WO2024102474A1

WO2024102474A1 - System and method for ai-optimized hvac control

Info

Publication number: WO2024102474A1
Application number: PCT/US2023/037153
Authority: WO
Inventors: Dongxiao ZHU; Masoud H. NAZARI; Caisheng WANG
Original assignee: Wayne State University
Priority date: 2022-11-11
Filing date: 2023-11-10
Publication date: 2024-05-16

Abstract

A system including a plurality of sensors and a building management system (BMS). The plurality of sensors are configured to receive occupancy data from at least a first room. The BMS includes at least one computer readable storage medium having instructions thereon to employ a gating mechanism to analyze the occupancy data via a recurrent neural network (RNN) to create a first occupancy prediction of the first room during a first upcoming time period. The RNN employs gated recurrent units (GRUs) and each of the GRUs employs at least an update gate and a reset gate. The instructions further cause adjustment of a heating, ventilation, and cooling (HVAC) system for the first upcoming time period based on the first occupancy prediction. The adjustment of the HVAC system includes adjustment of at least one of heating, cooling, and ventilation in the first room.

Description

SYSTEM AND METHOD FOR AI-OPTIMIZED HVAC CONTROL

CROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims priority to U.S. Provisional Application 63/424,540 filed November 11, 2022, the contents of which is incorporated herein by reference in its entirety.

TECHNICAL FIELD

[0002] This disclosure relates generally to Al integration for occupancy prediction in zones of a building to optimize HVAC control, energy efficiency, and climate management.

BACKGROUND

[0003] Buildings are intended to provide the occupants with a comfortable environment. From an occupant’s perspective, constant comfortable temperature and fresh air often define comfort. To achieve this, heating, ventilation, and air conditioning (HVAC) are generally employed.

[0004] In addition to comfort, however, energy' demands need to be considered. Buildings are generally considered major energy consumers and their energy demands may have significant impact on global energy. For example, in the United States, buildings may consume up to 40% of total energy demand. As such, a challenge for a Building Management Systems (BMS) is to provide a comfortable environment for occupants, while also minimizing energy consumption.

[0005] As mentioned, thermal comfort and air quality are generally driving forces that define occupant comfort. Thermal comfort, however, is not always achieved. Often, heating and cooling controlled by a BMS is based on occupancy predictions. For example, during “work hours’’, building temperatures may be set to a fixed temperature in an attempt to keep occupants comfortable. Then, during “off hours'’, the heating or cooling may be relaxed to minimize energy expenditures. Buildings, however, are often composed of a large amount of space or rooms. Further, the use of these “rooms” may vary' hourly.

[0006] BMSs have been used extensively to control on/off HVAC equipment time schedules based on outdoor temperatures and other variables. Existing systems, however, generally do not capture real-time building occupancy and the dynamic behavior of occupants. As such, rooms may be heated or cooled when there are no occupants therein. Additionally, rooms may not be properly heated or cooled when occupants are present. Both scenarios may lead to inefficiently using an HVAC system.

[0007] Similarly, ineffective occupancy predictions may have a negative impact when using other, or additional, environmental control systems. For example, ineffective occupancy predictions may lead to inefficient use of lighting. That is, rooms or zones may be lit when occupants are not present or, alternatively, rooms or zones may be unlit when occupants arrive. [0008] As such, there is a need for better occupancy detection

BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 illustrates an exemplary block diagram of an occupancy prediction system;

[0010] Figure 2 illustrates an exemplary technique for occupancy detection;

[0011] Figure 3 illustrates an exemplary HVAC control system employing an occupancy prediction system; and

[0012] Figure 4 illustrates an exemplary occupancy prediction workflow.

DESCRIPTION

[0013] With reference to Figure 1, an exemplary block diagram of an occupancy prediction system 100 is shown. The exemplary system 100 includes at least one computer readable storage medium 102 having instructions thereon embodied by a schematic 104. Further shown is at least one sensor 106 in a room 108. The sensor 106 detects whether or not the room 108 is occupied. Occupancy data 110 is provided from the sensor 106 to the occupancy prediction system 100 where it is analyzed to predict occupancy of the room 108 at a future time.

