Building Heating, Ventilation, And Air Conditioning Systems Research Articles

Fault detection research on novel transfer learning-based method for cross-condition, cross-system and cross-operation in public building HVAC sensors

Transfer learning (TL) has the inspiring potential for artificial intelligence in heating, ventilation and air conditioning (HVAC) system with insufficient data labels. However, traditional TL-based methods are limited when applied across different conditions, systems, and operations.Unfortunately, public building HVAC systems encounter challenges related to data acquisition and richness, making it difficult to obtain data from similar HVAC systems conditions, scenarios and operations. It proposes a novel TL-based method that combines energy and mass balance constraint equation (EBCe) to diagnose the sensor faults in HVAC systems across different systems, conditions and operations.Firstly, it utilizes laboratory data as the source domain data and constructes EBCe based on the common physical laws of HVAC system to reduce the data differences between laboratory and public buildings. Then, an laplacian kernel domain-adaptive neural network (LkDaNN) is proposed to generalize more efficiently feature differences between the source domain data and target domain data. Finally, experiment analyzes the non-fault and four control-sensors fault under both cross-operation and non-cross operation conditions. The experimental results demonstrate that the EBCe-LkDaNN method achieves satisfactory fault detection and diagnosis (FDD) performance.The overall FDD accuracy of porposed method can reach 90.72 % and 88.64 % under different cross-operation, respectively. Practical application of the EBCe-LkDaNN strategy for HVAC sensor FDD are discussed at last.

An ontology for automated fault detection & diagnostics of HVAC using BIM and machine learning concepts

This paper presents an ontology for AFDD (Automated Fault Detection and Diagnostics) of HVAC (Heating, Ventilation, and Air conditioning) systems in buildings called “AFDDOnto”. Presently, the AFDD models are mainly data-centric and often lack semantic information such as contextual information and spatial information; additionally, configuration information, analysis, and results used for model development are lost once developed. This impedes an effective mechanism for tracking changes and updating the model for future developments and use cases. The Proposed ontology can be used for AFDD model development, tracking changes, analytics, visualization, and digital twinning by enabling integration of BIM with BAS/BMS (Building Automation System/Building Management System) concepts and secondly to store AFDD configuration and analytics in the AFDDOnto. Select competency questions are constructed using SPARQL queries to access the proposed knowledge model. The proposed ontology has been tested against different measures using multiple metrics and a case study and further validated using a semi-structured survey of experts. Applied AI engineers, Facility managers, Asset managers, and building owners aiming to develop AFDD models for HVAC systems can benefit from adopting this ontology for HVAC maintenance, including analysis, model development, and knowledge management.

A novel rule mining method for knowledge discovery of relationships among indoor air quality, HVAC operation and occupants’ activities

This paper presents a novel data mining method to discover the occurrence time, co-occurrence and sequential association rules among indoor air quality (IAQ), Heating, Ventilation and Air Conditioning (HVAC) operation and occupants' activities. The novelty of this work includes a peak detection method to discover the pollutants' peaks and a rule mining method with extended rule selection criteria that considers both the time lags between events and relationships among multiple events. The peak detection method aimed to use moving windows and criteria to find the candidate peaks and the non-maximum suppression to filter the overlapped candidate peaks. The proposed rule mining method is based on the Apriori and Apriori-All with the rule selection criteria extended by labeling the occurrence time to events to find the rules among various occurrence times and multiple events and by limiting the time lags among the occurrences of the events to discover the sequential rules. Two datasets, one from 70 houses and one from a single commercial building, were used to evaluate the performance of this proposed data-driven method. The results showed that 93.19 % of the peak detection results matched the peaks detected from the data records. The CO2 peaked frequently at 12:00, 17:00 and from 19:00 to 24:00. The occurrence of occupants' activities could lead to pollutants’ concentrations and pollutants peaks within 2 hours in the residential buildings. The occurrence of a high Wifi connection number was frequently followed by a CO2 peak within 2 hours in the commercial building. These findings can support the development of control strategies for building HVAC systems to achieve better indoor air quality control.

