Fine-grained occupancy pattern estimation for HVAC control

Abstract

The present disclosure relates to utilizing a refined granularity occupancy estimation model for efficient heating, ventilation, and air conditioning (HVAC) control. A thermal state of a plurality of zones of a building is updated based on a building thermal model and information received from a temperature sensor of the building. A predicted occupant count for an upcoming plurality of time slots for each of the plurality of zones is updated using an actual occupancy count for each of the plurality of zones. A misprediction type distribution for an upcoming plurality of time slots for each of the plurality of zones is updated, the misprediction type distribution indicating mispredictions of true negatives, false positives, false negatives, and true positives. A total misprediction cost expectation is updated based on the predicted occupant counts and the misprediction type distribution. An HVAC power is determined for each of the plurality of zones to optimize occupant thermal comfort weighted by the predicted occupant counts while minimizing the total misprediction cost expectation. HVAC operation is controlled by the HVAC power.

Description

[0001] STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0002] This invention was made with Government support under Grant DE-EE0007682 awarded by the Department of Energy and Grant CNS-1553273 awarded by the National Science Foundation. The Government has certain rights in this invention. Technical Field

[0003] The present disclosure relates to utilizing fine-grained occupancy estimation models for efficient heating, ventilation, and air conditioning (HVAC) control. Background Art

[0004] HVAC is a major source of energy consumption in the United States. In 2017, approximately 30% of commercial buildings' energy consumption in the United States was for HVAC. Building operators often use static scheduling to control HVAC systems without a deeper understanding of how many people use the building at different times of the day. Furthermore, many HVAC systems operate by assuming maximum occupancy in each room. This can lead to significant energy waste, for example, when an HVAC system ventilates 30 people when only 10 are present in the room. Such widely used HVAC control designs miss opportunities for more accurate and efficient control. Summary of the Invention

[0005] In one or more illustrative examples, a method for controlling HVAC operation of a building to minimize energy consumption using predicted occupant counts and accounting for misprediction costs includes updating a thermal state of multiple zones of the building based on a building thermal model and information received from temperature sensors of the building; updating a predicted occupant count for an upcoming multiple time slots for each of the multiple zones using an actual occupancy count for each of the multiple zones; updating a misprediction type distribution for the upcoming multiple time slots for each of the multiple zones, the misprediction type distribution indicating mispredictions of true negatives, false positives, false negatives, and true positives; updating a total misprediction cost expectation based on the predicted occupant counts and the misprediction type distribution; determining an HVAC power for each of the multiple zones to optimize occupant thermal comfort weighted by the predicted occupant counts while minimizing the total misprediction cost expectation, the determination being constrained by a heat exchange configuration of the building and a minimum and maximum HVAC power available to each of the multiple zones; and controlling the HVAC operation of the building based on the HVAC power determined for each of the multiple zones.

[0006] In one or more illustrative examples, a system for controlling HVAC operation of a building to minimize energy consumption using predicted occupant counts and accounting for misprediction costs includes an HVAC controller programmed to update a thermal state of multiple zones of the building based on a building thermal model and information received from temperature sensors of the building; update a predicted occupant count for an upcoming multiple time slots for each of the multiple zones using an actual occupancy count for each of the multiple zones; update a misprediction type distribution for the upcoming multiple time slots for each of the multiple zones, the misprediction type distribution indicating mispredictions for true negatives, false positives, false negatives, and true positives; update a total misprediction cost expectation based on the predicted occupant counts and the misprediction type distribution; determine an HVAC power for each of the multiple zones to optimize occupant thermal comfort weighted by the predicted occupant counts while minimizing the total misprediction cost expectation, the determination being constrained by a heat exchange configuration of the building and a minimum and maximum HVAC power available to each of the multiple zones; and control the HVAC operation of the building based on the HVAC power determined for each of the multiple zones.

[0007] In one or more illustrative examples, a non-transitory computer-readable medium includes instructions for controlling heating, ventilation, and air conditioning (HVAC) operations of a building to minimize energy consumption using predicted occupant counts and accounting for misprediction costs, the instructions, when executed by an HVAC controller, causing the HVAC controller to update a thermal state of a plurality of zones of the building based on a building thermal model and information received from temperature sensors of the building; update a predicted occupant count for an upcoming plurality of time slots for each of the plurality of zones using an actual occupancy count for each of the plurality of zones; update a misprediction type distribution for the upcoming plurality of time slots for each of the plurality of zones, the misprediction type distribution indicating mispredictions for true negatives, false positives, false negatives, and true positives; update a total misprediction cost expectation based on the predicted occupant counts and the misprediction type distribution; determine an HVAC power for each of the plurality of zones to optimize occupant thermal comfort weighted by the predicted occupant counts while minimizing the total misprediction cost expectation, the determination being constrained by a heat exchange configuration of the building and minimum and maximum HVAC powers available to each of the plurality of zones; and control HVAC operations of the building based on the HVAC power determined for each of the plurality of zones. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 An example occupancy-aware HVAC control system for a building is illustrated;

[0009] Figure 2An example of occupancy dynamics is illustrated as a temporally inhomogeneous Markov chain unfolding for a day;

[0010] Figure 3 The proposed network combines temporal and contextual features to predict the next N Occupancy count;

[0011] Figure 4 The performance of the time-inhomogeneous Markov chain predictor is shown;

[0012] Figure 5 The performance of the linear regression predictor is shown;

[0013] Figure 6 The performance of sequential and contextual neural network predictors is shown;

[0014] Figure 7 The graph shows the average occupant count per day over a six-month period;

[0015] Figure 8 An example of the performance of an HVAC controller using ground truth and neural network predicted occupant counts is shown.