[0014] The schematic 104 represents occupancy data 110, received via at least one occupancy sensor 106, being input a recurrent neural network (RNN) 1 12, which is employed to analyze the occupancy data 1 10 to predict occupancy of the room at later times.

[0015] The RNN 112 may employ gated recurrent units (GRUs) 114, a dropout layer 116, and a dense layer 118. The GRUs include an update gate and a reset gate. The dropout layer 116 helps prevent or minimize overfitting. The dense layer 118 is a layer of neurons where each neuron receives input from a previous layer of neurons.

[0016] Unlike a traditional long short-term (LTSM) recurrent neural network, the RNN 112 generally does not employ input, output, and forget gates. Accordingly, the RNN 112 has fewer gates than a traditional LTSM RNN. It is noted that the RNN 112 is merely exemplary and other configurations of the RNN 112 employing its own set of GRUs may be employed. [0017] Upon analyzing the occupancy data 110, occupancy predictions 120 of future occupancy in the room 108 may be output to, for example, a BMS 122. The BMS 122 may then use the occupancy predictions 120 to adjust environmental controls of a building or room 108. For example, the occupancy predictions 120 may represent that the room 108 may be occupied at a future time. As such, the BMS 122 may, based on the occupancy prediction 116, adjust the environmental controls 120 the room 108 to adjust aspects of that room at the future time. For example, the room temperature may be raised or lowered to a desired temperature at that future time. Alternatively, the room temperature may be adjusted prior to that future time so that the room attains the target temperature at that future time. For example, it may take 10 minutes for a room’s temperature to reach the target temperature. As such, the BMS 122 may adjust the adjust environmental controls 10 minutes prior to the future time so the room 108 reaches the target temperature when the future time.

[0018] Similarly, based on the occupancy predictions 120, the BMS 122 may adjust the environmental controls if/when the occupancy predictions 120 predicts that the room 108 will not be occupied at a future time. Accordingly, heating and cooling costs can be minimized since the occupancy prediction 120 can be used to determine when the room 108 will likely be occupied.

[0019] Similarly, based on the occupancy predictions 120, the BMS 122 may adjust environmental controls to affect lighting or other environmental factors to minimize energy usage.

[0020] While occupancy detection system 100 of Figure 1 is shown as distinct from BMS 122, other examples may have the occupancy detection system 100 integrated into the BMS 122.

[0021] Referring now to Figure 2, an exemplary technique 200 for occupancy detection is shown. The technique 200 includes receiving occupancy data from a plurality of sensors at block 202. The occupancy data may be received on at least one computer readable storage medium.

[0022] Upon receiving the occupancy data, process control proceeds to block 204, where employing a gating mechanism to analyze the occupancy data via a trained recurrent neural network (RNN) to create a first occupancy prediction of a first zone (e.g., a room in a building or a portion of a geographic area) during a first upcoming time period is carried out. To carry⁷ out the analysis, the trained RNN employs gated recurrent units (GRUs), where each of the GRUs employs at least an update gate and a reset gate. As will be understood, instructions embodied on at least one computer readable storage medium may cause the analysis to be carried out.

[0023] The predicted occupancy (e.g., a first occupancy prediction) is a prediction of whether or not the first zone will be occupied during an upcoming time frame (i.e., a first upcoming time period).

[0024] After analysis of the occupancy data, process control may proceed to block 206, where causing adjustment of a heating, ventilation, and cooling (HVAC) system based on the first occupancy prediction may be carried out. The adjustment adjusts at least one of heating, cooling, and ventilation in the first zone. As such, energy costs may be minimized. For example, the occupancy prediction may identify that during an upcoming time period (e.g., minutes or hours ahead of the cunent time) that the first room will not be occupied. As such, the HVAC system may be adjusted since there is no need to keep the temperature in the first zone at a level that is comfortable to a user. Alternatively, if it is predicted that the first room will be occupied at the upcoming time period based on the occupancy prediction, the HVAC system can be adjusted so that the environment in the zone is comfortable to a user during the upcoming time period.

