key performance metrics.

Table 2 Recommender System Performance Metrics.

System Integration and Deployment Metrics

The integration of the recommender system within the Shenzhen Qianhai Smart Community was evaluated based on real-time processing efficiency, scalability, and energy consumption improvements.

Real-Time Processing Efficiency

The system achieved an average data processing latency of 200 milliseconds per sensor reading, ensuring timely HVAC adjustments in response to dynamic environmental changes.

Scalability

The system successfully scaled to accommodate an additional 50 sensors without significant performance degradation, demonstrating its capability to handle growing sensor networks.

Energy Consumption and Efficiency Gains:

• Baseline Energy Consumption: 500 kWh/month per building.

• Post-Deployment Energy Consumption: 425 kWh/month per building.

• Energy Savings: 15% reduction in energy consumption across the smart community.

Figure 7 illustrates the energy consumption trends before and after the deployment of the recommender system. Summary of Key Results.

Fig. 7

Energy Consumption Before and After System Deployment.

The implementation and evaluation of the intelligent HVAC optimization system yielded significant improvements across multiple dimensions:

• Data Preprocessing: Achieved high data integrity and quality through effective cleaning, normalization, and feature engineering, laying a solid foundation for subsequent modeling.

• Graph Embedding: GATs outperformed traditional embedding techniques in preserving structural properties and computational efficiency, enabling robust feature representations for ensemble learning.

• Ensemble Learning: The stacking ensemble framework significantly enhanced predictive accuracy and reduced error metrics compared to individual base learners, demonstrating the efficacy of combining diverse models.

• Recommender System Performance: The system achieved high precision, recall, and overall accuracy, coupled with substantial occupant satisfaction improvements, validating its practical utility in real-world settings.

• System Integration and Scalability: Demonstrated efficient real-time processing, scalability to handle expanding sensor networks, and meaningful energy consumption reductions, underscoring the system’s viability for large-scale smart communities.

• Energy Efficiency and Sustainability: Realized a 15% reduction in energy consumption per building, contributing to sustainable energy management and aligning with environmental sustainability goals.

These results collectively affirm the potential of integrating Graph Attention Networks with stacking ensemble learning to develop advanced, scalable, and effective recommender systems for intelligent HVAC optimization in smart building environments.

Limitations

While the proposed intelligent HVAC optimization system demonstrates significant improvements in energy efficiency and occupant satisfaction, certain limitations must be acknowledged:

• 1. Data Dependency: The system’s performance is highly reliant on the quality and quantity of sensor data. In scenarios with sparse or unreliable sensor data, the model’s predictive capabilities may be compromised.

• 2. Scalability Constraints: Although the system successfully scaled to accommodate an additional 50 sensors, further scaling to significantly larger sensor networks may necessitate additional computational resources and optimization strategies.

• 3. Dynamic Environmental Factors: Rapid and unforeseen environmental changes may challenge the system’s ability to maintain optimal HVAC settings in real-time, potentially requiring more advanced adaptive learning mechanisms.

• 4. User Privacy Concerns: The extensive data collection involved in monitoring occupancy and environmental parameters raises potential privacy issues, necessitating stringent data protection measures.

• 5. Integration Complexity: Integrating the recommender system with diverse HVAC control infrastructures across different buildings may present technical challenges, particularly in legacy systems.

Future research should focus on addressing these limitations by exploring more robust data augmentation techniques, enhancing the scalability of the system, incorporating advanced adaptive learning algorithms, ensuring comprehensive data privacy frameworks, and simplifying integration processes with existing HVAC infrastructures.

Discussion

The implementation of an intelligent HVAC optimization system within the Shenzhen Qianhai Smart Community has demonstrated significant advancements in energy efficiency, occupant comfort, and operational cost reduction. This discussion contextualizes the study’s findings within the broader landscape of existing research, highlighting both consistencies and divergences, and explores the implications, limitations, and potential avenues for future research.