[0016] Figure 9 illustrates an example adaptive model predictive HVAC control architecture;

[0017] Figure 10 illustrates an example process for controlling HVAC operations of a building to minimize energy consumption using predicted occupant counts and accounting for misprediction costs; and

[0018] Figure 11 Illustrated are example computing devices for performing various aspects of HVAC operations described in detail herein. DETAILED DESCRIPTION

[0019] Embodiments of the present disclosure are described herein. However, it should be understood that the disclosed embodiments are merely examples, and other embodiments may take various forms and alternative forms. The figures are not necessarily to scale; some features may be enlarged or minimized to show details of particular components. Therefore, the specific structural and functional details disclosed herein should not be interpreted as limiting, but merely as a representative basis for teaching those skilled in the art to adopt the embodiments in different ways. As will be understood by those of ordinary skill in the art, the various features illustrated and described with reference to any one of the figures may be combined with the features illustrated in one or more other figures to produce embodiments that are not explicitly illustrated or described. The combination of illustrated features provides representative embodiments for typical applications. However, various combinations and modifications of features consistent with the teachings of the present disclosure may be desirable for specific applications or implementations.

[0020] Some buildings include coarse-grained occupancy sensors that provide binary data related to building occupancy patterns (e.g., whether a space is occupied). Finer-grained occupancy sensors are increasingly being deployed in modern commercial buildings. These fine-grained occupancy sensors provide real-time occupancy information, not just whether a space is occupied. This occupancy count information can benefit building HVAC control to improve building energy efficiency, enhance occupant thermal comfort, and enhance building management. By using fine-grained occupancy sensor data, predictive models can be designed to capture the spatiotemporal characteristics of occupancy dynamics. Using such predictive models, model predictive control algorithms for HVAC control can be implemented, resulting in significant energy savings.

[0021] Figure 1 An example occupancy-aware HVAC control system 100 for a building 102 is illustrated. As shown, building 102 includes occupancy sensors 108 configured to sense occupancy of building 102. An occupant count prediction model 112 receives the number of occupants sensed in the building from occupancy sensors 108. Based on current real-time and historical sensed occupant information, occupant count prediction model 112 outputs an occupant count prediction indicating the predicted number of occupants in each zone of building 102 at various time slots. An HVAC controller 106 is configured to receive the occupant count prediction, including temperature information in the building, and the building status, and maintain a building thermal model 110. The building thermal status is detected via a building monitoring system, such as temperature sensors located in zones of building 102. The building thermal model 110 may be based on building 102 data (such as the floor plan, building materials, HVAC system deployment, etc.). Using the occupant count prediction and building status, HVAC controller 106 provides control input to the building HVAC system 104. Actuation of the building HVAC system 104, in turn, affects the building status. Thus, the occupancy-aware HVAC control system 100 adaptively adjusts the building HVAC 104 based on predicted occupant counts and real-time building thermal conditions to maintain occupant comfort and conserve energy. It should be noted that the illustrated HVAC control system 100 is an example, and systems with more, fewer, or differently arranged elements may be used.

[0022] Building 102 may be an enclosed structure where the climate is controlled by occupancy-aware HVAC control system 100. In many examples, building 102 has walls, a roof, windows, and in many cases, more than one floor. Building 102 may include one or more rooms and doors or other openings through which occupants can enter and exit. In some examples, building 102 may be divided into logical zones. Each zone may include a subset of one or more rooms or other areas of building 102. In some systems 100, zones of building 102 are defined or otherwise associated with areas of building 102 that are under separate HVAC control.

[0023] Building HVAC 104 refers to the heating, ventilation, and air conditioning systems in building 102 used to heat, ventilate, and / or cool areas of building 102. In addition to temperature and humidity control, building HVAC 104 can also be used to improve the quality of environmental conditions inside building 102, such as removing smoke, odors, dust, airborne bacteria, carbon dioxide, and performing oxygen replenishment.

[0024] HVAC controller 106 includes one or more devices configured to direct building HVAC system 104 to control the environment of building 102. HVAC controller 106 may include one or more processors configured to execute instructions, commands, and other routines that support the processes described herein. Such instructions and other data may be maintained in a non-volatile manner using various types of computer-readable storage media. Data may be loaded from the storage media into memory to be accessed by the processor. Computer-readable storage media (also referred to as processor-readable media or storage devices) include any non-transitory media (e.g., tangible media) that contributes to providing instructions or other data that can be read by the processor of the telematics controller. Computer-executable instructions may be compiled or interpreted from a computer program created using a variety of programming languages ​​and / or technologies, including, without limitation, Java, C, C++, C#, Objective C, Fortran, Pascal, JavaScript, Python, Perl, and PL / SQL, singly or in combination.

[0025] Occupancy sensors 108 may include various types of devices configured to detect the presence or absence of occupants in rooms of building 102. As some examples, occupancy sensors 108 may include one or more of ultrasonic sensors, passive infrared (PIR) sensors, or red, green, and blue (RGB) cameras. As a specific example of a motion sensor, a depth sensor (e.g., a Kinect sensor such as the Microsoft Xbox One) may be used, for example, mounted on a ceiling near a doorway, to estimate occupancy. The depth sensor may be used in combination with an occupancy estimator algorithm (e.g., a fine-grained occupancy estimator using Kinect (FORK)). In such an example, real-time depth frames from occupancy sensors 108 placed at each doorway are provided to the algorithm, allowing the algorithm to estimate and update the occupancy count of the room (or area) in response to occupants entering or exiting through the doorway.

