Systems and methods for controlling the movement of an elevator group

Abstract

A control system for controlling the movement of an elevator in an elevator group is based on a partial trajectory of a person and uses a neural network trained for extended destination prediction of the person to generate a polynomial of the extended destination prediction. The polynomial has at least two dimensions, the at least two dimensions including a first dimension of the destination of the person and a second dimension of the time interval for the person to reach the destination of the first dimension. The control system optimizes the scheduling of the elevator group based on the polynomial and further controls the elevator group according to the scheduling.

Description

Technical Field

[0001] The present disclosure generally relates to elevator destination prediction and scheduling, and more particularly to a system and method for controlling the movement of an elevator group. Background Art

[0002] Today, elevators are often used daily in office and residential buildings. The elevators in an elevator group (also referred to as "cars") are coupled to a control system that determines the destination of the elevator and controls the scheduling of the elevator.

[0003] Some control systems are configured to predict the destination of an elevator. Such predicted destinations include information about the future arrival of passengers at the elevator. Based on the predicted destination, the elevator is scheduled. Generally, scheduling a group of elevators (also referred to as group elevator scheduling (GES)) is a combinatorial optimization problem for an elevator group having two or more cars. The goal of the most common instances of this problem is to assign elevator cars to passengers who request an elevator car by pressing an up or down button. Upon receiving a request, the scheduler assigns a car to each passenger such that a performance metric (e.g., average waiting time (AWT) for all passengers) is minimized. AWT is defined as the length of the time interval from the moment a passenger makes a request until the car arrives, averaged over many requests of the passenger. To this end, various scheduling methods have been identified.

[0004] Most existing schedulers only serve known requests and destinations, that is, they completely ignore the future arrival of passengers. Some recently proposed schedulers use algorithms that take into account the statistics of the future arrival of passengers, see, for example, U.S. Patent 9,834,405. However, statistical estimation of the future arrival time of passengers is laborious and generally does not provide an accurate representation of the actual future requests for elevator service.

[0005] Therefore, there is a need for a system that can more accurately determine the future arrival of passengers for scheduling elevators. Summary of the Invention

[0006] Some embodiments aim to provide systems and methods for predicting the destinations of future passengers and scheduling elevator cars in a group elevator system (GES), and more particularly, to determine prediction information regarding the arrival time of future passengers at any floor of a building. In addition, some embodiments aim to determine the scheduling of a GES including a group of elevator cars based on the prediction information, where the scheduling optimizes a performance metric by minimizing the average waiting time (AWT) for all passengers.

[0007] Some embodiments conceive the problem of predicting the occurrence of future elevator requests and the time of occurrence of future elevator requests as an extended destination prediction problem. The destination prediction problem aims to estimate the final destination of a trajectory based on a partial observation of the passenger's trajectory. Thus, the destination prediction problem can be used to: estimate a person's final destination based on the person's current movement, and if the final destination is within an elevator request area (such as near an elevator call button or otherwise associated with an elevator request), then consider this person as a future passenger. However, accurate destination prediction is a well-known difficult problem. Moreover, it is not sufficient to only predict whether a person will request elevator service, i.e., it is also necessary to estimate the time of such a future elevator request.

[0008] Therefore, the extended destination prediction problem extends the destination prediction problem by predicting not only the final destination of the trajectory but also the time to reach the final destination. Additionally, the extended destination prediction problem replaces the prediction of whether a person's destination is associated with an elevator request with predictions of various possibilities of the destination, which is computationally simpler as described below and can benefit the design and training of the extended destination prediction.

[0009] Furthermore, some embodiments are based on model-based or deep learning-based methods for solving the extended destination prediction problem. Some embodiments are based on the recognition that AI-based destination prediction for elevator scheduling is difficult. More specifically, it is difficult to make such a change to predict a passenger's destination through AI-based destination prediction because a person (i.e., a passenger) in a typical office environment has a large degree of freedom of movement, and the person's destination choice and incomplete trajectory can lead to various final destinations. Moreover, in an office environment, the annotation of the final destination is spurious, i.e., any final destination may be an intermediate destination and vice versa. In this way, the prediction of whether the final destination of a partial trajectory is an elevator request area is challenging. Additionally, it is difficult to achieve the prediction of the distribution of arrival times (e.g., in a Gauss-Bernoulli distribution) through a neural network because a neural network cannot easily provide the required type of probability distribution as its continuous output.

[0010] To this end, some embodiments are based on the recognition that a neural network is trained to output the result of a destination prediction in the form of a polynomial distribution. In some implementations, the polynomial is two-dimensional, having one dimension that defines the destination and another dimension that defines the arrival time. The two-dimensional polynomial distribution can be visualized as a matrix that has, for example, destination names for columns or rows, time intervals for rows or columns, and the values of the matrix that define the probability of arriving at the corresponding destination during the corresponding time interval. The time interval can be represented by a time instance at the start, middle, or end of the interval. Further, the sum of all the values in the matrix equals "1", which reflects the probabilistic nature of the polynomial distribution. To achieve this, some embodiments define an exhaustive list of destinations to ensure that if a person's movement is detected, the person will have a final destination in the defined list of destinations. More specifically, the area served by the elevator system (e.g., the floors in an office building) is quantified as a tessellation of locations, referred to herein as tessellation elements. Unlike semantic destinations such as a kitchen, a hallway, a meeting room, etc., such quantification ensures an exhaustive list of destinations and better matches various motion sensing methods.

