To cite this article: Samaneh Nadali and Tsi Hao Yong 2025 J. Phys.: Conf. Ser. 3140 052022
Journal of Physics: Conference Series 3140 (2025) 052022
IOP Publishing
doi:10.1088/1742-6596/3140/5/052022
Samaneh Nadali, Tsi Hao Yong
ADF Technologies Sdn.Bhd, Malaysia.
Samanehnadali@adftech.com.my, thyong@adftech.com.my
Abstract. Optimizing the differential pressure (DP) setpoint in HVAC systems is vital for achieving energy efficiency and maintaining system performance. Conventional strategies typically rely on fixed thresholds or worst-case assumptions, setting DP setpoints high enough to ensure adequate flow to the most demanding or distant air handling units (AHUs). While this guarantees system reliability, it often leads to energy overuse, reduced operational efficiency, and higher costs, particularly during periods of low demand. These static methods also fail to respond to real-time variations in occupancy, weather, and internal loads. This paper proposes a data-driven, adaptive control strategy that integrates machine learning (ML) and genetic algorithms (GA) to intelligently optimize DP setpoints in commercial buildings. The framework includes two main components: (1) an XGBRegressor model that forecasts next-day cooling load at 15-minute intervals using inputs such as weather conditions, historical energy usage, and occupancy trends; and (2) a GA that dynamically determines the most energy-efficient DP setpoints based on these forecasts. Unlike traditional approaches, this method continuously adjusts the setpoints in response to actual operating conditions, ensuring both energy savings and occupant comfort. The system also provides clear visualization tools for operators to monitor key HVAC elements, including compressors, pumps, and cooling towers. Implemented in a high-end commercial building in Kuala Lumpur, the proposed approach demonstrated up to 19% energy savings and 21% peak demand reduction. These results highlight the potential of combining ML and GA for smarter, more efficient HVAC control across various building types.
Keywords: Differential Pressure Setpoint, Energy Efficiency, HVAC System, XGBRegressor Algorithm, Genetic Algorithm
1. Introduction
Keeping commercial buildings comfortable while reducing energy use is an ongoing challenge in the building industry. HVAC systems responsible for heating, ventilation, and air conditioning are among the largest energy consumers in such buildings. A key factor in their performance is the differential pressure (DP) setpoint, which controls the flow of chilled water. If not properly managed, incorrect DP setpoints can lead to inefficiencies, higher energy costs, and greater equipment wear. Despite technological progress, many buildings still rely on fixed or manually adjusted DP settings. These static methods cannot adapt to real-time changes in weather, occupancy, or cooling demand, often resulting in wasted energy or discomfort. This highlights the need for smarter, automated systems that dynamically adjust DP setpoints in response to changing conditions.
DP control is critical for energy efficiency, occupant comfort, and system reliability. Optimizing DP in variable air volume (VAV) and chilled water systems has been widely studied. This review examines and categorizes a range of approaches from past research, offering a foundation for further advancement.
Early strategies based on static setpoints and manual tuning were found to be inefficient, as shown by studies such as Raftery et al. (2018) and Cai & Braun (2014). Model Predictive Control (MPC) gained traction as a more advanced solution, with applications including Bayesian tuning (Lu et al., 2020), occupancy-based control (Aftab et al., 2017), and predictive modelling (Hong & Kim, 2018). Reinforcement learning (Yang et al., 2015) also showed promise for adaptive control. More recently, deep learning has been applied to cooling and DP control, with studies by Merabet et al. (2021) and Chen et al. (2024) addressing issues like sensor faults and system robustness.
Optimization algorithms such as genetic algorithms (GA) and particle swarm optimization (PSO) have also been widely used. Research by Nassif et al. (2005), Seong (2019), and Kusiak et al. (2010) demonstrates how these methods enhance both efficiency and comfort. Hybrid strategies that combine model-based and learning-based approaches such as those by Ding et al. (2024) and Dowling & Zhang (2020) offer even greater adaptability and fault detection capabilities.