[0026] The building thermal model 110 may be a dynamic model that describes how the thermal state of the building 102 changes. These changes may be based on factors such as HVAC power and the number of occupants. An example model is shown in the equation below:

[0027]

[0028] in:

[0029] It is a region i Time slot k The state vector at the end (including temperature, surface and internal nodes);

[0030] It means in the time slot k column vector of external (e.g., outside temperature and solar gain) and internal (e.g., occupant) gain disturbance vectors for the period;

[0031] It is a sign n Time slots in each region k The length of the indoor temperature at the end is n Column vector of ; and

[0032] A 、 B u 、 B d and C is a fixed parameter matrix specified by the building configuration.

[0033] It is important to note that these fixed parameters not only describe how the state of a region i is affected by the state, action, and disturbance of the region, but also take into account the heat exchange between regions. For example, Can be 、 and The linear model is determined by , where the parameters are matrices A 、 B u and B d No. i OK. Some elements of represent the heat load due to heat dissipation by occupants, while other elements describe the heat exchange with the surrounding environment. Assume that each occupant dissipates the same amount of heat, and the formula for indoor human heat dissipation is ,in is the heat dissipated by the occupant in a time slot, and It is a time slot k in i The actual number of occupants in the area. For areas that do not directly exchange heat with the environment, B d The corresponding elements in can be set to zero.

[0034] The occupant counting prediction model 112 is configured to receive occupant count information from the occupancy sensor 108 and predict the next N Time slot area i . In the example, the occupant count is a forced function that does not depend on any state variables other than time. The impact of the occupants is dynamically manifested as internal heat gain in the building's thermal state. By using the occupant count prediction model 112 to accurately predict the number of occupants within a timeframe, the HVAC controller 106 can make intelligent decisions to offset this disturbance and stabilize the thermal conditions of the building 102. As described herein, the statistical model can be trained on real-world data that can be used to predict occupancy changes such that the current time slot is the same as the current time slot. t At the end of the period, and using the previous occupancy counts, the model 112 can predict the next N Time slot area i This is denoted in this paper as .

[0035] Figure 2 An example of occupancy dynamics is shown as a time-inhomogeneous Markov chain unfolding over a day. For the baseline, occupancy dynamics can be modeled using a time-inhomogeneous Markov chain, where states represent occupant counts and transitions between states refer to occupancy changes over time. Given the nature of occupancy, the Markov chain is naturally time-inhomogeneous: the probability distribution of the next state depends on the current state and the current time. Therefore, in time slot t , given the current count m1 , the future occupant count for the next time slot is m 2 The probability is:

[0036] .

[0037] In the example, a fixed time slot duration of 30 minutes and occupant counts ranging from 0 to 50 (maximum count) can be used so that the time-inhomogeneous Markov chain can be unfolded in time, as Figure 2 , where the occupancy state for a whole day is represented as a chain of 48 states. Here, it is assumed that the state dynamics do not differ from day to day. Therefore, the structure of the chain can repeat itself after 48 states.

[0038] Using historical occupancy data to estimate transition probabilities, we can make N-step predictions by looking at the expected future counts N steps later:

[0039]

[0040] Some states may not appear once in the training data, but appear after training. In this case, the closest state transition can be used.

[0041] Markov chains require that the occupancy dynamics obey assumptions such as the Markov property, which may not hold in practice. Therefore, we can further implement a linear regression model to make predictions, such as the following:

[0042]

[0043] exist t + 1 can be fed back into the regression function to predict the t + 2 occupied. Expand N After the step, it can be predicted N future occupant counts. In this model, there are T coefficient and bias terms as parameters to be trained.

[0044] Figure 3 The figure shows the proposed network that combines temporal and contextual features to predict the next N occupancy counts. Since occupancy dynamics are inherently nonlinear, as the period of data that can be collected increases, statistical models with higher representative power can be used while generalizing well without overfitting.

[0045] With a neural network, prediction can be turned into a regression problem. The recurrent part of the neural network is a function that takes as input the occupant counts in the T previous time slots and concatenates them to the final N neurons—which represent N The future occupancy count value of the time slot. Figure 3 As shown in the rectangular box in , it consists of two hidden layers with 32 units (i.e., neurons) each and rectified linear unit (ReLU) activation. (It should be noted that this neural network configuration is only an example, and neural networks utilizing a different number of neurons and / or a different number of layers can be used. In addition to sequential occupant counts, other contextual information such as the time of day and day of the week in the prediction can also be utilized. Therefore, the fully connected layer can combine the categorical features Transformed into a one-hot encoding (e.g., the time of day feature is discretized into 24 columns per hour) and concatenated again to the final N units. So, Figure 3 The pyramid-like architecture in illustrates how temporal and contextual information can be combined to make a final occupancy prediction at the top level. The architecture is flexible to incorporate more specific features that may be available in different applications and uses of the building 102 space.

[0046] The occupancy-aware model predictive HVAC controller 106 has two goals: minimize HVAC energy consumption and optimize occupant thermal comfort. t At the beginning, the model predictive HVAC controller 106 considers the future N The HVAC control is performed for each time slot and the HVAC power sequence within this range is calculated to optimize these two objectives.