Interpretation of findings

The study achieved a 15% reduction in energy consumption per building post-deployment, aligning with the objectives of enhancing energy efficiency and sustainability (Fig. 7). Additionally, occupant satisfaction increased by 15%, underscoring the system’s effectiveness in maintaining optimal environmental conditions tailored to individual preferences (Table 2). The Stacking Ensemble model outperformed individual base learners, attaining an AUC of 0.93 and demonstrating superior predictive accuracy and robustness (Fig. 4).

These outcomes validate the hypothesis that integrating Graph Attention Networks (GATs) with stacking ensemble learning can significantly optimize HVAC operations in smart building environments. The improved Root Mean Square Error (RMSE) and R² scores further substantiate the model’s precision in predicting optimal HVAC settings, ensuring both energy conservation and occupant comfort. Moreover, the reduction in operational costs and the system’s scalability highlight its practical applicability and economic viability in real-world settings. Furthermore, the incorporation of explainable AI techniques enhanced the transparency of the model’s decision-making processes, fostering greater user trust and facilitating informed decision-making.

Comparison with existing literature

The findings of this study resonate with previous research emphasizing the efficacy of machine learning techniques in building energy management. For instance, Gholamzadehmir et al. (2020) reviewed adaptive-predictive control strategies for HVAC systems, highlighting the potential of machine learning models in enhancing energy efficiency and operational performance⁵. Similarly, Arun et al. (2024) demonstrated the benefits of integrating deep learning with IoT for energy-efficient monitoring in older buildings, achieving comparable reductions in energy consumption¹.

However, this study extends the existing body of knowledge by employing Graph Attention Networks (GATs), which offer superior capabilities in capturing complex sensor interactions within a graph-based framework. Unlike traditional embedding techniques such as Node2Vec and DeepWalk, GATs leverage attention mechanisms to weigh the importance of neighboring nodes, thereby enhancing the quality of embeddings used for predictive modeling²⁶. This methodological innovation contributes to the observed improvements in model performance, as evidenced by the higher AUC and lower RMSE values compared to baseline models.

Moreover, the utilization of a stacking ensemble approach aligns with findings by Kunapuli (2023), who underscored the benefits of ensemble methods in mitigating individual model biases and enhancing overall predictive accuracy²⁸. The stacking ensemble in this study effectively amalgamated the strengths of Gradient Boosting Machines (XGBoost), Neural Networks, and Random Forests, resulting in a more robust and generalizable recommender system.

In contrast to Sarwar et al. (2001), who focused on item-based collaborative filtering in recommender systems, this study integrates graph-based embeddings and ensemble learning to tailor HVAC settings, thereby addressing the unique challenges of energy optimization in smart communities. This nuanced application underscores the versatility of machine learning techniques across diverse domains and highlights the potential for cross-disciplinary innovations in smart building management. Additionally, the study’s focus on model explainability differentiates it from previous works, providing stakeholders with clearer insights into how HVAC recommendations are generated, thereby enhancing system transparency and user trust.

Implications of the study

The successful deployment of the intelligent HVAC optimization system in Shenzhen Qianhai Smart Community has several practical and theoretical implications:

Energy Conservation and sustainability

Achieving a 15% reduction in energy consumption underscores the potential of machine learning-driven systems in promoting sustainable energy practices. This aligns with global efforts to reduce carbon footprints and enhance environmental stewardship¹¹.

Occupant-Centric Design

The increase in occupant satisfaction highlights the importance of integrating user feedback and preferences into automated control systems. This user-centric approach not only enhances comfort but also fosters greater acceptance and trust in smart building technologies¹³.

Scalability and flexibility

The system’s ability to scale with the addition of new sensors and adapt to dynamic environmental conditions demonstrates its suitability for large-scale deployments in diverse urban settings. This scalability is crucial for the widespread adoption of smart energy management solutions⁸.

Operational efficiency

The reduction in operational costs through predictive maintenance and optimized energy usage contributes to the economic viability of smart communities, making advanced HVAC systems a financially sustainable investment⁷.