[0011] As used herein, a tessellation is an arrangement of shapes that fit closely together in a pattern without gaps or overlaps. Examples of such shapes include squares, rectangles, polygons, etc. When the tessellation uses squares, such a tessellation can also be referred to as a regular grid, or simply as a grid. When the tessellation uses rectangles of different sizes, such a tessellation can also be referred to as an irregular grid. In different embodiments, the tessellation elements can have the same or different shapes to better fit the layout of the area served by the elevator system.

[0012] In some embodiments, the sensing and tracking of passengers on a floor is achieved by radio antennas placed every few meters in the building. Thus, "location N" or "tessellation element N" indicates that the passenger was detected by antenna number N. If multiple antennas detect a passenger, the number of sensing antennas can be determined by a fixed tie-breaking order of the antennas. In some other embodiments, people are detected by means of other sensors such as camera sensors.

[0013] Additionally, the time dimension is also quantified as a set of time intervals that define a prediction time period (also referred to as the prediction horizon). The quantified time period and granularity can depend on the computing power and / or be defined by application requirements. Thus, the prediction horizon is configurable. For example, in some embodiments, the destination prediction is for the next 10 seconds, and thus, the quantified time period in the polynomial distribution is 10 seconds.

[0014] In some embodiments, training a prediction neural network to output prediction estimates in the form of a polynomial has several advantages. First, the polynomial replaces individual predictions of the probability of reaching an elevator request zone with a joint prediction of reaching any destination (where the elevator request zone is one of the destinations). Thus, when two conditions are met, the probability of reaching the elevator request zone is naturally determined relative to other destinations. The two conditions include that the list of destinations is exhaustive and that the sum of all values in the polynomial equals "1". The first requirement is met through spatial quantization, and the second requirement is implemented during the training of the neural network. In this way, the extended destination prediction of the probability of reaching the elevator request zone is simplified. Second, the time dimension of the polynomial eliminates the need for a continuous output for the arrival time as required in existing systems (e.g., in a Gaussian-Bernoulli distribution). This is because the different values for the elevator request zone determined for different time instances represent samples on a continuous distribution of arrival times with corresponding probabilities defined by the values for the elevator request zone. In fact, the probability distribution of the arrival time can be derived by interpolating the values of the probability of reaching the elevator request zone at different time instances. However, different embodiments are implemented to use the values arriving at different time instances without constructing a probability distribution. This is advantageous because such an implementation does not limit the extended destination prediction to any particular type of distribution.

[0015] Furthermore, the time dimension of the polynomial distribution can define a period of interest for destination prediction. In one embodiment, this period corresponds to the maximum period within which a person will reach the elevator. In another embodiment, such a period is determined by the scheduler. This is advantageous because it enables replacing the prediction of whether and when a passenger will reach the elevator request zone with a prediction of where the passenger will be in the next prediction period. Such a reformulation of the problem simplifies the training of the AI-based extended destination prediction estimator.

[0016] In some embodiments, historical trajectory data can be utilized to train a neural network model and predict the destinations and arrival times of potential future passengers according to partial trajectories of them. In the scheduling task for an elevator car at any floor at the current time, the entire arrival stream considered includes the arrivals of current passengers (who have arrived and issued service requests) and the predicted arrivals of future passengers, where the latter are referred to as the continuation set of the current arrival stream.

[0017] In some embodiments, an elevator group is scheduled by using a continuation set and considering current service requests and future service requests. Such a continuation set is obtained based on the execution of a simulation (e.g., Monte Carlo simulation). Each continuation set includes a combination of future requests and current requests. Individually, each continuation set is deterministic because both the current request and the future requests are considered actual requests. However, collectively, the different combinations of future requests in different continuation sets enable some embodiments to capture the probabilistic nature of future requests.

[0018] To this end, some embodiments are based on the recognition that not all destinations in a polynomial are associated with the area service provided by the elevator. Thus, in such embodiments, the polynomial is filtered based on destinations related to the area served by the elevator group. Based on the filtering, a continuation set is obtained by sampling a polynomial generated from a destination predictor based on destinations related to the area served by the elevator, such that an accurate continuation set is obtained for destinations related to the area served by the elevator.

[0019] In some embodiments, a destination predictor is utilized to determine a polynomial for predicting future arrivals of passengers. Such a destination predictor is implemented according to a deep learning-based method (deep neural network (DNN)). The deep learning-based method has the ability to adapt to non-verbal applications such as destination / trajectory prediction. Examples of DNNs include, but are not limited to, recurrent neural networks (RNNs), long short-term memory (LSTM)-based neural networks, bi-LSTM-based neural networks, and neural networks based on the transformer architecture.

[0020] Some embodiments are based on the recognition that a transformer architecture can facilitate extended destination prediction. The transformer architecture of DNNs is used in various speech-related applications such as language translation, natural language understanding, and document generation. The transformer architecture does not have any recurrent primitive units present in RNN models, but instead relies on more efficient attention modules for sequentially correlating elements. Some embodiments are based on the recognition that extended destination prediction, due to the mapping nature of the transformer architecture, can be viewed as a transcription of a partial trajectory, which makes the transformer architecture suitable for the extended destination prediction task. Thus, in some embodiments, the destination predictor uses a neural network architecture based on transformers. Both the position component and the timing component in the sensed trajectory data are discretized. A continuation set is formed by filtering the predicted destination relative to the elevator platform. To handle the uncertainty in the arrival time, Monte Carlo simulation is applied to generate multiple continuation sets by sampling from a polynomial distribution generated by the neural network. Once a complete arrival stream of current and potential future passengers in the form of multiple continuation sets is obtained, the scheduling of the elevator is determined, e.g., by using the method described in U.S. Patent 9,834,405, a method and system for scheduling elevator cars in a group elevator system with uncertain information about the arrival of future passengers.