[0047] Regarding the first goal, the occupancy-aware HVAC controller 106 can reduce energy waste based on occupancy predictions, especially when zero occupancy is correctly predicted during working hours—for example, when occupants arrive late in the morning or leave early in the evening. During time slot k, the HVAC power of a zone i can be constrained as follows:

[0048]

[0049] in and denote the cooling and heating capacities of the building HVAC 104, respectively. If φ is positive, it indicates that the HVAC system is heating the zone; otherwise, the building HVAC 104 is cooling the zone. Refers to the amount of heat flux from the building HVAC 104 acting on the area. The value will imply more HVAC energy consumption, so The L1 norm of is used as a proxy for energy consumption. The total power consumption can be expressed as .

[0050] Regarding the second goal, the occupant-aware HVAC controller 106 can optimize the thermal comfort of the occupants. Predicted Mean Vote (PMV) is a common comfort measurement method that is standardized in ISO 7730. The PMV model estimates the average occupant comfort level using the following function:

[0051]

[0052] in

[0053] M is the metabolic rate of the occupant;

[0054] T a is the temperature;

[0055] T r is the mean radiant temperature (set equal to T a );

[0056] v is the relative air velocity;

[0057] P a is the relative humidity; and

[0058] I cl is the occupant's clothing insulation factor.

[0059] The range of PMV may be defined herein as between -3 (indicating cold) and 3 (indicating hot), with 0 being neutral. However, this is arbitrary, and other scales may be used. To simplify the description, let To describe the time slot k Central region i individual occupant comfort.

[0060] In order for the HVAC to work efficiently, weighted occupant comfort is taken into account, which means that: (i) if a zone of the building 102 is occupied, the thermal comfort of the occupants is ensured, and (ii) the more occupants are present in the zone, the more comfortable the indoor environment becomes. Therefore, the goal of providing group occupant comfort can be formulated as .

[0061] There is a trade-off between these two objectives. For example, if the outside weather is cold, for example, below 0°C, then in order to make the occupants comfortable, the HVAC should heat the zone. However, this will incur high energy costs. Therefore, a weight is used to sum up and balance the two objectives, and the model predictive HVAC controller 106 problem is thus formulated as follows:

[0062] (1)

[0063] in Constrain each region i HVAC power. Represent the above equation, where the actual occupant count is However, at the beginning of time slot t, for the controller 106, the actual occupant count is unknown. Therefore, the predicted occupant count is used as an input parameter to the controller 106 to determine the HVAC power, and the model predictive HVAC control problem is labeled ,in In the example, three predictors can be used: a time-inhomogeneous Markov chain, a linear regression, and a sequential and contextual neural network to predict t based on historically collected data and real-time occupancy information at the beginning of time slot t. Since the objective and constraints are convex functions, they can be solved by existing solvers, such as Gurobi used in OpenBuild.

[0064] Based on the previous description, the predictive HVAC controller 106 uses the occupancy-aware model predictor 112 to determine the occupancy counts for each zone of the building 102 based on the predicted occupancy counts for the future time slots. i However, due to the uncertainty of occupancy patterns, the performance may be affected by the prediction accuracy.

[0065] To analyze this performance, we define mispredictions and mispredictions of occupant counts. A misprediction refers to a predictor incorrectly predicting the occupied or unoccupied state of a future time slot, and a misprediction of occupant counts refers to a predictor incorrectly estimating the number of occupants in a region of a future time slot. Therefore, two metrics can be used to measure the performance of a predictor: misprediction type distribution and prediction error. Mispredictions can be classified into one of four categories: false positives, false negatives, true positives, and true negatives, which are shown in Table 1.

[0066]

[0067] If both the true value and the predicted value of occupancy are occupied, the misprediction type is labeled as a true positive. If both the true value and the predicted value of occupancy are unoccupied, the misprediction type is labeled as a true negative. If the predictor incorrectly predicts occupied as unoccupied or unoccupied as occupied, the misprediction is recorded as a false negative or false positive, respectively. The prediction error can be used to measure the misprediction of the occupant count and is equal to the absolute value of the difference between the predicted number of occupants and the ground truth.

[0068] Figure 4 Plotting the performance of a time-inhomogeneous Markov chain predictor. Figure 5 Plotting the performance of the linear regression predictor. Figure 6 The performance of sequential and contextual neural network predictors is shown. Because correctly predicting the occupancy state does not affect control performance, the combination of true positives and true negatives is true. It can be observed that each predictor has a relatively higher probability of incorrectly predicting the occupancy index at the beginning and end of the day than at the middle of the day. Meanwhile, during the working hours of the day, the predicted occupancy index is always true, while the predictor error is small at the beginning and end of the day and large during working hours.

[0069] Figure 7 The average occupant count for one day over six months is illustrated. As can be seen, during periods of large misprediction probability, there are few occupants in the area, which is a possible reason why the predictor mispredicts occupant status with a higher probability than during periods of high occupant density. Figure 7 The monthly variations in occupancy patterns in the example data can be explained by seasonal changes in staffing.

[0070] Based on the prediction accuracy of the three predictors, the performance of the occupancy-aware HVAC controller 106 with any of the three predictors can be identified. To measure the performance, a measurement metric can be defined as energy efficiency multiplied by effective PMV improvement to measure the combination of energy efficiency and how much occupant comfort is provided. To calculate the time slot k Energy efficiency and effective PMV improvements: (i) based on time slots k Actual detection of occupant counts in and the initial building state , solve a time-slot version of Equation (1) to determine the optimal HVAC power and comfort value ; and (ii) calculating the energy efficiency and effective PMV improvement as follows:

[0071]

[0072]

[0073] in F are the extreme values ​​of PMV (e.g., -3 for cool and 3 for heat). Using the measurement metric, if the HVAC consumes energy when the zone is unoccupied, its efficient energy consumption ratio is 0 because the optimal energy consumption is 0, and if there are occupants, the control performance is determined by the PMV improvement and energy efficiency of the model-predicted HVAC control using real occupant counts.