Comparing the life cycle costs of a traditional and a smart HVAC control system for Australian office buildings

Many smart technologies have been introduced in buildings with the aim to reduce the energy and GHG emissions associated with their operation, particularly through improved control systems for regulating heating, ventilation and air conditioning (HVAC) equipment. Despite their energy saving potential, only a few studies have comprehensively assessed the costs associated with their practical implementation from a life cycle perspective. Accordingly, this study quantifies and compares the life cycle costs of a smart HVAC control system with that of a traditional control system, in the context of an Australian office building. For both systems, the required hardware are specified based on the characteristics of these systems and the layout of the serviced spaces in the reference building. The costs incurred over the period of assessment are quantified using the net present cost (NPC) approach. To evaluate the effects of these control systems on the operational energy costs of the building HVAC system, the control logics of both these systems are modelled through building energy simulations. The results show that, over the period of assessment, the smart control system incurred a higher total cost compared to the traditional control system. However, the findings from the simulations show that the HVAC energy cost savings achieved through the specification of the smart control system offset the additional cost incurred to deploy this system over the traditional control system. The smart control system resulted in HVAC operational cost savings between 9 % and 10 % compared to the traditional control system. Sensitivity analyses indicated that the total life cycle costs varied between −27 % and +50 %, with the discount rate and energy price increase rate being the most influential parameters.

Successful application of predictive information in deep reinforcement learning control: A case study based on an office building HVAC system

Reinforcement Learning (RL), a promising algorithm for the operational control of Heating, Ventilation, and Air Conditioning (HVAC) systems, has garnered considerable attention and applications. However, traditional RL algorithms typically do not incorporate predictive information for future scenarios, and only a limited number of studies have examined the enhancement and impact of predictive information on RL algorithms. To address the issue of coupling RL and predictive information in HVAC system operation optimization, we employed an open-source framework to examine the impact of various predictive information strategies on RL outcomes. We propose a joint gated recurrent unit (GRU)-RL algorithm to handle situations where a time-series exists in state space. The results from four classic test cases demonstrate that the proposed GRU-RL method can reduce operating costs by approximately 14.5% and increase comfort performance by 88.4% in indoor comfort control and cost-management tasks. Moreover, the GRU-RL method outperformed the conventional DRL method and was merely augmented with prediction information. In indoor temperature regulation, the GRU-RL algorithm improves control efficacy by 14.2% compared to models without predictive information and offers an approximately 5% improvement over traditional network models. Finally, all models were made open source for easy replication and further research.

Noise isolation in building HVAC system using recyclable plastic bottles

The paper proposes an improvement of the acoustic environment in a sustainable building by utilizing differently sized discarded bottles. The HVAC (heating, ventilation and air conditioning) ducts in buildings serve as conduits for purified air, but they also generate noise as a result of turbulent flow caused by internal irregularities within the ducts. This study involves the modification in the design of ducts by utilizing the discarded or recyclable bottles. By incorporating recyclable bottles along the periphery of the duct, there is decrease in the generation of flow noise. The work presents a comprehensive investigation that employs numerical, experimental, and theoretical methods to assess the sound transmission loss (STL) within the redesigned air duct. The four-microphone impedance tube technique is utilized to measure the STL. The findings reveal that connecting bottles along the length of the air duct at specific intervals leads to a significant reduction in low-frequency noise. Additionally, increasing the number of bottles arranged radially along the air duct leads to a further enhancement and widening of the bandwidth of sound transmission loss (STL).

MF^2: Model-free reinforcement learning for modeling-free building HVAC control with data-driven environment construction in a residential building

Reinforcement Learning (RL) has advanced energy-efficient control of building Heating, Ventilation and Air Conditioning (HVAC) systems. Constructing a suitable RL environment for buildings is a crucial challenge. Compared to widely-used simulation-based environments, data-driven approaches offer higher training efficiency but face convergence difficulties due to influential factors, limiting their current application.To explore data-driven construction of RL environments for building HVAC systems, this study proposes two strategies for controlling room temperature setpoints in a residential building. XGBoost and Long Short-Term Memory Network (LSTM) are trained for energy consumption and room temperature prediction. One strategy predicts parameters for on-off states, while the other for power-on states. The XGBoost models are integrated into an OpenAI Gym environment. The first strategy achieves 0.8634 R2 and 0.2423 Root Mean Squared Error (RMSE) for energy consumption prediction. The R2 of room air temperature models are approximately 0.99 and the RMSE are lower than 0.31. The second strategy achieves 0.9181 R2 and 0.1042 RMSE for energy consumption prediction and similar performance for room temperature prediction. Deep Q-learning (DQN) and Deep Deterministic Policy Gradient (DDPG) algorithms are separately trained using these environments. Results show that the first strategy fails to induce the correct training of RL models, while the second strategy successfully induces a useable DDPG model for controlling building HVAC systems but fails to induce a useable DQN model. We analyze the reasons behind these observations. Compared to the original room temperature setpoint method, the DDPG-based HVAC control logic achieves a 10.06% energy-saving effect while ensuring comfort.