[0021] Accordingly, one embodiment discloses a control system for controlling the movement of elevators in an elevator group, the control system comprising: a first input interface configured to receive a current request from a passenger for service provided by the elevator group; a second input interface configured to receive a partial trajectory of the movement of a person moving on a floor in an area served by the elevator group. The control system further comprises a processor configured to: after receiving the partial trajectory of the movement of the person, execute, based on the partial trajectory, a neural network trained for extended destination prediction of the person having the movement to generate a polynomial for the extended destination prediction of the person, wherein the polynomial has at least two dimensions, the at least two dimensions including a first dimension of the destination of the person and a second dimension of the time interval for the person to reach the destination of the first dimension, and wherein the sum of the values of the polynomial is normalized to 1. Further, the floor is quantized to form checkerboard layout elements of the destination of the first dimension, and wherein at least one checkerboard layout element is associated with the area served by the elevator group. The processor is further configured to optimize the scheduling of the elevator group for serving the current request from the passenger and the future requests of the person to obtain elevator service at a time instance provided in the polynomial, with a probability of a corresponding value determined for the checkerboard layout element associated with the area served by the elevator group in the polynomial; and to control the elevator group according to the scheduling.

[0022] In addition, the time interval of the second dimension quantifies the prediction time period, and wherein the prediction time period is the maximum time that the control system considers for future requests of the person.

[0023] In another embodiment, the processor is further configured to sample from a polynomial distribution determined for a checkerboard layout element associated with an area served by an elevator group to generate a plurality of combinations of request times, wherein the occurrence frequency of a particular time instance among the plurality of combinations of request times is the probability of a future request for elevator service within a particular time interval, and the probability has a value at the intersection of the particular time interval and the checkerboard layout element associated with the area served by the elevator group in the polynomial. The actual time of the request is assumed to be the start or midpoint of the time interval of the polynomial distribution. The processor is further configured to: combine each of the plurality of combinations of request times with the current request to generate a plurality of continuation sets; and determine a scheduling of the elevator group to optimize a performance metric for at least some of the passengers in all combinations of the continuation sets.

[0024] In another embodiment, the partial trajectory includes trajectory data containing only position information, wherein each trajectory is represented by a sequence of checkerboard layout indices, wherein the checkerboard layout indices are obtained by discretizing a range of position coordinates, and wherein the range of position coordinates represents longitude and latitude associated with the checkerboard layout element.

[0025] In another embodiment, the partial trajectory includes trajectory data containing both position information and timing information, wherein each trajectory is represented by a sequence of tuples of checkerboard layout indices, and wherein each tuple contains checkerboard layout indices obtained by discretizing a range of position coordinates and a prediction time period.

[0026] In another embodiment, the timing information of the trajectory is extracted by considering relative time with respect to elapsed time, which is relative to their start timestamps.

[0027] In another embodiment, each trajectory is represented by a sequence of symbols, and wherein each symbol is mapped from a tuple of checkerboard layout indices indicating position information and timing information.

[0028] In another embodiment, the continuation sets are formed by filtering the destinations by only considering the area served by the elevator group.

[0029] In another embodiment, the neural network has a transformer architecture, wherein the transformer architecture uses position encoding and attention mechanisms that enable parallelization.

[0030] Another embodiment discloses a method for controlling the movement of elevators in an elevator group, the method comprising: obtaining a current request from a passenger for a service provided by the elevator group; obtaining a partial trajectory of the movement of a person moving on a floor in an area served by the elevator group. The method further comprises, after receiving the partial trajectory of the movement of the person, performing, based on the partial trajectory, a neural network trained for extended destination prediction of the person having the movement to generate a polynomial for the extended destination prediction of the person, wherein the polynomial has at least two dimensions, the at least two dimensions including a first dimension of the destination of the person and a second dimension of the time interval for the person to reach the destination in the first dimension. Further, the floor is quantized to form checkerboard arrangement elements for the destination in the first dimension, and at least one checkerboard arrangement element is associated with the area served by the elevator group. The method further comprises optimizing the scheduling of the elevator group serving the current request from the passenger and the future requests of the person to obtain elevator service at the time interval provided in the polynomial, with the probability of the corresponding value determined by the checkerboard arrangement element associated with the area served by the elevator group in the polynomial; and controlling the elevator group according to the scheduling.

[0031] In another embodiment, the method further comprises performing a Monte Carlo simulation of the arrival process by sampling from the polynomial determined for the checkerboard arrangement element associated with the area served by the elevator group to generate multiple combinations of request times, wherein the occurrence frequency of a particular time interval among the multiple combinations of request times is the probability of a future request for elevator service at the particular time interval, the probability having the value at the intersection of the particular time interval and the checkerboard arrangement element associated with the area served by the elevator group in the polynomial; combining each of the multiple combinations of request times with the current request to generate multiple continuation sets; and determining the scheduling of the elevator group to optimize the performance metric for at least some of the passengers in all combinations of the continuation sets.

[0032] The presently disclosed embodiments will be further illustrated with reference to the accompanying drawings. The drawings shown are not necessarily to scale, the emphasis instead being generally on illustrating the principles of the presently disclosed embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1A Illustrating the environment of a control system for controlling the movement of an elevator group according to an embodiment of the present disclosure.

[0034] Figure 1B Illustrating a block diagram of a control system according to an embodiment of the present disclosure.

[0035] Figure 1CIllustrate the quantization of floors in a checkerboard arrangement element of a regular grid for polynomial distribution according to an exemplary embodiment of the present disclosure.

[0036] Figure 1D Illustrate the quantization of floors in a checkerboard arrangement element of an irregular grid for polynomial distribution according to an exemplary embodiment of the present disclosure.