[0074] Figure 8 An example of the performance of the HVAC controller 106 using ground truth and neural network predicted occupant counts is shown. As shown, the example lasts for a period of four days. The neural network can be a sequential and contextual neural network as mentioned above.

[0075] As described above, HVAC control has two objectives: minimizing energy consumption and maximizing group occupant comfort, and there is a trade-off between these two objectives. If there are no occupants in a time slot, the HVAC controller 106 should aim to minimize energy consumption and consider a weighted sum of minimizing energy consumption and group occupant comfort when there are occupants in the time slot. However, due to mispredictions, the HVAC controller 106 uses the predicted occupant count to optimize HVAC control with an incorrect objective. Therefore, to handle mispredictions, especially random false positives / negatives, the HVAC controller 106 should determine an HVAC power that is robust to mispredictions and incurs minimal misprediction costs.

[0076] Figure 9 An example adaptive model predictive HVAC control architecture is illustrated. To make the control decisions of the HVAC controller 106 more robust to random mispredictions, the HVAC controller 106 may employ such an architecture to adapt the misprediction type distribution and corresponding misprediction costs in different time slots.

[0077] The control diagram is shown as a closed control loop. As shown, the HVAC controller 106 determines the HVAC power, which further changes the temperature of the manipulated airflow and the building thermal state. Occupancy sensors 108 sense occupant events and calculate actual occupant counts, which are used by the predictor 112 to predict the number of future occupants. Environmental sensors in the building 102 detect and output real-time building thermal conditions. This information from the occupant count predictor 112 and environmental sensors in the building 102 is forwarded as feedback to the HVAC controller 106 to improve control efficiency. Specifically, for an upcoming time slot, the HVAC controller 106 generates a misprediction type distribution by sampling historical predictions and actual occupant count information. Based on the misprediction type distribution and the predicted occupant count information for a given time slot, the HVAC controller 106 determines the HVAC power by minimizing the expected misprediction cost.

[0078] Regarding misprediction classification and distribution, the misprediction type distribution can be updated at different time slots of a day based on the real-time collected prediction and actual occupant count data. N For each time slot, the historical prediction and true occupancy status data are sampled at the same time of the day, and the frequencies of different types of mispredictions are counted. 、 、 and Time slots k Area i There are probabilities of misprediction types true negative, false positive, false negative, and true positive in . The distribution of these four types of misprediction can be forwarded to the misprediction cost optimization part.

[0079] It is worth noting that when there is a large probability of false positive / false negative mispredictions due to incorrect weights between the two targets, computing The HVAC controller 106 may be sensitive to mispredictions. Therefore, the misprediction cost optimization portion aims to determine the HVAC power based on the probability of correctly or incorrectly predicting the occupancy state for a given predicted state to minimize the misprediction cost expectation.

[0080] Recall that there is a time slot k Central region i True occupant count The model predictive HVAC control target is defined as:

[0081]

[0082] Significantly, in the time slot k At the beginning, the computation only has the future N time slots and nPredicted occupant counts for each region.

[0083] make and When using Solution area i and time slots k The energy consumption and PMV of equation (1) are given here. and The control target value is defined as

[0084]

[0085] Then the misprediction cost can be defined as .

[0086] area i and region k The expected misforecast cost of can be defined as:

[0087] (2)

[0088] Based on the predicted occupancy state information, mathematical equations for these different misprediction cost expectations can be defined. 、 、 and You can refer to the misprediction costs for the different types of mispredictions (false positives, false negatives, true positives, and true negatives, respectively). These can be calculated as described in the paper.

[0089] As shown in Table 1, for both true negative and false positive mispredictions, there is no occupant in the area, so the misprediction cost should be wasted energy, which can be expressed as:

[0090] (3)

[0091] For true positive predictions, the control decision The target value determined is , where the predicted value is used as the actual value because the predictor makes a true positive prediction. If the occupancy state is correctly predicted as occupied, then the optimal HVAC power in zone i during time slot k is also affected by the other Therefore, in the case of true positive prediction, the optimal target value for region i and time slot k can be defined as The expected optimal target value for each possible situation. ,make is the probability that the situation exists. Given a situation with some correct or incorrect prediction information, the problem formula can be determined and the optimal target value can be calculated, which uses This leads to = true positives), where is an indicator function that shows the type of misprediction. The misprediction cost can be defined as:

[0092] = true positive) (4)

[0093] In the above function, for n regions and N Given the prediction and misprediction distributions for time slots, is a precomputed constant value, and Also with A linear function of . Therefore, yes A linear function of .

[0094] For false negative predictions, , which uses , because the predictor infers the state to be unoccupied. Time slot k and region i The optimal target value of is also an expected optimal target value. The equation can be defined as:

[0095] =false negative)(5)

[0096] This equation (5) is also related to the HVAC power The relevant linear equation, .

[0097] exist N time slots and n Over regions, the total misprediction cost is expected to be minimized, defined as According to equation (2), and and or and Based on the previous definition of these four variables, they are linearly related to the control decision variables Therefore, the objective function is linearly related to HVAC power. Therefore, the problem of determining HVAC power to minimize the total misprediction cost can be formulated as follows:

[0098]

[0099] Since the objective function of the total misprediction cost minimization problem is linear in the decision variables, and since all constraints are also linear, Equation (6) is convex and can be solved using a convex optimizer.