[0025] As an alternative, the analyzed occupancy data may instead, or in addition, be employed for other purposes. For example, the occupancy data could be employed to control lighting in a room (a.k.a., zone). Based on the analyzed occupancy data, the lights in the room may be turned on when occupancy is predicted, and off when occupancy is not predicted.

[0026] Still further, the occupancy prediction may be employed to predict when an outside zone will be occupied, for example, by user(s), vehicle(s), and/or other animal(s). Accordingly, exemplary environments where such occupancy predictions can be employed include, for example, outdoor festival grounds or fairgrounds, parking lots or streets, and zoos, to name just a few.

[0027] After adjustments are made to environmental settings (e.g., HVAC controls, lighting, and/or etc.), process control may come to an end. Alternatively, the RNN may continue to receive occupancy data to make further occupancy predictions as time passes on. [0028] It is noted that the technique 200 may also include employing the gating mechanism to analyze the occupancy data via the RNN to create a plurality of occupancy predictions during a plurality' of upcoming time periods. For example, occupancy predictions can be made for an entire week or more. Take for example, an upcoming work week. The occupancy data may be employed to predict what times the zone (e.g., a room) will and will not be occupied during the week. These predictions may then be used to control environmental settings (e.g., HVAC controls, lighting, and/or etc.) throughout the week.

[0029] Still further, the technique 200 may be carried out to make occupancy predictions for a plurality⁷ of zones (e.g., a plurality of rooms in a building or a plurality of spaces in an outdoor space). For example, the occupancy data may be employed to predict what rooms will be occupied, and at what time those rooms will be occupied, during the week.

[0030] With reference now to Figure 3, an exemplary HVAC control system 300 is shown. The system 300 includes a management system (e.g., Building Management System (BMS) 302) in communication with a plurality of sensors 304-N. As will be discussed below, occupancy data 308 received from the sensors 304-N is employed by the BMS 302 to make occupancy predictions for a plurality of zones 310-N (e.g., rooms in a building). The predictions are then employed to control an HVAC system and/or other types of environmental control systems such a lighting system (not show n) to effect the environment in the zones 310- N. After training the BMS 302, occupancy data 308 received at, for example, five minute increments, may be used by the BMS 302 to make occupancy forecasts (e.g., six hours into the future).

[0031] One or more sensors 304-N are placed in each respective zone 310-N. The sensors 304-N are configured to detect occupancy in each of zones 310-N, and the respective occupancy data 308 is relayed to the BMS 302 for processing in order to make occupancy predictions for each of the respective zones 310-N. Any effective occupancy sensor 304-N may be employed. For example, the sensors 304-N may detect light disturbances, carbon dioxide, infrared, doppler changes, and/or other phenomenon to determine if a zone is occupied.

[0032] Regarding the HVAC system, it may include a Building Automation Controller network (BACnet Controller) 314, a variable air volume (VAV) controller 316, and a plurality of timers/thermostats 318-N. The BACnet controller 314 enables communication among the BMS 302 and the VAV controller 316.

[0033] The VAV controller 316 controls an air handling system 322. For example, if the BMS 302 predicts that one or more zones will not be occupied for two hours, the temperature set point of the corresponding VAV controller 316 can be relaxed in those zones by, for example, 7°F, during the unoccupied period. So that the zones are comfortable upon occupant arrival, the set point on the VAV controller 316 can return to its nominal value (e.g., 72°F) before occupant arrival to ensure occupant’s comfort upon arrival. This relaxation (e.g., 7°F) of the HVAC system can result in energy savings and support the power grid.

[0034] The system 300 may also include a gateway 324, an energy manager 326, and an override controller 328. The gateway 324 serves as a liaison between the network on which the sensors 304-N reside and the network on which the BMS 302 resides.

[0035] The energy manager will send the occupancy-based control signals to the BMS. The override controller 328 allows the BMS 302 to control the HAV C system or for manual control of the HVAC system, depending on its setting.