Privacy-Preserving Regulation Capacity Evaluation for HVAC Systems in Heterogeneous Buildings Based on Federated Learning and Transfer Learning

Heating, ventilation, and air conditioning (HVAC) systems in buildings have great potential to provide regulation capacity that is leveraged to maintain the balance of supply and demand in the power system. In order to make full use of HVAC’s regulation capacity, it is important to accurately evaluate it ahead of time. Because physical model-based approaches are hard to implement and highly personalized for each building, data-driven approaches are preferable for this capacity evaluation. However, given the insufficient data for individual buildings and buildings’ potential unwillingness to share their data because of privacy concerns, it is extremely challenging to build a high-performance data-driven regulation capacity evaluation model. In this paper, we propose a privacy-preserving framework that combines federated learning and transfer learning to evaluate the regulation capacity of HVAC systems in heterogeneous buildings. Specifically, a classified federated learning algorithm is proposed to build capacity evaluation models of HVAC systems for different building types. Each building trains its model locally without sharing data with other buildings to preserve privacy. The algorithm also tackles data insufficiency and achieves high evaluation accuracy. In addition, we design a cross-type transfer learning algorithm to enhance model generalization and further address data deficiency. A protocol is created for the above two algorithms to protect privacy and security. Finally, numerical case studies are conducted to validate the proposed framework.

Air quality and comfort constrained energy efficient operation of multi-zone buildings

Maintaining indoor air quality (IAQ) through effective ventilation is essential for the well-being and productivity of building occupants. Control strategies aimed at improving the efficiency of heating, ventilation and air conditioning (HVAC) systems must jointly determine ventilation and heating and cooling processes. In this paper, we study the problem of minimizing the energy consumption of the HVAC system in a multi-zone building, while meeting thermal comfort and IAQ requirements. We first perform a steady state analysis of the zonal carbon dioxide (CO2) concentration and the temperature dynamics. The resulting expressions are convex in the zonal mass flow rates and zonal temperatures. Guided by the steady state solutions for meeting the thermal comfort constraints, we develop two control policies for improving the energy efficiency of building HVAC systems while jointly satisfying indoor temperature and IAQ constraints. We compare the performance of our proposed approaches with those of multiple baseline approaches which implement separate regimes for controlling zonal temperature and IAQ for a typical work-day in a multi-zone campus building. We have evaluated the performance of our proposed approaches under varying levels of flexibility in zonal temperatures. We have shown that zonal temperature flexibility can result in energy savings up to 32% (for the same control strategies) as compared to the case where no such flexibility is permitted. Our proposed approaches were seen to offer potential savings of nearly 29% compared to the baseline.

Similarity learning-based fault detection and diagnosis in building HVAC systems with limited labeled data

Machine learning has been widely adopted for fault detection and diagnosis (FDD) in heating, ventilation and air conditioning (HVAC) systems over the past decade due to the ever-increasing availability of massive building operational data. Machine learning-based FDD is flexible and accurate but heavily relies on the availability of sufficient labeled data to develop supervised or unsupervised models. However, collecting labeled data is usually labor-intensive for various types of faulty conditions, significantly limiting the practical implementation of machine learning-based FDD. Therefore, this study proposes a similarity learning-based method using Siamese networks to improve the performance of machine learning-based FDD in applications with limited labeled data. Unlike the conventional supervised approach, the proposed Siamese networks contain two identical long short-term memory subnetworks which take a pair of multivariate time-series samples from the building energy management system as input. The number of training samples can be significantly augmented by generating pairs randomly. In this way, the generalization ability of the machine learning-based FDD is significantly improved in practical applications. Two case studies were designed and conducted using experimental data when labeled data were limited and imbalanced to validate the proposed similarity learning-based method. In case 1, the proposed method improves the fault diagnostic accuracy by at most 45.7% compared with the baseline model when the number of labeled data is limited. In case 2, the proposed method demonstrated better generalization ability when the labeled data is imbalanced.