The growing integration of AI technologies including ML, RL, and GAs has significantly advanced DP setpoint optimization. However, many existing studies still separate prediction from control and are often limited to simulations or laboratory conditions.
This study addresses that gap by presenting a real-world hybrid framework combining machine learning and genetic algorithms for automated DP setpoint control. An XGBRegressor forecasts next-day cooling loads in 15-minute intervals using weather, historical energy, and occupancy data. A GA then optimizes DP setpoints based on these forecasts, improving energy efficiency without compromising comfort.
Implemented in a high-end commercial building in Kuala Lumpur, this approach led to up to 19% energy savings and a 21% reduction in peak demand. By integrating forecasting and optimization into a single system, this work offers a scalable, intelligent solution for smarter and more efficient HVAC operation. The following section details the methodology used.
2. Methodology
This study presents an integrated methodology for optimizing DP setpoints in HVAC systems by combining ML for predictive cooling load forecasting and a GA for optimal control. The framework consists of seven key stages: data collection, preprocessing, ML model development, model training, cooling load prediction, GA-based optimization, and control implementation.
2.1 Data collection
Raw data were collected from a commercial building’s Building Management System (BMS) as well as from external weather services. The weather data were sourced from AccuWeather [11]. All the data is provided by [12], a company specializing in smart building management solutions in Malaysia. The dataset encompasses a range of variables, including environmental conditions such as outdoor temperature and humidity. It also includes operational parameters like indoor temperature, HVAC control setpoints, and occupancy schedules. Additionally, energy-related measurements such as chilled water flow, air handling unit (AHU) fan power, and differential pressure were recorded to support detailed modeling of the building’s cooling demand. The data cover a duration of one full year and were recorded at 15-minute intervals to ensure sufficient resolution for predictive modeling and control optimization.
2.2 Data preprocessing
Data quality is essential for reliable model training, and several preprocessing steps were applied to ensure consistency and accuracy. The data was resampled to uniform 15-minute intervals to align all input sources. Missing values were imputed using forward or backward fill methods, depending on the context and data behavior. To maintain a consistent scale across features, normalization was performed. Feature engineering techniques were employed, including the creation of time-of-day and day-of-week indicators to capture temporal patterns. Finally, outlier detection and removal were conducted using z-score and interquartile range (IQR) methods to eliminate anomalous data and ensure clean inputs for modeling.
2.3 Machine learning model development
An XGBoost Regression (XGBRegressor) model was selected due to its robustness, interpretability, and strong performance on structured datasets. The model is designed to predict the building’s cooling demand 24 hours in advance, corresponding to 96 time steps at a 15-minute resolution. Input features included forecasted weather conditions (temperature and relative humidity), time-based variables such as hour of the day, weekend or weekday indicators, and holidays, as well as past energy usage patterns including cooling energy consumption and dew point trends. Additionally, occupancy-related variables and control setpoints were incorporated to enhance predictive accuracy.
2.4 Model training
The dataset was split into 80% training and 20% testing sets. The model was trained with a learning rate of 0.1, a maximum depth of 6, 200 estimators, and early stopping after 20 rounds. Training performance was evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the $R^2$ score. The trained model achieved a strong fit and was subsequently deployed for next-day cooling load forecasting.
2.5 Prediction of cooling load
Using the trained XGBoost model and forecasted inputs, the system generates 15-minute interval predictions for cooling load over the next 24 hours. These predicted loads are used as inputs for the optimization stage.
2.6. DP Setpoint optimization via genetic algorithm
The predicted cooling load is used as input to GA, which determines the optimal DP reset schedule for the same 24-hour forecasting horizon. The GA was configured with a population size of 100 and ran for 50 generations. A crossover rate of 0.8 and a mutation rate of 0.05 were applied to promote exploration and diversity within the solution space. The selection of individuals for reproduction was performed using a tournament selection method, balancing convergence speed with solution quality.
3. Control Implementation
The optimized DP reset schedule is transmitted to the building’s control system in real time. Operators can monitor key parameters through a web-based dashboard, which displays predicted versus actual cooling loads, optimized versus baseline DP values, and energy usage trends along with system alerts. This interface supports informed decision-making and enables proactive system management.