[0037] Figure 1E Illustrate a graph of the quantization of elapsed time of a trajectory for destination prediction according to an embodiment of the present disclosure.

[0038] Figure 1F Illustrate a Monte Carlo simulation for obtaining a continuation set based on destination prediction according to an embodiment of the present disclosure.

[0039] Figure 1G Illustrate an exemplary representation of a continuation set based on destination prediction according to an exemplary embodiment of the present disclosure.

[0040] Figure 1H Illustrate a schematic diagram of scheduling carriages of an elevator group based on a continuation set according to an exemplary embodiment of the present disclosure.

[0041] Figure 2A Illustrate a block diagram of a destination predictor according to an embodiment of the present disclosure.

[0042] Figure 2B Illustrate a block diagram of a neural network according to another exemplary embodiment of the present disclosure.

[0043] Figure 2C Illustrate the layout of floors in a building according to an embodiment of the present disclosure, where multiple trajectories of multiple pedestrians are used for training a destination predictor, and a partial trajectory of one pedestrian is used for testing the destination predictor.

[0044] Figure 2D Illustrate a plot comparing the probabilities of destination prediction determined by a neural network based on a transformer architecture, a neural network based on long short-term memory (LSTM), and a neural network based on bidirectional long short-term memory (BiLSTM) according to an exemplary embodiment of the present disclosure.

[0045] Figure 2E Illustrate a plot comparing the probabilities of destination prediction determined by a neural network based on a transformer architecture, an LSTM-based neural network, and a BiLSTM-based neural network according to another exemplary embodiment of the present disclosure.

[0046] Figure 3 Illustrate the steps of a method implemented for controlling the movement of an elevator group according to some embodiments of the present disclosure.

[0047] Figure 4 Illustrate the steps of a method implemented for Monte Carlo simulation according to some embodiments of the present disclosure, the method being for determining a schedule for controlling the movement of an elevator group.

[0048] While the above-identified drawings illustrate presently disclosed embodiments, other embodiments are also contemplated as noted in the discussion. The present disclosure presents illustrative embodiments by way of illustration, not limitation. Many other modifications and embodiments that fall within the scope and spirit of the principles of the presently disclosed embodiments can be designed by those skilled in the art. Detailed Description

[0049] In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. However, it will be apparent to one of ordinary skill in the art that the present disclosure may be practiced without these specific details. In other instances, devices and methods are shown only in block diagram form in order to avoid obscuring the present disclosure.

[0050] As used in this specification and the claims, the terms "such as", "for example", and "like" and the verbs "comprising", "having", "including", and their other verb forms when used in conjunction with a list of one or more components or other items will each be construed as open-ended, meaning that the list is not to be considered as excluding other, additional components or items. The term "based on" means at least partially based on. Further, it is to be understood that the language and terminology used herein are for the purpose of description and should not be regarded as limiting. Any headings used within this description are for convenience only and have no legal or limiting effect.

[0051] The present disclosure discloses a control system that provides a solution to the previously described GES problem. The GES control system uses a scheduler to assign elevator cars to passengers so as to minimize a performance metric (e.g., average waiting time (AWT) for passengers). Currently, many schedulers only serve known requests (or currently existing requests) and known destinations to optimize AWT. These requests are typically sent directly by passengers to the GES system using floor call and car call buttons. However, it would be advantageous if the scheduler could determine the schedule for elevator cars by considering future requests for elevator service. The AWT can be significantly minimized by considering information associated with future requests.

[0052] For example, a first passenger on the sixth floor of a building requests service from an elevator group to go to the first floor of the building. Consider that the elevator car assigned to serve the first passenger is currently on the fifteenth floor. Consider that within the next five seconds, a second passenger on the tenth floor will also request service to go to the first floor. In this case, the scheduler determines this information associated with the arrival of the future passenger (i.e., the second passenger) in advance, and the scheduler can determine a dispatch such that the elevator car on the fifteenth floor can wait for five seconds until the future passenger (i.e., the second passenger) requests service to go to the first floor. When receiving a request from the second passenger, the elevator car can first pick up the second passenger from the tenth floor and then pick up the first passenger on the sixth floor. Thus, the two passengers can reach the first floor together. Therefore, it is desirable to have some information about future arrivals to further optimize the AWT of passengers.

[0053] However, there are problems associated with using information about future arrivals for dispatching. It is difficult to predict the probability that a future passenger will issue a service request and the probability of the possible time of the service request. In addition, even if these probabilities are somehow determined or predicted, it is still difficult to use these probabilities for dispatching purposes.

[0054] To address these problems, some schedulers can use algorithms that consider known services (i.e., existing or current service requests), known destinations, and the probability of future passenger arrivals. These schedulers can consider such arrivals by first generating a set of probability distributions for the arrivals of future passengers to any floor of the building, where the set of probability distributions is characterized by probability variables that specify the arrival information of future passengers (i.e., the probability that a future passenger will issue a service request and the probability of the possible time of the service request).

[0055] Some embodiments are based on the understanding that the scheduler can determine arrival information based on the set of probability distributions, where the set of probability distributions (i.e., the arrival information of future passengers) is assumed to follow some known distribution form; for example, a Gaussian - Bernoulli probability distribution. Since the arrivals of future passengers are assumed, the set of probability distributions is inaccurate. Therefore, the scheduler may not be able to effectively optimize the AWT for current and future passengers.

[0056] To address such problems, the present disclosure proposes a control system that predicts the future arrivals of passengers by replacing the statistical estimation of the probability distribution of future arrivals.