Distributed MPC-ILC Thermal Control Design for Large-Scale Multi-Zone Building HVAC System

With building heating, ventilation, and air conditioning (HVAC) systems accounting for 50% the energy consumption in the building sector in the United States, there is a need to develop and implement optimal control strategies for building HVAC systems that reduce energy consumption while achieving thermal comfort for users. In the author's previous work, an integrated model predictive control (MPC) and iterative learning control (ILC) design approach was presented that took advantage of both controllers. It did not rely on model accuracy compared to conventional MPC and reduced the learning curve compared to conventional ILC. Albeit the previous numerical results showed fast convergence in most of the VAV subsystems, the gap between room air temperature and its set point remains noticeable in several zones. This paper further proposes an approach to extend the centralized MPC-ILC controller to take into account the distributed factor and the spatial distribution of the thermal zones of the VAV system. The improved control strategy allows all VAVs to interact with each other and contribute collectively to the overall convergence of the whole system. The proposed controller is implemented on a thirty-two zone VAV reheat system and compared with different controllers, including our previous MPC-ILC design. The outcomes show that the proposed controller results in faster and closer convergence among all zones even when the number of subsystems is large.

Smart Filter Performance Monitoring System

Air filters are widely used in residential and industrial applications. It is designed to remove particulate pollutants in the air to supply cleaner air to either occupants or industrial equipment. Without air filters, the occupants might suffer from polluted air, and expensive industrial equipment could be damaged by contaminants. However, air filters are installed and operated with a limited performance monitoring system, and the efficacy of air filters is unknown after replacement. Here, we propose and prototype a smart filter performance monitoring system that costs less than 200 US dollars and can report filtration efficiency, differential pressure, temperature, and RH (relative humidity) in real time. Three case studies are presented: an air purifier, a teaching building HVAC (heating, ventilation, and air conditioning) system, and a large-scale air cleaning system. Applications of the proposed system cover from household to industrial, which not only verified the concept of the system but also proved its feasibility and advantages of the system. With the proposed system, residential customers can rest assured with their air purifiers or HVAC furnace filter; industrial customers can monitor the filtered air cleanness that will enter their internal combustion engines or gas turbines. Moreover, filter monitoring data can establish a database for researchers to validate the filter models or train a machine learning model for filter performance prediction. Filter or air-cleaning device manufacturers can improve their products or recommend suitable products to their customers based on big data contributed by this filter monitoring system.

A Control Framework to Enable a Commercial Building HVAC System for Energy and Regulation Market Signal Tracking

Commercial buildings are great demand response resources in the energy and regulation markets. Many commercial building heating, ventilation and air conditioning (HVAC) systems are composed of a chiller producing chilled water, and multiple Air-Handling Units (AHUs) distributing cooled air to thermal zones. Demand response of such systems has the potential to closely track energy and regulation market signals. However, existing control methods are mostly focused on fan control, neglecting the impact of the air loop fan control on the power consumption of water loop pumps and chillers. This paper presents a complete modeling of a commercial building HVAC system consisting of chiller, water pump and multiple AHUs, and develops a two-level control to follow five-minute energy market signals and four-second frequency regulation signals. The first-level coarse control provides commands for both water and air loop variables, the second-level fine control adjusts the fan commands while maintaining the water loop inputs within the five-minute control period. As a result, the water-loop power serves as a basis to track energy market signals and the air-loop fan control can be adjusted flexibly for frequency regulation. A high-fidelity model in Dymola is built to validate the model and control. The validation is conducted through co-simulation between Matlab and Dymola model via the Building Controls Virtual Test Bed (BCVTB). The simulation results show significant tracking improvement with the proposed two-level control framework.

Evaluating optimal control of active insulation and HVAC systems in residential buildings

Recently, active insulation systems (AIS) have been conceptualized in building envelopes to optimally modulate thermal resistance in response to changing environmental conditions. Building flexibility can further be improved if the building is also equipped with optimized heating, ventilating, and air conditioning (HVAC) control. In this work, we investigate the annual potential benefits of jointly optimizing AIS and HVAC system controls in both heating and cooling days over all climate zones (CZs) in the U.S. To reduce the computational complexity of applying model predictive control (MPC) to annual operations and detailed whole-building energy models, timeseries clustering was used to identify a set of representative days for optimizing in each climate zone. To isolate the increase in benefits from this joint optimization, we compare the performance to cases where the AIS and HVAC controls are optimized separately. Results indicate savings potential in all CZs, with the largest annual average savings of 9.02% and 4.02% observed in the cooling days with large daily temperature swings and heating days with cold sunny conditions, respectively. Savings patterns across climate zone, day types, and HVAC modes (i.e., heating or cooling) are also discussed along with the implications of important system design variables.