4. Result and Discussion
To evaluate the effectiveness of the proposed ML + GA-based DP reset strategy, a 12-month field study was conducted in a commercial building located in Kuala Lumpur. The system’s performance was compared against a baseline HVAC operation using a conventional fixed DP setpoint strategy. Three primary evaluation dimensions were considered: energy savings, peak demand reduction, and thermal comfort compliance. In addition to outcome metrics, this section also provides performance indicators of the machine learning model and insights into broader implications and potential scalability of the proposed system.
4.1. ML model performance evaluation
The XGBoost model, used for real-time cooling load prediction, was evaluated on a holdout test set (20% of the data). It achieved a Mean Absolute Error (MAE) of 2.13 kW, Root Mean Squared Error (RMSE) of 3.21 kW, and a coefficient of determination ($R^2$) of 0.92. These metrics confirm the model’s robustness and ability to generalize, accurately forecasting both baseline and peak demand variations critical for downstream control. The high predictive performance ensures that the GA optimization module receives reliable input, directly contributing to system efficiency. Compared to reported metrics in similar works (e.g., RMSE of ~4.5-5.2 kW in Hong & Kim, 2018; Merabet et al., 2021), this model demonstrates competitive accuracy suitable for field deployment. The high predictive performance ensures that the GA optimization module receives reliable input, directly contributing to system efficiency. Compared to reported metrics in similar works (e.g., RMSE of ~4.5-5.2 kW in Hong & Kim, 2018; Merabet et al., 2021), this model demonstrates competitive accuracy suitable for field deployment.
4.2 Monthly energy consumption comparison
The monthly energy consumption (in kWh) under both methods is shown below.
The results shown in Table 1 are based on a full 12-month real-time field deployment in a commercial building in Kuala Lumpur. The system consistently achieved monthly energy savings averaging 19.1%, enabled by dynamic optimization of the differential pressure (DP) setpoint using live cooling load forecasts and GA-based control. This real-world validation distinguishes our work from controlled test-bed evaluations and highlights its effectiveness in an actual operational environment. The consistent savings observed across all months underscore the practical viability and robustness of the proposed framework. Furthermore, these empirical results confirm the value of predictive, real-time HVAC optimization in delivering measurable energy savings while ensuring system stability and occupant comfort.
4.3 Monthly Peak Demand Comparison
The ML + GA strategy led to an average 21.0% reduction in monthly peak demand, which is especially significant in regions with demand-based electricity tariffs. This reduction was achieved by smoothing load profiles and minimizing abrupt system ramp-ups through better prediction and DP modulation. By forecasting load and adjusting setpoints ahead of time, the system flattens demand peaks that would otherwise trigger higher utility costs.
4.4 Comfort compliance
Indoor environmental quality was monitored to ensure that energy savings didn’t compromise occupant comfort.
The proposed method significantly improved thermal comfort, maintaining temperature within the desired range ($22-25^{\circ}\text{C}$) 95.6% of the time, compared to 87.2% under the fixed DP system. Moreover, the system reduced temperature fluctuation from $\pm2.5^{\circ}\text{C}$ to $\pm1.2^{\circ}\text{C}$, indicating greater stability and responsiveness. This result reflects the value of day-ahead load prediction, which allows the system to proactively adjust airflow and pressure instead of reacting to real-time deviations.
4.5 Real-Time Optimization Capability
Unlike offline or day-ahead optimization systems, the proposed framework is implemented in real time, with predictions and optimized DP schedules updated continuously based on streaming data. This real-time loop enables dynamic responses to sudden changes in occupancy, weather, or system faults. Operators can monitor and override the control logic through a web-based dashboard integrated with the building’s BMS. The advantage of real-time implementation lies in its resilience and adaptability – enabling minute-level decision-making rather than relying on pre-scheduled actions. This also aligns with emerging standards in digital twin and smart grid interaction frameworks.