[0057] Figure 1A Illustrate the environment of a control system 101 for controlling the movement of an elevator group 109 according to an embodiment of the present disclosure.

[0058] The control system 101 can predict the arrival information of potential future passengers based on the trajectories of the passengers. The predicted arrival information is used to determine the dispatching (dispatching 107) of the elevator cars (such as cars 125 and 127) of the elevator group 109 to serve passengers going to the desired floors in the set of floors 129a, 129b, 129c, 129d, 129e, and 129f (also referred to as the set of floors 129a to 129f) in a building (not shown in Figure 1A ).

[0059] To this end, the control system 101 can receive from one or more passengers one or more current passenger requests 119 for the service provided by the elevator group 109, and receive the partial trajectories 103 of the movement of each of one or more persons (i.e., one or more passengers) moving on the floors (e.g., the floors in the set of floors 129a to 129f) including the area served by the elevator group 109. The partial trajectory 113 can include data associated with the movement of a potential future passenger 115 on any of the floors in the set of floors 129a to 129f. The control system 101 can use sensors 117 to capture the movement of the potential future passenger 115 in order to determine the partial trajectory 113 of the potential future passenger 115, where each partial trajectory may be incomplete and its final destination remains to be predicted.

[0060] In addition, the control system 101 can utilize the partial trajectory 113 as an input to a neural network 103, which can predict the arrival information of potential future passengers as an output in the form of a polynomial 111, where the arrival information can include all possible destinations and corresponding time instances.

[0061] The polynomial 111 is a two-dimensional entity that has one dimension corresponding to the destinations (e.g., dest1, dest2, and dest3) and another dimension corresponding to the arrival times (e.g., t1 and t2). Specifically, the two-dimensional polynomial distribution can be a matrix that has the destination names for the columns or rows and the time instances for the rows or columns, where the values of the matrix define the probability of arriving at the corresponding destination at the corresponding time interval. The destinations in the first dimension are obtained by quantifying each floor in the set of floors 129a to 129f into checkerboard layout elements. Since the destinations are obtained by quantifying the floors (e.g., 129a), the destinations can correspond to different locations on that floor. For example, destination 1 can be the cafeteria located on that floor, destination 2 can be the service request area located on that floor, and destination 3 can be the kiosk located on that floor. At the same time, the time dimension is also quantified into a set of time intervals (e.g., t1 and t2) that define the time period of interest for the prediction.

[0062] Thus, based on the partial trajectory 113, the neural network 103 performs an extended destination prediction for a person with such movement at different destinations and at each time interval. Further, the sum of all the values in the matrix equals "1", and "1" reflects the probabilistic nature of the multinomial distribution. For example, the sum of the values in the multinomial 111 is 0.03 + 0.01 + 0.6 + 0.02 + 0.04 + 0.3 = 1. This enables the neural network 103 to determine the probability that a passenger will reach a specific destination (say, an elevator request area) relative to other destinations. Thus, accurate destination prediction is performed.

[0063] In some embodiments, the neural network 103 may be configured to use an interested time period and a quantization value to determine the arrival of a passenger at different time intervals. For example, if the interested time period for prediction is "5 seconds" and the quantization equals "1 second". Then, based on the multinomial 111, the neural network 103 can predict the arrival of future passengers associated with 5 time intervals (e.g., t1, t2, t3, t4, and t5). Thus, the probability of reaching a specific destination equals the sum of the values of the destination over all time intervals. Further, the control system 101 uses such information to determine the schedule 107 to control the operation of the elevator group 109.

[0064] More specifically, after receiving the partial trajectory 113, the neural network 103 of the control system 101 can generate a multinomial 111 for the extended destination prediction of one or more persons. The multinomial 111 may have at least two dimensions, and the at least two dimensions include a first dimension of the destination of the person and a second dimension of the time interval at which each person arrives at the destination in the first dimension. Each floor in the set of floors 129a to 129f can be quantized to form a checkerboard arrangement element corresponding to the destination in the first dimension of each floor. The multinomial 111 may be determined for an entire floor (e.g., floor) 129a, and each checkerboard arrangement element may be associated with a different area of the floor. At least one of the different areas corresponds to an area served by the elevator group 109, and the probability of the corresponding value of the multinomial 111 may be determined for each checkerboard arrangement element associated with the area served by the elevator group 109. The list of destinations in the multinomial 111 is exhaustive, and the sum of the values in the multinomial equals "1", which ensures that a person wandering on the floor (e.g., 129a) can always be located in at least one of the destinations in the list of destinations.

[0065] In addition, some embodiments are based on the recognition that the AWT for one or more passengers for service can be optimized based on information associated with one or more passengers who have currently requested elevator service, as well as the time instances provided in polynomial 111, and the probability of the corresponding values determined for the checkerboard arrangement elements associated with the area served by elevator group 109, for one or more passengers (also referred to as one or more unserved passengers) who request elevator service. Such a request from a current passenger is also referred to as a current passenger request 119. In addition, one or more unserved passengers not carried by an elevator car (e.g., carriages 125 and 127) are referred to as existing unserved passengers 121. Therefore, the control system 101 needs to determine the scheduling for all passengers including the unserved passengers so as to shorten the AWT for all passengers.

[0066] To this end, the control system 101 uses a scheduler 105, which determines a schedule 107 for the service provided by elevator group 109 such that one or more performance metrics (such as the AWT for all passengers) are optimized. The scheduler 105 obtains information on the current passenger requests 119 and information on the existing unserved passengers 121, together with the predicted arrival information of potential future passengers (i.e., polynomial 111), to determine the schedule 107 for the controller 123 coupled to elevator group 109.