MTL-Deep-STF: A multitask learning based deep spatiotemporal fusion model for outdoor air temperature prediction in building HVAC systems

Buildings consume large quantities of energy. Reducing building energy consumption is essential to achieving carbon neutrality goals. Building energy consumption is strongly influenced by various meteorological factors for which outdoor air temperature is a proxy. This study proposes a data-driven multitask-learning deep spatiotemporal fusion model (MTL-Deep-STF) that predicts outdoor air temperature to provide baseline data against which improvements to conserve energy in heating, ventilation, and air conditioning (HVAC) systems can be proposed and assessed. MTL-Deep-STF is a generic multiple input–multiple output model that predicts temperatures for successive time points in separate high-level tasks. Each task uses a multilayer perceptron to combine meteorological data from neighboring observation stations. CNN-GRU is used to extract spatially local long-term and short-term temporal features that are shared between different tasks via CNN. Comparative validation experiments using real world datasets for complex terrains showed that the model was highly accurate in both single step and multistep predictions. One-step-ahead predictions for winter (MAE = 0.428 °C, RMSE = 0.705, Acc ≤2 °C = 97.832%) and summer (MAE = 0.467 °C, RMSE = 0.706, Acc ≤2 °C = 97.899%) datasets were particularly accurate. This accuracy was coupled with high robustness ( R 2 = 0.976 for winter and R 2 = 0.981 for summer). Use of the model can conserve building energy. • The spatiotemporal fusion model MTL-Deep-STF uses data from neighboring stations to predict air temperature at multiple future time steps. • MTL-Deep-STF treats temperature prediction for successive prediction horizons as distinct tasks. • MTL-Deep-STF extracts global spatial features of multiple stations through fully connected networks. • MTL-Deep-STF is extremely accurate and robust in predicting outdoor temperatures and is therefore suitable for controlling HVAC systems, particularly in complex terrains.

Reducing the spread of COVID-19 transmission through analysis of the evolving building ventilation systems guidance

PurposeThe purpose of this study was to examine how data from the World Health Organization, United States Environmental Protection Agency and Center for Disease Control have evolved with relation to engineering controls for heating, ventilation and air-conditioning (HVAC) systems to mitigate the spread of spread of aerosols (specifically related to the COVID-19 pandemic) in occupied buildings.Design/methodology/approachA document analysis of the pandemic-focused position documents from the aforementioned public health agencies and national HVAC authorities was performed. This review targeted a range of evidence from recommendations, best practices, codes and regulations and peer-reviewed publications and evaluated how they cumulatively evolved over time. Data was compared between 2020 and 2021.FindingsThis research found that core information provided early in the pandemic (i.e. early 2020) for engineering controls in building HVAC systems did not vary greatly as knowledge of the pandemic evolved (i.e. in June of 2021). This indicates that regulating agencies had a good, early understanding of how airborne viruses spread through building ventilation systems. The largest evolution in knowledge came from the broader acceptance of building ventilation as a transmission route and the increase in publications and ease of access to the information for the general public over time.Originality/valueThe promotion of the proposed controls for ventilation in buildings, as outlined in this paper, is another step toward reducing the spread of COVID-19 and future aerosol spread viruses by means of ventilation.

FMI real-time co-simulation-based machine deep learning control of HVAC systems in smart buildings: Digital-twins technology

As heating, ventilation, and air conditioning (HVAC) systems have become one of the most contributing systems in energy consumption in the world, the control of these large-scale systems remains a challenging duty due to the decoupling effects of control variables. Accordingly, the penetration of these types of systems in all-smart buildings has increased in recent years. Furthermore, the application of digital twin as a fast-growing concept is being developed. In HVAC systems, independent and accurate control of temperature and humidity of the indoor air has been playing an undeniable role in reducing energy consumption. In this paper, to have cost-effective energy management in a single-zone HVAC system, a new reliable digital twin proximal policy optimization (PPO)–based model-independent nonsingular terminal sliding-mode control (MINTSMC) methodology has been proposed. Moreover, due to the nonlinear characteristics of HVAC systems, MINTSMC tends to handle the un-modeled system dynamics and disturbances. For regulating parameters of proposed control, an efficient PPO algorithm has been developed due to its actor-critic-based reinforcement learning. Extensive examinations and comparative analyses with particle swarm optimization designed sliding-mode control and proportional–integral–derivative controller have been made using digital twin of the proposed controller to show the importance, accuracy, and application of this method in the comfort and energy management achievement of HVAC control systems. A digital signal processor computing device has been utilized for implementation by utilizing hardware-in-loop (HIL) in the concept of the digital twin. To determine the interface between established models, software-in-loop, and HIL, the Functional Mock-up Interface has been utilized. The outcomes revealed a superior performance of suggested digital twin-based controller than the compared control methodologies in the compensation of unknown uncertainties, fast-tracking, and smooth response.