4.6 Broader Implications and Scalability
Unlike offline or day-ahead optimization systems, the proposed framework is implemented in real time, with predictions and optimized DP schedules updated continuously based on streaming data. This real-time loop enables dynamic responses to sudden changes in occupancy, weather, or system faults. Operators can monitor and override the control logic through a web-based dashboard integrated with the building’s BMS. The advantage of real-time implementation lies in its resilience and adaptability – enabling minute-level decision-making rather than relying on pre-scheduled actions. This also aligns with emerging standards in digital twin and smart grid interaction frameworks.
5. Conclusion & Future Work
This study proposed an intelligent HVAC control strategy that optimizes DP reset setpoints using a hybrid of XGBRegressor and GA. By predicting cooling demand based on weather, energy usage, and occupancy data, the system dynamically adjusts DP setpoints in real time. Tested over a year in a Kuala Lumpur commercial building, it achieved 19.1% energy savings, 21% peak demand reduction, improved comfort compliance, and reduced equipment runtime. The results demonstrate the approach’s efficiency and scalability. Future work may explore integration with demand response programs, multi-objective optimization, and deployment in other building types such as hospitals, shopping malls, educational campuses, where energy-intensive HVAC operations and varying occupancy patterns make real-time optimization especially valuable.
References
- [1] Raftery, P., Li, S., Jin, B., Ting, M., Paliaga, G., & Cheng, H. (2017). Evaluation of a cost-responsive supply air temperature reset strategy. Energy and Buildings, 158, 123-136. https://doi.org/10.1016/j.enbuild.2017.09.044
- [2] Cai, J., & Braun, J. E. (2014). A heuristic supply air temperature reset strategy for VAV systems employing variable-capacity DX cooling equipment. Proceedings of the 2014 International High Performance Buildings Conference.
- [3] Taylor, S. T. (2019). VAV system duct main design. ASHRAE Journal, 61(4), 18-28.
- [4] Rahnama, S., Afshari, A., Bergsøe, N. C., Sadrizadeh, S., & Hultmark, G. (2018). Experimental study of the pressure reset control strategy for energy-efficient fan operation – Part 2: Variable air volume ventilation system with decentralized fans. Energy and Buildings, 172, 249-256. https://doi.org/10.1016/j.enbuild.2018.04.024
- [5] Zhao, T., Ma, L., & Zhang, J. (2016). An optimal differential pressure reset strategy based on the most unfavorable thermodynamic loop on-line identification for a variable water flow air conditioning system. Energy and Buildings, 128, 1-12. https://doi.org/10.1016/j.enbuild.2016.06.065
- [6] Lu, Q., Kumar, R., & Zavala, V. M. (2020). MPC controller tuning using Bayesian optimization techniques. arXiv preprint arXiv: 2009.14175. https://arxiv.org/abs/2009.14175
- [7] Aftab, M., Chen, C., Chau, C.-K., & Rahwan, T. (2017). Automatic HVAC control with real-time occupancy recognition and simulation-guided model predictive control in low-cost embedded system. Energy and Buildings, 154, 141-156.
- [8] Hong, G., & Kim, B. S. (2018). Development of a data-driven predictive model of supply air temperature in an air-handling unit for conserving energy. Energies, 11(2), 407. https://doi.org/10.3390/en11020407
- [9] Ahn, B. C., & Mitchell, J. W. (2001). Optimal control development for chilled water plants using a quadratic representation. Energy and Buildings, 33(4), 371-378. https://doi.org/10.1016/S0378-7788(00)00091-7
- [10] Yang, L., Nagy, Z., Goffin, P., & Schlueter, A. (2015). Reinforcement learning for optimal control of low exergy buildings. Applied Energy, 156, 577-586. https://doi.org/10.1016/j.apenergy.2015.07.050
- [11] AccuWeather, “Kuala Lumpur Weather Forecast,” Available: https://www.accuweather.com/en/my/kuala-lumpur/233776/weather-forecast/233776. [Accessed: 1-Jul-2025].
- [12] ADF Technologies Sdn Bhd. https://www.adftech.com.my/.