[0067] Elevator group 109 includes a plurality of carriages (such as carriages 125 and 127) installed in a building having a plurality of floors (floor sets 129a to 129f). After receiving the schedule 107, the controller 123 issues a command to carriages 125 and / or 127 to move through floor sets 129a to 129f to pick up the one or more passengers accordingly.

[0068] Figure 1B Block diagram illustrating a control system 101 according to an embodiment of the present disclosure.

[0069] The control system 101 may include a first input interface 131 configured to receive a current request for the service provided by elevator group 109 from one or more passengers. The first input interface 131 may include a dashboard, and the dashboard includes means (such as buttons, touchscreens, etc.) that can be used by one or more passengers to request service (such as by pressing an up button or a down button). In another embodiment, the first input interface 131 may include a destination control (DC) panel so that the destination floor is determined before entering the elevator.

[0070] In addition, the control system 101 may include a second input interface 133, which may be configured to receive a partial trajectory 113 of the movement of one or more passengers moving on a floor (e.g., floor 29a) that includes information about the area served by the elevator group 109. The second input interface 133 may communicate directly with a sensor 117 (e.g., a motion sensor) to obtain the partial trajectory 113 of the movement of one or more passengers on the floor.

[0071] Each of the first interface 131 and the second interface 133 is connected to other components of the control system 101 (such as a processor 137, a memory 139, etc.) via a bus 135. In addition, the control system 101 includes a processor 137, which is configured to execute instructions stored in the memory 139. The processor 137 may be a single-core processor, a multi-core processor, a computing cluster, or any number of other configurations. The memory 139 may be a random access memory (RAM), a read-only memory (ROM), a flash memory, or any other suitable memory system. The processor 137 may be connected to other components of the control system 101 via the bus 135.

[0072] The instructions may implement a method for controlling the movement of the elevator group 109. To this end, the memory 139 may include an instruction set of a neural network 103 for implementing a destination predictor 141. The destination predictor 141 may be implemented to predict the destination of one or more persons. The neural network 103 may obtain the partial trajectory 113 from the second input interface 133. After obtaining the partial trajectory 113, the processor 137 may execute the neural network 103 trained for the extended destination prediction of the person by generating a polynomial 111 for the extended destination prediction of the person. The processor 137 may be further configured to optimize the schedule 107 to serve the current requests from one or more passengers and the future requests of the one or more passengers at the time instances provided in the polynomial 111, with the probabilities of the corresponding values determined for the checkerboard layout elements associated with the area served by the elevator group 109 in the polynomial 111.

[0073] The instructions of the memory 139 further include a scheduler 105, which is configured to assign a car (e.g., car 125 or car 127) to each passenger so as to minimize a performance metric (e.g., the average waiting time (AWT) for all passengers). To this end, the processor 137 is further configured to execute the scheduler 105 to control the elevator group 109 according to the schedule 107.

[0074] Sensor

[0075] The sensor 117 can be installed in the areas of the set of floors 129a, 129b, 129c, 129d, 129e, and 129f where future passengers will arrive. In some example embodiments, the sensor 117 can be a motion detector, such as a camera that detects human motion or presence (e.g., surveillance cameras often placed in the aisles and lobbies on each of the floors 129a to 129f of a building) or a proximity sensor (such as a radio antenna).

[0076] In one or more example embodiments, radio antennas serving as the sensor 117 are placed every few meters in the set of floors 129a to 129f of the building. Thus, "Location N" or "Checkerboard Arrangement Element N" indicates that a passenger is detected by antenna number N. If multiple antennas detect a passenger, the exact number of the primary sensing antenna can be determined by the fixed tie-breaking order of the antennas.

[0077] In addition, the sensor 117 can be configured to detect one or more persons at multiple locations in the building, not necessarily only at the elevator doors or aisles leading to the elevator group 109. In such a case, when a person is detected at a location l (e.g., in a lobby 50 meters away from the elevator platform), the probability p that this person will request elevator service can be determined by correlating the sensed data with the actual service requests. i . In addition, the control system 101 utilizes the correlation to determine the schedule 107 in an accurate manner.

[0078] In addition, some embodiments are based on the recognition that the trajectories of passengers on the set of floors 129a to 129f (e.g., partial trajectory 113) can be represented in a way that enables the control system 101 to analyze these trajectories to determine the schedule 107. Such trajectories can be obtained from the sensor 117. The partial trajectory 113 includes trajectory data that contains only position information. In some implementations, each trajectory is represented by a sequence of checkerboard arrangement indices, where the checkerboard arrangement indices are obtained by discretizing the range of position coordinates. In addition, the range of position coordinates represents the longitude and latitude associated with the checkerboard arrangement elements.

[0079] The control system 101 can store these trajectories of one or more passengers wandering on each floor of the building in a database. In some cases, the records of the trajectories obtained from the sensor 117 regarding both position and timestamp are continuous variables. Therefore, it may not be feasible to directly use such continuous variables for a sequence-to-sequence predictor. Thus, the position data and timing data can be preprocessed by discretizing them into checkerboard arrangement indices.

[0080] In an example embodiment, given a set of position coordinates

[0081] {(xl , y l ); l = 1, 2, ..., L},

[0082] where x and y represent longitude and latitude respectively. Let x min = min l∈{1,2，...，L} x l , x max = max l∈{1，2，...，L} x l , y min = min l∈{1，2，...，L} y l , y max = max l∈{1，2，...，L} y l . Then, the chessboard arrangement index p X × N Y of any given tuple of coordinates (x l , y l ) in the rectangle can be determined as l where where the quantization interval The quantization interval The quantization interval The positive integer N X and N Y can be made reasonably large as long as computing resources permit. Additionally, given a set of trajectory data, N X and N Y can be determined by using cross - validation. This embodiment uses this data in the neural network 103 of the control system 101 to generate a polynomial 111 for controlling the elevator group 109.