[0100] The pseudo code for the adaptive model predictive HVAC controller 106 algorithm is shown in Algorithm 1. t At the beginning of the process, the building thermal state is updated and n The future of the region N Then, the misprediction classification and distribution are partially updated N time slots and n The probability of different types of misprediction for each zone is calculated. Based on the predicted occupant counts and the misprediction type distribution, the objective of minimizing the expected total misprediction cost is updated. Equation (6) can then be solved to determine the optimal HVAC power that is robust to random misprediction.

[0101] Based on empirical performance analysis using the predictor and control of the HVAC controller 106, occupancy-aware control shows very similar performance during time slots with small misprediction probabilities compared to using true occupant counts. However, for time slots with large misprediction probabilities, it is difficult to estimate the occupancy state with high accuracy due to the small number of true occupants. Here, the probability-based adaptive HVAC controller 106 approach essentially provides a probabilistic guarantee: if ϵ percent of all misprediction type data follows the sampled misprediction type distribution, then our solution can minimize the total misprediction cost for ϵ percent of future time slots. Due to the law of large numbers, using long-term occupancy data, ϵ percent of time slot mispredictions are close to the true distribution. Therefore, this approach can minimize the expected misprediction cost over time.

[0102] Figure 10 An example process 1000 is illustrated for controlling HVAC operations of a building 102 to minimize energy consumption using predicted occupant counts and accounting for misprediction costs. In an example, process 1000 may be performed by operation of HVAC controller 106, as discussed in detail herein.

[0103] At operation 1002, the HVAC controller 106 updates the thermal state of the plurality of zones of the building 102 based on the building thermal model 110 and information received from the temperature sensors of the building 102. In an example, the building thermal model 110 is a dynamic model that describes how the building thermal state changes based on the building interior and building exterior temperature disturbances during a plurality of time slots and the heat exchange configurations of the plurality of zones of the building 102.

[0104] At operation 1004, the HVAC controller 106 uses the actual occupancy counts for each of the plurality of zones to update the predicted occupant counts for the upcoming plurality of time slots for each of the plurality of zones. In an example, the HVAC controller 106 may use occupancy sensors 108 near each doorway of the building to determine the current occupancy counts, such that real-time depth frames from the occupancy sensors 108 are provided to estimate the occupancy counts for the zones in response to occupants entering or exiting through the doorways. The predicted occupant counts may be modeled using a predictor 112 that includes one or more of: a time-inhomogeneous Markov chain in which each state represents an occupant count, transitions between states represent occupancy changes over time, and a probability distribution for the next state depends on the current state and the current time; a linear regression model that uses a regression function to predict the next occupancy count from the current occupancy count; or a recurrent neural network that combines temporal and contextual features to predict the next occupancy count based on the current occupancy count.

[0105] At operation 1006, the HVAC controller 106 updates, for each of the plurality of zones, a misprediction type distribution for the upcoming plurality of time slots, the misprediction type distribution indicating mispredictions of true negatives, false positives, false negatives, and true positives. For example, the misprediction type distribution for the plurality of time slots may be updated for each of the plurality of zones based on a predicted occupant count and an actual occupancy count for the plurality of time slots as subsequently measured.

[0106] At operation 1008, the HVAC controller 106 updates the total misprediction cost expectation based on the predicted occupant count and the misprediction type distribution. In an example, the misprediction cost expectation is updated for a plurality of time slots in each of the plurality of zones using a set of linear equations relating HVAC power to true negatives, false positives, false negatives, and true positives.

[0107] At operation 1010, the HVAC controller 106 determines an HVAC power for each of the plurality of zones to optimize occupant thermal comfort weighted by the predicted occupant count while minimizing an expected total misprediction cost. This determination may be subject to factors such as the heat exchange configuration of the building 102 and the minimum and maximum HVAC power available for each of the plurality of zones.

[0108] At operation 1012, the HVAC controller 106 controls the HVAC operation of the building 102 based on the HVAC power determined for each of the plurality of zones. After operation 1012, the process 1000 ends. However, it should be noted that the process 1000 is iterative, and some or all of the described operations of the process 1000 may be performed simultaneously and / or may be repeated in a loop-by-loop manner to control the operation of the HVAC system over time.

[0109] Therefore, the real-time occupant counts received from occupancy sensors 108 and the predicted future occupant counts in each zone of building 102 determined using predictive model 112 are useful for providing efficient HVAC control. On the one hand, the accurate binary occupancy state, such as whether occupied or not, informs HVAC controller 106 when to turn building HVAC 104 on or off. On the other hand, due to the heat emissions of occupants themselves, HVAC controller 106 can use occupant counts to save energy and perform precise heating and cooling accordingly. In this framework, occupancy sensors 108 can collect real-time occupant counts for different zones of building 102, where this data set can be used to design predictive model 112 and create a predictive HVAC controller 106.

[0110] For the HVAC controller 106 based on model predictive control (MPC), an MPC-based optimization problem is solved to determine the HVAC power in the future time slot. The MPC controller 106 discretizes the time line into multiple time slots and t Consider the future at the beginning N A time slot can be used to control the power of k To index, ( For modeling purposes, it can be assumed that there is a n regions, and Refers to the length of n The column vector of n Time slots in each region k Heating / cooling power during the period.