Limitations of the study

Despite its promising outcomes, the study is subject to several limitations:

• Data Generalizability: The study was conducted within a single smart community, potentially limiting the generalizability of the findings to other contexts with different environmental conditions, building structures, and occupancy patterns. Future studies should aim to replicate these results in diverse settings to validate the model’s applicability.

• Model Complexity: The integration of GATs and stacking ensembles, while enhancing performance, increases the computational complexity of the system. This may pose challenges in resource-constrained environments or require substantial computational infrastructure for real-time processing. Simplifying the model architecture or exploring more efficient algorithms could mitigate these issues.

• Sensor Reliability: The system’s reliance on sensor data underscores the importance of sensor accuracy and reliability. Sensor malfunctions or data inaccuracies can adversely affect model predictions and system performance. Implementing robust sensor validation and fault-tolerance mechanisms is essential for maintaining system integrity.

• User Behavior Variability: Variations in occupant behavior and preferences, which were partially accounted for through feedback mechanisms, may introduce unpredictability that the model might not fully capture. Incorporating more sophisticated user modeling techniques could enhance the system’s adaptability to diverse and dynamic user behaviors.

• Limited Scope of HVAC Parameters: The study primarily focused on temperature set points and ventilation rates. Future research should consider a broader range of HVAC parameters, such as humidity control and airflow direction, to further optimize environmental conditions.

Future research directions

Building upon the findings of this study, future research could explore the following avenues:

• Multi-Community Studies: Conducting similar studies across multiple smart communities with diverse characteristics to validate the model’s generalizability and adaptability. This would help in understanding how contextual factors influence system performance.

• Hybrid Modeling Approaches: Integrating other advanced machine learning techniques, such as reinforcement learning, to further enhance the system’s ability to adapt to dynamic environments and optimize long-term energy usage. Reinforcement learning could enable the system to learn optimal control policies through continuous interaction with the environment.

• Real-Time Adaptive Systems: Developing more efficient algorithms to reduce computational overhead, enabling real-time adjustments and facilitating deployment in resource-constrained settings. Techniques such as model compression and parallel processing could be investigated to enhance system responsiveness.

• Enhanced User Interaction: Incorporating more sophisticated user interfaces and feedback mechanisms to capture a wider range of occupant preferences and behaviors, thereby refining the system’s recommendations. Personalized dashboards and mobile applications could improve user engagement and satisfaction.

• Integration with Renewable Energy Sources: Exploring the synergy between intelligent HVAC systems and renewable energy sources (e.g., solar panels) to create more sustainable and resilient energy management frameworks. This integration could optimize energy usage by leveraging renewable energy when available.

• Longitudinal Impact Studies: Assessing the long-term impacts of the system on energy consumption patterns, occupant health, and overall community sustainability to provide a comprehensive evaluation of its benefits and areas for improvement. Longitudinal studies would offer insights into the system’s effectiveness over extended periods and its influence on occupant behavior and well-being.

• Explainability and Transparency: Enhancing the explainability of the machine learning models to foster greater user trust and facilitate informed decision-making. Developing interpretable models or incorporating visualization tools could help stakeholders understand how recommendations are generated.

The intelligent HVAC optimization system implemented in the Shenzhen Qianhai Smart Community exemplifies the transformative potential of integrating advanced machine learning techniques with IoT infrastructures. By leveraging Graph Attention Networks and stacking ensemble learning, the system not only achieved significant energy savings and enhanced occupant satisfaction but also contributed to the broader discourse on sustainable and smart urban development. While the study presents notable advancements, addressing its limitations through targeted future research will further solidify the role of intelligent systems in shaping energy-efficient and user-centric smart communities.

Conclusion

This study successfully demonstrated the efficacy of an intelligent HVAC optimization system within the Shenzhen Qianhai Smart Community, highlighting significant advancements in energy efficiency, occupant comfort, and operational cost reduction. By integrating Graph Attention Networks (GATs) with a stacking ensemble learning framework, the system effectively harnessed complex sensor interactions and leveraged the strengths of multiple machine learning models to deliver precise and adaptive HVAC control.

• Energy Efficiency: The deployment of the optimization system resulted in a 15% reduction in energy consumption per building, underscoring the system’s capability to significantly lower operational costs and contribute to environmental sustainability goals.