[0083] As used herein, a chessboard arrangement is an arrangement of shapes that fit closely together in a pattern without gaps or overlaps. Examples of such shapes include squares, rectangles, polygons, etc. When the chessboard arrangement uses squares, such a chessboard arrangement can also be referred to as a regular grid, or simply as a grid. When the chessboard arrangement uses rectangles of different sizes, such a chessboard arrangement can also be referred to as an irregular grid. In different embodiments, the chessboard arrangement elements can have the same or different shapes to better fit the layout of the area served by the elevator system.

[0084] Figure 1C Illustrate the quantization of floors in the regular chessboard arrangement elements 143 of the regular grid for polynomial distribution according to an exemplary embodiment of the present disclosure.

[0085] Some embodiments are based on the recognition that the rectangular checkerboard arrangement of elements 143 facilitates the calculation of the trajectories of one or more passengers on a floor. To this end, each floor in the set of floors 129a to 129f is quantified based on a constant value such that the area of each floor is equally divided. Based on such quantification, the partial trajectories 113 corresponding to each checkerboard arrangement element of the quantified floors can be analyzed by the control system 101 to generate polynomials 111. Based on the polynomials 111, a schedule 107 for controlling the elevator group 109 is determined. One or more checkerboard arrangement elements 145c of the rectangular checkerboard arrangement elements 143 correspond to service areas associated with requests for service provided to the elevator group. Examples of service areas include the area of the elevator landing and / or the area in which the first interface 131 is arranged.

[0086] Figure 1D Illustrates the quantification of floors in the checkerboard arrangement elements 145 of an irregular grid for polynomial distribution according to an example embodiment of the present disclosure. One or more checkerboard arrangement elements 145d of the irregular checkerboard arrangement elements 145 correspond to service areas associated with requests for service provided to the elevator group.

[0087] Some embodiments are based on the recognition that the irregular checkerboard arrangement elements 145 can accurately cover the entire area of a floor and facilitate the accurate calculation of the trajectories of one or more passengers on that floor. However, the calculation can be very complex. In an embodiment, the set of floors 129a to 129f is quantified with irregular checkerboard arrangement elements 145 for determining the trajectories of one or more passengers on each floor in the set of floors 129a to 129f.

[0088] In some embodiments, only the information associated with a number of checkerboard arrangement elements (e.g., regular checkerboard arrangement elements 143 or irregular checkerboard arrangement elements 145) of the quantified floors that correspond to the area served by the elevator group 109 in the polynomials 111 can be used to determine the schedule 107 for controlling the elevator group 109. In such a case, a number of checkerboard arrangement elements of the quantified floors that are not associated with the area served by the elevator group 109 can be discarded when determining the schedule 107 for controlling the elevator group 109.

[0089] In some other embodiments, the information associated with all the checkerboard arrangement elements (e.g., regular checkerboard arrangement elements 143 or irregular checkerboard arrangement elements 145) of the quantified floors can be used to determine the schedule 107 for controlling the elevator group 109.

[0090] Figure 1E Illustrates a graph of the quantification of the elapsed time of a trajectory for destination prediction according to an embodiment of the present disclosure.

[0091] Figure 1E The exemplary trajectory shown has a start time t0 147. The current timestamp is t 149, the quantization interval is Δt, and the prediction duration is T 151. The resulting checkerboard arrangement index of the elapsed time of the trajectory is indicated by 153. Let t m represent the checkerboard arrangement index. In an example embodiment, for a given upper bound Q on the elapsed time, the control system 101 may determine the checkerboard arrangement index t of the elapsed time at the current timestamp t 149 for a given tuple of position coordinates in a given trajectory m . Further, for the checkerboard arrangement index t m , the start timestamp can be recorded as t0 147, where the quantization interval where N where N T is an integer representing the maximum checkerboard arrangement index of the elapsed time. N T can be made reasonably large, as long as computational resources permit. Further, given a set of trajectory data, N T can be determined using cross-validation.

[0092] Some embodiments consider both position information and timing information to represent a trajectory as expressed in Equation (1):

[0093] S = ((p1, t1), (p2, t2),...,(p n , t n )) (1)

[0094] where n is the length of the trajectory, and the last tuple (p n , t n ) may correspond to the destination, provided that the trajectory is complete. The processor 137 will retain only the checkerboard arrangement index itself (without parentheses and commas) from the representation of the trajectory S in Equation (1).

[0095] In some embodiments, the checkerboard arrangement index may be represented using longitude and latitude. For example, for a tuple of coordinates (x l , y l ), the checkerboard arrangement index p l1 (correspondingly, p l2 ) may be used for longitude (correspondingly, latitude), where the checkerboard arrangement index p l1 is determined as (correspondingly, ).

[0096] Thus, the same trajectory S can be represented as expressed in Equation (2).

[0097] S = ((p11 , p 12 , (p, t1) 21 , p 22 , (p, t2), …, (p n1 , p n2 , t n )) (2)

[0098] The processor 137 may retain only the checkerboard arrangement index itself (without parentheses and commas) from the above trajectory representation of equation (2).

[0099] In some embodiments, the tuples of the checkerboard arrangement indices that make up the trajectory are mapped to new single symbols. Thus, each trajectory in the partial trajectory 113 can be represented by a sequence of ordinal numbers, where each symbol is mapped from a tuple of checkerboard arrangement indices indicating position information and timing information. For example, the above trajectory S is represented as S = q1q2...q n . Since this representation shortens the original trajectory (and thus, when there is the same number of trajectories, information is lost), in order to achieve comparable generalization performance, a larger set of trajectory data may be required to train the prediction model composed of the destination predictor 141.