Model predictive control of heating, ventilation, and air conditioning (HVAC) systems: A state-of-the-art review

Due to the fast advancement of communication and information technology, intelligent buildings have garnered great interest. These buildings can forecast weather, ambient temperature, and sun irradiation and can modify heating, ventilation, and air conditioning (HVAC) operations appropriately, based on current and previous data. This change is intended to reduce HVAC system energy usage while maintaining an appropriate degree of thermal comfort and indoor air quality. Since its inception, model predictive control (MPC) has been one of the prospective solutions for HVAC management systems to reduce both costs and energy usage. Additionally, MPC is becoming increasingly practical as the processing capacity of building automation systems increases and a large quantity of monitored building data becomes available. MPC also provides the potential to improve the energy efficiency of HVAC systems via its capacity to consider limitations, to predict disruptions, and to factor in multiple competing goals such as interior thermal comfort and building energy consumption. Although substantial research has been conducted on MPC in building HVAC systems, there is a shortage of critical reviews and a lack of a comprehensive framework that formulates and defines the applications. This article provides a comprehensive state-of-the-art overview of MPC in HVAC systems. Detailed discussions of modeling approaches and optimization algorithms are included. Numerous design aspects such as prediction horizon, occupancy behavior, building type, and cost function, that impact MPC performance are discussed in detail. The technical characteristics, advantages, and disadvantages of various types of modeling software are discussed. The primary objective of this work is to highlight critical design characteristics for the MPC control scheme and to give improved suggestions for future research. Moreover, numerous prospective scenarios have been suggested that might provide future research direction.

Discrete-Time-Distributed Adaptive ILC With Nonrepetitive Uncertainties and Applications to Building HVAC Systems

Aiming to addressing the nonrepetitive uncertainties of multiagent systems, this work proposes a discrete-time-distributed adaptive iterative learning control (DDAILC) scheme for an output consensus problem, where two fundamental requirements in the traditional distributed iterative learning control (ILC) methods, i.e., the identical initial states and the repetitive desired trajectories, are removed. Furthermore, the algorithm design and analysis are directly aimed at discrete-time nonlinear multiagent systems, rather than continuous-time ones, to meet the needs of practical implementations. The iteration-varying trajectory of the virtual leader is included in the learning control protocol for a compensation. The adaptive parameter-updating law works along the iteration dimension by using a general consensus error that contains the output data of adjacent agents. To ensure the estimation of the control gain to be nonzero, a semisaturator is utilized in the parameter-updating law. The convergence of the output consensus is shown rigorously. Both numerical and practical examples are used to test the theoretical results. Moreover, the DDAILC efficiently improves performance of the building heating, ventilation, and air conditioning (HVAC) system by utilizing both the distributed topology and the repetitive dynamic characteristic.

End-to-End Deep Reinforcement Learning Control for HVAC Systems in Office Buildings

The heating, ventilation, and air conditioning (HVAC) system is a major energy consumer in office buildings, and its operation is critical for indoor thermal comfort. While previous studies have indicated that reinforcement learning control can improve HVAC energy efficiency, they did not provide enough information about end-to-end control (i.e., from raw observations to ready-to-implement control signals) for centralized HVAC systems in multizone buildings due to the limitations of reinforcement learning methods or the test buildings being single zones with independent HVAC systems. This study developed a model-free end-to-end dynamic HVAC control method based on a recently proposed deep reinforcement learning framework to control the centralized HVAC system of a multizone office building. By using the deep neural network, the proposed control method could directly take measurable parameters, including weather and indoor environment conditions, as inputs and control indoor temperature setpoints at a supervisory level. In some test cases, the proposed control method could successfully learn a dynamic control policy to reduce HVAC energy consumption by 12.8% compared with the baseline case using conventional control methods, without compromising thermal comfort. However, an over-fitting problem was noted, indicating that future work should first focus on the generalization of deep reinforcement learning.