[0036] Regarding the BMS 302, it may include at least one computer readable storage medium 330 and one or more processors 332. Via the storage medium(s) 330 and processor(s) 332, the BMS 302 analyzes occupancy data 308 received via the sensors 304-N to make occupancy predictions. Initially, the sensors 304-N are employed to gather occupancy data 308 from the zones 310-N to train the BMS 302. To overcome issues (e.g., long training periods) that may arise when large amounts of data is employed to make predictions, the BMS 302 employs a gated recurrent unit (GRU) architecture (e.g., a one-layer GRU architecture). The GRU architecture employs an update gate (Z) and a reset gate (r) and recognizes patterns that occur over changing time scales. Since the number of gates employed are minimized, the number of parameters per gate are also minimized. As such, the GRU architecture minimizes the amount of training parameters needed during training.

[0037] The update gate may take the form of Zt = o( W_z t-i + LWCt). Z_t is an input vector, h_t is an output vector, and Xt is a reset vector. W and U are parameter matrices and vector and a is a logistic function. The update gate may be employed to determine how much of a new input should be used to update a hidden state.

[0038] The reset gate may take the form of r_t = a(W_rht-l + UXt . The reset gate may be employed to determine how much of the hidden state should be forgotten.

[0039] A cell gate may take the form of G = tanh(W_c(h_t-l * r) + UXt), while a new state may take the form of ht = (Zt * Ct) + ((1 - Zt) * h_t-i).

[0040] As opposed to a network that employs an input gate, output gate, and reset gate (e.g., long short-term memory network (LTSM)), the amount of gates employed by the GRU architecture are minimized. As such, the number of parameters per gate are also minimized, thus increasing training and prediction efficiency since there are fewer training parameters.

[0041] Adaptive moment estimation (ADAM) optimization may be used to train the network weights. Sigmoid may be employed as the activation function and a dropout of , for example, 10% may be applied. This may serve as a regularization technique to remove neurons whose probability is equal to the dropout value from the neural network. This may help build a more robust network that is less affected by input fluctuations. Dropout refers to the number of neurons in the network that are randomly and temporarily deactivated during a training epoch. Accordingly, this evenly distributes the inputs across the network and may prevent the contribution of any one node from dominating the output. Networks trained by using dropout are generally more resistant to overfitting than networks trained without using dropout. A binary cross-entropy loss function may be used for back-propagation.

[0042] Referring now to Figure 4, an exemplary occupancy prediction workflow 400 is represented. The workflow includes a data preparation module 402 where historical occupancy data is received 404. The historical occupancy data can come from a plurality of sensors deployed in zones for which occupancy predictions is desired. Further, the historical occupancy data may include occupancy data received over days, weeks, months, or more.

[0043] Once received, the historical occupancy data is split 406 and portion is provided to a training module 408 and a portion may be provided to the testing module 410. Effectively, a portion of the historical occupancy data will be used to train the GRU architecture 412 and another portion of the historical occupancy data will be used to test/validate the occupancy predictions provided by the GRU architecture 412. As an example, the training module 408 may receive seven weeks of occupancy data taken at five minute intervals associated with the plurality of rooms, while the testing module 410 may receive one week of historical data, also taken at five minute intervals. Other examples may include data from different time periods gathered at different intervals.

[0044] At the training module 408, the portion of historical data received is prepared as training samples 414 and assessed or analyzed by the GRU architecture 412 to output occupancy prediction data 416. For example, the occupancy predictions may be a plurality of occupancy predictions for a variety of rooms over a plurality of upcoming time frames.

[0045] Regarding the testing module 410, the portion of historical data received is prepared as testing samples 418. The prepared testing samples are then provided, as testing data, to a Classification model 420 along with the occupancy predictions 416. In turn, the testing sample data 418 and occupancy predictions 416 are classified, as understood in the art, for processing. [0046] After the data is classified 420, a validation model 422 determines whether the occupancy prediction(s) were valid or not. If the result indicates that the GRU architecture has not been effectively trained, additional historical occupancy data can be provided until the GRU architecture is effectively trained.