[0111] Figure 11An example computing device 1100 is illustrated for performing various aspects of HVAC operations described in detail herein. Algorithms and / or methods of components of system 100, such as HVAC controller 106, may be implemented using such a computing device 1100. Computing device 1100 may include memory 1102, a processor 1104, and non-volatile storage 1106. Processor 1104 may include one or more devices selected from a high-performance computing (HPC) system, including a high-performance core, a microprocessor, a microcontroller, a digital signal processor, a microcomputer, a central processing unit, a field programmable gate array, a programmable logic device, a state machine, a logic circuit, an analog circuit, a digital circuit, or any other device that manipulates (analog or digital) signals based on computer-executable instructions residing in memory 1102. Memory 1102 may include a single memory device or multiple memory devices, including, but not limited to, random access memory (RAM), volatile memory, non-volatile memory, static random access memory (SRAM), dynamic random access memory (DRAM), flash memory, cache memory, or any other device capable of storing information. Non-volatile storage 1106 may include one or more persistent data storage devices such as a hard drive, an optical drive, a tape drive, non-volatile solid-state devices, cloud storage, or any other device capable of persistently storing information.

[0112] The processor 1104 can be configured to read and execute computer-executable instructions residing in the program instructions 1108 of the memory 1102 and embodying the algorithms and / or methods of one or more embodiments of the present invention, which are stored in the non-volatile storage device 1106. The program instructions 1108 may include an operating system and applications. The program instructions 1108 may be compiled or interpreted from a computer program created using a variety of programming languages ​​and / or technologies, including, without limitation, Java, C, C++, C#, Objective C, Fortran, Pascal, JavaScript, Python, Perl, and PL / SQL, either singly or in combination.

[0113] When executed by processor 1104, the computer-executable instructions of program instructions 1108 may cause computing device 1100 to implement one or more algorithms and / or methods disclosed herein. Non-volatile storage 1106 may also include data 1110 that supports the functions, features, and processes of one or more embodiments described herein.

[0114] The processes, methods, or algorithms disclosed herein may be deliverable to / implemented by a processing device, controller, or computer, which may include any existing programmable electronic control unit or dedicated electronic control unit. Similarly, the processes, methods, or algorithms may be stored in a variety of forms as data and instructions executable by a controller or computer, including but not limited to information permanently stored on non-writable storage media such as ROM devices and information reversibly stored on writable storage media such as floppy disks, magnetic tapes, CDs, RAM devices, and other magnetic and optical media. The processes, methods, or algorithms may also be implemented in software executable objects. Alternatively, the processes, methods, or algorithms may be embodied in whole or in part using suitable hardware components, such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), state machines, controllers, or other hardware components or devices, or a combination of hardware, software, and firmware components.

[0115] While exemplary embodiments have been described above, it is not intended that these embodiments describe all possible forms encompassed by the claims. The words used in the specification are descriptive, not restrictive, and it should be understood that various changes may be made without departing from the spirit and scope of the present disclosure. As previously described, features of the various embodiments may be combined to form further embodiments of the present invention that may not be explicitly described or illustrated. Although various embodiments may have been described as providing advantages or being preferred over other embodiments or prior art implementations in terms of one or more desired characteristics, those skilled in the art recognize that one or more features or characteristics may be compromised to achieve the desired overall system properties, depending on the specific application and implementation. These properties may include, but are not limited to, cost, strength, durability, lifecycle cost, marketability, appearance, packaging, size, applicability, weight, manufacturability, ease of assembly, and the like. As such, to the extent that any embodiment is described as less desirable than other embodiments or prior art implementations with respect to one or more characteristics, such embodiments are not outside the scope of the present disclosure and are desirable for the particular application.

Claims

1. A method for controlling heating, ventilation and air conditioning (HVAC) operations of a building to minimize energy consumption using predicted occupant counts and accounting for misprediction costs, the method comprising: updating a thermal state of a plurality of zones of a building based on a thermal model of the building and information received from temperature sensors of the building; updating a predicted occupant count for an upcoming plurality of time slots for each of the plurality of regions using an actual occupancy count for each of the plurality of regions; updating, for each of the plurality of regions, a misprediction type distribution for the upcoming plurality of time slots, the misprediction type distribution indicating mispredictions of true negatives, false positives, false negatives, and true positives; updating the total misprediction cost expectation based on the predicted occupant count and the misprediction type distribution; determining an HVAC power for each of the plurality of zones to optimize occupant thermal comfort weighted by predicted occupant counts while minimizing an expected total misprediction cost, the determining step being constrained by a heat exchange configuration of the building and minimum and maximum HVAC powers available to each of the plurality of zones; and HVAC operation of the building is controlled based on the HVAC power determined for each of the plurality of zones.

2. The method according to claim 1, wherein: The building thermal model is a dynamic model that describes how the building thermal state changes based on building interior and building exterior temperature disturbances and heat exchange configurations of multiple zones of the building during the multiple time slots.

3. The method of claim 1 , further comprising determining actual occupancy counts using occupancy sensors proximate each doorway of the building, such that real-time depth frames from the occupancy sensors are provided to an occupancy estimator algorithm that estimates occupancy counts for areas in response to occupants entering or exiting through the doorways.

4. The method of claim 1 , further comprising modeling the predicted occupant count using one or more of: A temporally nonuniform Markov chain where each state represents an occupant count, transitions between states refer to occupancy changes in time, and the probability distribution of the next state depends on the current state and the current time; A linear regression model that uses a regression function to predict the next occupancy count from the current occupancy count; or A recurrent neural network that combines temporal and contextual features to predict the next occupancy count from the current occupancy count.

5. The method of claim 1, further comprising updating a misprediction type distribution of the plurality of time slots for each of the plurality of regions based on predicted occupant counts and actual occupancy counts of the plurality of time slots as subsequently measured.