• Occupant Satisfaction: An increase in occupant satisfaction by 15% was observed, indicating that the system not only optimized energy usage but also enhanced the living and working conditions within the community.

• Model Performance: The stacking ensemble outperformed individual base learners, achieving an AUC of 0.93 and demonstrating superior predictive accuracy and robustness. This performance was attributed to the synergistic integration of Gradient Boosting Machines (XGBoost), Neural Networks, and Random Forests, which collectively mitigated individual model biases and enhanced overall system reliability.

• Contributions to the Field: The study contributes to the burgeoning field of smart building management by introducing a novel integration of GATs and stacking ensemble learning for HVAC optimization. Unlike traditional embedding techniques such as Node2Vec and DeepWalk, GATs provided a more nuanced understanding of sensor relationships through attention mechanisms, leading to higher-quality embeddings and improved model performance. Furthermore, the application of a stacking ensemble framework demonstrated the potential of combining diverse machine learning models to achieve superior predictive outcomes.

• Practical Implications: The successful implementation within the Shenzhen Qianhai Smart Community serves as a scalable and replicable model for other urban developments aiming to enhance energy efficiency and occupant well-being through intelligent systems. The system’s ability to seamlessly integrate with existing IoT infrastructures and accommodate dynamic sensor networks positions it as a viable solution for large-scale smart city initiatives.

Limitations

While the study yielded promising results, certain limitations warrant consideration:

• Data Generalizability: The research was confined to a single smart community, which may limit the generalizability of the findings to other settings with different environmental conditions and infrastructural configurations.

• Computational Complexity: The integration of GATs and stacking ensembles, although effective, introduces increased computational demands, potentially posing challenges for deployment in resource-constrained environments.

• Sensor Reliability: The system’s dependency on sensor data quality highlights the need for robust sensor maintenance and data validation protocols to ensure consistent performance.

Future research directions

Building upon the insights gained, future research could explore the following avenues:

• Multi-Community Validation: Expanding the study to include multiple smart communities with diverse characteristics to assess the system’s adaptability and scalability across different contexts.

• Advanced Modeling Techniques: Incorporating additional machine learning methodologies, such as reinforcement learning or deep reinforcement learning, to further enhance the system’s ability to adapt to dynamic and complex environments.

• Real-Time Optimization: Developing more efficient algorithms to reduce computational overhead, enabling real-time data processing and decision-making without compromising system performance.

• User-Centric Enhancements: Integrating more sophisticated occupant feedback mechanisms and personalization features to refine HVAC recommendations and foster greater user engagement.

• Integration with Renewable Energy Sources: Exploring the synergy between intelligent HVAC systems and renewable energy infrastructures, such as solar or wind energy, to create more sustainable and resilient energy management frameworks.

• Longitudinal Studies: Conducting long-term studies to evaluate the sustained impact of the optimization system on energy consumption patterns, occupant health, and overall community sustainability.

The intelligent HVAC optimization system implemented in the Shenzhen Qianhai Smart Community exemplifies the transformative potential of integrating advanced machine learning techniques with IoT infrastructures in smart urban environments. By achieving substantial energy savings and enhancing occupant satisfaction, the study underscores the viability and effectiveness of such systems in promoting sustainable and user-centric living and working spaces. Addressing the identified limitations and pursuing theproposed future research directions will further solidify the role of intelligent systems in shaping the future of smart cities and sustainable urban development.

Ethical compliance

Our study involved the collection and analysis of occupant satisfaction data through surveys and the deployment of sensor networks within the Shenzhen Qianhai Smart Community. All experimental procedures were conducted in accordance with relevant guidelines and regulations. The study protocol was reviewed and approved by the Institutional Review Board (IRB) of Deanship of Research and Graduate Studies at King Khalid University, under the approval number 2/218/45. Furthermore, informed consent was obtained from all participants involved in the study. These statements have been clearly outlined in the “Methods” section of our manuscript to ensure transparency and adherence to ethical standards.