[0047] When the GRU architecture is effectively trained, it is deployed as a trained GRU architecture 424. At which time, real-time occupancy data 426 can be input into the trained GRU architecture 424 and analyzed. Once analyzed, the trained GRU architecture 424 outputs future occupancy predictions 428. The future occupancy predictions 428 may, for example, predict how occupancy may vary over a future time period (e.g., 6 hours). These predictions may then be employed by an environmental control system (e.g., a BMS system) to control, for example, an HVAC system, lighting system, or the like. Accordingly, the occupancy predictions contribute to energy efficiency improvement and demand response capability for smart buildings or other infrastructure.

[0048] Since the trained GRU architecture 424 may receive real-time data 426, the system may become more refined as time goes on.

[0049] With reference to the Figures, although the Figures represent some possible approaches, the Figures are not necessarily to scale and certain features may be exaggerated, removed, or partially sectioned to better illustrate and explain the present disclosure. Further, the descriptions set forth herein are not intended to be exhaustive, otherwise limit, or restrict the claims to the precise forms and configurations shown in the Figures and disclosed in the detailed description.

[0050] When introducing elements of various embodiments of the disclosed materials, the articles “a,” “an,” "‘the,” and “said” are intended to mean that there are one or more of the elements. The terms “comprising,” “including,” and “having” are intended to be inclusive and mean that there may be additional elements other than the listed elements. Further, terms such as “first”, “second”, and “third” do not denote a sequence unless explicitly noted otherwise. That is, unless noted otherwise, terms such “first”, “second”, and “third” merely distinguish between elements, features, or the like. Furthermore, any numerical examples in the discussion above are intended to be non-limiting, and thus additional numerical values, ranges, and percentages are within the scope of the disclosed embodiments.

[0051] While the disclosed materials have been described in detail in connection with only a limited number of embodiments, it should be readily understood that the embodiments are not limited to such disclosed embodiments. Rather, the disclosed embodiments can be modified to incorporate any number of variations, alterations, substitutions or equivalent arrangements not heretofore described, but which are commensurate with the spirit and scope of the disclosed materials. Additionally, while various embodiments have been described, it is to be understood that disclosed aspects may include only some of the described embodiments. Accordingly, that disclosed is not to be seen as limited by the foregoing description, but is only limited by the scope of the appended claims.

[0052] It should be understood that a compute device or system (e.g., occupancy detection system 100 of Figure 1 and BMS 302 of Figure 3), a system, and/or a processor as described herein may include a conventional processing apparatus known in the art, which may be capable of executing preprogrammed instructions stored in an associated memory, all performing in accordance with the functionality described herein. To the extent that the methods described herein are embodied in software, the resulting software can be stored in an associated memory and can also constitute means for performing such methods. Such a system or processor may further be of the type having ROM, RAM, RAM and ROM, and/or a combination of non-volatile and volatile memory⁷ so that any software may be stored and yet allow storage and processing of dynamically produced data and/or signals.

[0053] It should be further understood that an article of manufacture in accordance with this disclosure may include a non-transitory computer-readable storage medium having a computer program encoded thereon for implementing logic and other functionality described herein. The computer program may include code to perform one or more of the methods disclosed herein. Such embodiments may be configured to execute via one or more processors, such as multiple processors that are integrated into a single system or are distributed over and connected together through a communications network, and the communications network may be wired and/or wireless.

[0054] A BMS or occupancy detection system may include any computing system. The systems may communicatively connect with and transfer information with respect to other devices and respective storage mediums. The systems may be in continuous or periodic communication with other elements/devices. The systems may include a local, remote, or cloud-based server or a combination thereof and may be in communication with and provide information to any or a combination of the elements/devices. The systems may further provide a web-based user interface (e.g., an internet portal) to be displayed by the user interface.

[0055] The systems may be connected via any wired or wireless connections between two or more endpoints, for example, to facilitate transfer of information. Connection may include a local area network, for example, to communicatively connect the devices, with a network. Connection may include a wide area network connection, for example, to communicatively connect a server with a network. Connection may include a wireless connection, e.g., radiofrequency (RF), near field communication (NFC), Bluetooth communication, Wi-Fi, or a wired connection, for example, to communicatively connect the devices.