6. The method of claim 1 further comprising updating misprediction cost expectations for a plurality of time slots for each of the plurality of zones using a system of linear equations relating HVAC power to each of true negatives, false positives, false negatives, and true positives.

7. A system for controlling heating, ventilation and air conditioning (HVAC) operations of a building to minimize energy consumption using predicted occupant counts and accounting for misprediction costs, the system comprising: An HVAC controller, the HVAC controller is programmed to: updating a thermal state of a plurality of zones of a building based on a thermal model of the building and information received from temperature sensors of the building; updating a predicted occupant count for an upcoming plurality of time slots for each of the plurality of regions using an actual occupancy count for each of the plurality of regions; updating, for each of the plurality of regions, a misprediction type distribution for an upcoming plurality of time slots, the misprediction type distribution indicating mispredictions of true negatives, false positives, false negatives, and true positives; updating the total misprediction cost expectation based on the predicted occupant count and the misprediction type distribution; determining an HVAC power for each of the plurality of zones to optimize occupant thermal comfort weighted by predicted occupant counts while minimizing an expected total misprediction cost, the determining step being constrained by a heat exchange configuration of the building and minimum and maximum HVAC powers available to each of the plurality of zones; and HVAC operation of the building is controlled based on the HVAC power determined for each of the plurality of zones.

8. The system according to claim 7, wherein: The building thermal model is a dynamic model that describes how the building thermal state changes based on building interior and building exterior temperature disturbances and heat exchange configurations of multiple zones of the building during the multiple time slots.

9. The system according to claim 7, wherein: The HVAC controller is further programmed to determine actual occupancy counts using occupancy sensors near each doorway of the building, such that real-time depth frames from the occupancy sensors are provided to an occupancy estimator algorithm that estimates occupancy counts for areas in response to occupants entering or exiting through the doorways.

10. The system according to claim 7, wherein: The HVAC controller is further programmed to model the predicted occupant count using one or more of the following: A temporally nonuniform Markov chain where each state represents an occupant count, transitions between states refer to occupancy changes in time, and the probability distribution of the next state depends on the current state and the current time; A linear regression model that uses a regression function to predict the next occupancy count from the current occupancy count; or A recurrent neural network that combines temporal and contextual features to predict the next occupancy count from the current occupancy count.

11. The system according to claim 7, wherein: The HVAC controller is further programmed to update the misprediction type distribution of the plurality of time slots for each of the plurality of zones based on the predicted occupant count and the actual occupancy count of the plurality of time slots as subsequently measured.

12. The system according to claim 7, wherein: The HVAC controller is further programmed to update misprediction cost expectations for a plurality of time slots for each of the plurality of zones using a system of linear equations relating HVAC power to each of true negatives, false positives, false negatives, and true positives.

13. A non-transitory computer readable medium comprising instructions for controlling heating, ventilation, and air conditioning (HVAC) operations of a building to minimize energy consumption using predicted occupant counts and accounting for misprediction costs, the instructions, when executed by an HVAC controller, cause the HVAC controller to: updating a thermal state of a plurality of zones of a building based on a thermal model of the building and information received from temperature sensors of the building; updating a predicted occupant count for an upcoming plurality of time slots for each of the plurality of zones using an actual occupancy count for each of the plurality of zones; updating, for each of the plurality of regions, a misprediction type distribution for an upcoming plurality of time slots, the misprediction type distribution indicating mispredictions of true negatives, false positives, false negatives, and true positives; updating the total misprediction cost expectation based on the predicted occupant count and the misprediction type distribution; determining an HVAC power for each of the plurality of zones to optimize occupant thermal comfort weighted by predicted occupant counts while minimizing an expected total misprediction cost, the determining step being constrained by a heat exchange configuration of the building and minimum and maximum HVAC powers available to each of the plurality of zones; and HVAC operation of the building is controlled based on the HVAC power determined for each of the plurality of zones.

14. The medium according to claim 13, wherein The building thermal model is a dynamic model that describes how the building thermal state changes based on building interior and building exterior temperature disturbances and heat exchange configurations of multiple zones of the building during the multiple time slots.

15. The medium of claim 13, further comprising instructions that, when executed by the HVAC controller, cause the HVAC controller to determine actual occupancy counts using occupancy sensors proximate each doorway of a building, such that real-time depth frames from the occupancy sensors are provided to an occupancy estimator algorithm that estimates occupancy counts for areas in response to occupants entering or exiting through doorways.

16. The medium of claim 13, further comprising instructions that, when executed by the HVAC controller, cause the HVAC controller to model the predicted occupant counts using one or more of: a time-inhomogeneous Markov chain in which each state represents an occupant count, transitions between states refer to occupancy changes in time, and a probability distribution for a next state depends on a current state and a current time; A linear regression model that uses a regression function to predict the next occupancy count from the current occupancy count; or A recurrent neural network that combines temporal and contextual features to predict the next occupancy count from the current occupancy count.

17. The medium of claim 13, further comprising instructions that, when executed by the HVAC controller, cause the HVAC controller to update a misprediction type distribution for the plurality of time slots for each of the plurality of zones based on predicted occupant counts and actual occupancy counts for the plurality of time slots as subsequently measured.

18. The medium of claim 13, further comprising instructions that, when executed by the HVAC controller, cause the HVAC controller to update misprediction cost expectations for a plurality of time slots for each of the plurality of zones using a set of linear equations relating HVAC power to each of true negatives, false positives, false negatives, and true positives.

Images