Claims

CLAIMS What is claimed is:

1. A system comprising: a plurality of sensors configured to receive occupancy data from at least a first room; and a building management system (BMS) having at least one computer readable storge medium, the at least one computer readable storage medium having instructions thereon to: employ a gating mechanism to analyze the occupancy data via a recurrent neural network (RNN) to create a first occupancy prediction of the first room during a first upcoming time period, wherein the RNN employs gated recurrent units (GRUs), and wherein each of the GRUs employs at least an update gate and a reset gate; and adjust a heating, ventilation, and cooling (HVAC) system for the first upcoming time period based on the first occupancy prediction, wherein adjustment of the HVAC system includes adjustment of at least one of heating, cooling, and ventilation in the first room.

2. The system of claim 1 wherein the first occupancy prediction predicts that the first room will be occupied during the first upcoming time period.

3. The system of claim 2 wherein the RNN does not employ a forget gate.

4. The system of claim 1 wherein the adjustment of the HVAC system occurs prior to the first upcoming time period.

5. The system of claim 1 wherein the at least one computer readable storage medium has further instructions thereon to create a plurality of occupancy predictions for a lurality of respective additional rooms during the first upcoming time period and a plurality of additional upcoming time periods using the occupancy data received via the plurality of sensors.

6. The system of claim 4 wherein the plurality of respective additional rooms has at least one sensor of the plurality of sensors therein.

7. The system of claim 1 wherein the RRN that employs the GRUs is a one-layer GRU network.

8. The system of claim 1 wherein the first upcoming time period is within six hours of a current time period.

9. A computer readable storage medium having instructions thereon to: receive occupancy data from a plurality of sensors; employ a gating mechanism to analyze the occupancy data via a recurrent neural network (RNN) to create a first occupancy prediction of a first zone during a first upcoming time period, wherein the RNN employs gated recurrent units (GRUs), and wherein each of the GRUs employs at least an update gate and a reset gate; and cause adjustment of a heating, ventilation, and cooling (HVAC) system based on the first occupancy prediction, wherein the adjustment adjusts at least one of heating, cooling, and ventilation in the first zone.

10. The computer readable storage medium of claim 9 wherein the first occupancy prediction predicts that the first zone will be occupied during the first upcoming time period, and wherein the first zone is a room in a building.

11. The computer readable storage medium of claim 9 wherein the RNN does not employ a forget gate.

12. The computer readable storage medium of claim 9 wherein the adjustment of the HVAC system occurs prior to the first upcoming time period.

13. The computer readable storage medium of claim 9 having further instructions thereon to create a plurality of occupancy predictions for a plurality of respective additional rooms during a plurality of upcoming time periods using the occupancy data received via the plurality of sensors, and wherein the plurality of occupancy predictions predict occupancy over the plurality of upcoming time periods.

14. The computer readable storage medium of claim 9 wherein the RRN that employs the GRUs is a one-layer GRU network.

15. A method comprising: receiving occupancy data from a plurality of sensors; employing a gating mechanism to analyze the occupancy data via a recurrent neural network (RNN) to create a first occupancy prediction of a first room during a first upcoming time period, wherein the RNN employs gated recurrent units (GRUs), and wherein each of the GRUs employs at least an update gate and a reset gate; and causing adjustment of a heating, ventilation, and cooling (HVAC) system based on the first occupancy prediction, wherein the adjustment adjusts at least one of heating, cooling, and ventilation in the first room.

16. The method of claim 16 wherein the first occupancy prediction predicts that the first room will be occupied during the first upcoming time period, and wherein the method further comprises employing the gating mechanism to analyze the occupancy data via the RNN to create a plurality of occupancy predictions during a plurality of upcoming time periods.

17. The method of claim 16 wherein the RNN does not employ a forget gate.

18. The method of claim 16 wherein the adjustment of the HVAC system occurs prior to the first upcoming time period.

19. The method of claim 16 further comprising creating a plurality of occupancy predictions for a plurality of respective additional rooms during the first upcoming time period using the occupancy data received via the plurality of sensors.

20. The method of claim 16 wherein the RRN that employs the GRUs is a one-layer GRU network.