[0100] Set of continuations

[0101] In various embodiments, the scheduler 105 is configured to optimize the scheduling of an elevator group that serves current requests from passengers and future requests of people, to obtain elevator service at a time interval provided in a polynomial, with a probability corresponding to a value determined for a checkerboard arrangement element associated with the service area in the polynomial. In this way, the scheduler 105 is a probabilistic scheduler that takes into account the probability of future passenger requests. Different embodiments use different implementations of the probabilistic scheduler. However, considering future requests in a probabilistic manner is challenging. For this reason, some embodiments generate samples from the polynomial 111 to represent the probability of future passenger requests in a deterministic manner, thereby simplifying the calculation.

[0102] For example, one embodiment considers the probability of future requests generated by the destination predictor 141 by generating a set of continuations (or continuations 170) of the current arrival stream, where the current arrival stream corresponds to the current passenger request 119. In this embodiment, the scheduler 105 determines the schedule 107 based on the polynomial 111, which includes time information and destination information associated with requests from one or more passengers. To this end, the scheduler 105 performs a simulation (e.g., Monte Carlo simulation) and forms a set of continuations that includes each current request and future request associated with the service of an elevator.

[0103] In some embodiments, a continuation set is formed by filtering the destination-associated values of polynomial 111 by considering only the areas of floors 129a through 129f served by elevator group 109. For example, in polynomial 111, dest2 among destinations dest1, dest2, and dest3 is an area served only by elevator group 109. In such a case, polynomial 111 is filtered such that only the information regarding dest2 in polynomial 111 is considered to form the continuation set, and the information regarding other destinations (i.e., dest1 and dest3) not related to the areas served by elevator group 109 is discarded.

[0104] For example, let H(t) be the set of passengers who have arrived by time t but are still waiting to be served, and W(H(t)) be the cumulative waiting time of the passengers in H(t), where W c (·|·) represents the waiting time of one or more passengers considering that another set of zero or more passengers is also assigned to the same car c. Further, let H c (t) be the set of passengers who have been assigned to car c by time t but are still waiting to be served. Further, for passenger h, the margin increase in the waiting time when assigned to car c is given by Equation (3):

[0105] ΔW c (h)≡W c (H c (t)∪h|H c (t)∪h)-W c (H c (t)|H c (t)). (3)

[0106] Equation (3) can be further expanded as expressed in Equation (4):

[0107]

[0108] where the first term is the time required to serve passenger h with car c, and the remaining terms in the above summation account for the increase in waiting time caused by passenger h to the passengers already in set H c (t) when passenger h is also assigned to c.

[0109] The continuation set can be mathematically defined as represented in Equation (5):

[0110]

[0111] In some embodiments, the output of destination predictor 141 can be used to generate M continuation sets These continuation sets may not necessarily be of the same length.

[0112] For each continuation set The optimal cumulative waiting time (CWT) is expressed by equation (6):

[0113]

[0114] in represents the continuation set (of passengers) assigned to car c The average waiting time (AWT) for this allocation can be determined as:

[0115]

[0116] In some embodiments, the duration of the n continuation sets can be adjusted according to the available computing resources. i <t的乘客h i The same elevator cars are assigned to each continuation set. In addition, any practical method for minimizing the AWT (eg, clear system method, immediate assignment, and reallocation mode) may be used.

[0117] Immediate allocation

[0118] In this mode, the current passenger h is temporarily assigned 250 to car c with a margin waiting time (MWT) 251

[0119]

[0120] The range of g includes all passengers temporarily assigned to car c.

[0121] Note that future passengers are omitted in the first term. However, when future passengers The allocations have an impact on their waiting times when the slack waiting times are determined as follows:

[0122]

[0123] in Indicates that at time t i+k The set of future passengers who have arrived previously and have been assigned to car c.

[0124] Then, the current passenger h is temporarily assigned to the continuation set and can consider the known passenger h and the unknown future passengers The mutual influence between them.

[0125] The immediate assignment mode has a relatively low complexity compared to the exhaustive search, being linear in the number of future arrivals, but it is not necessarily optimal for all passengers in the continuation set as it only considers the passengers who have arrived before the current passenger is assigned at time t. before time t i+k have already arrived.

[0126] Immediate assignment for current passengers and re - assignment for future passengers

[0127] The immediate assignment mode requires that the assignment for the current passenger is made immediately and never reconsidered. However, there is no such assignment restriction for future passengers. This allows for at least in principle a reconsideration of the assignment. However, this may lead to a significant increase in calculations and may also not correspond to the way the scheduling is being performed.

[0128] For example, one of the n continuation sets may actually occur in the future, even if this is unlikely but not impossible. In this case, the assignment for the request is made in the immediate mode and re - assignment is not allowed. So, if a good partition has been determined in the re - assignment mode, it may be missed in the immediate assignment mode, which is why re - assignment is likely not to be used when scheduling future passengers.

[0129] Re - assignment mode

[0130] When an efficient procedure is available for determining the optimal assignment of passengers for the entire continuation set, it can also be used efficiently for the Monte Carlo evaluation on the extended continuation set, with the computational time increasing accordingly. associated with all increases in computational time.

[0131] Regardless of which mode is used, the Monte Carlo scheduling method operates in a rolling - horizon manner. After passenger h has been assigned (temporarily or permanently) using n sets at time t i at time t i when the next passenger h arrives at time t i+1 at time t i+1 a new continuation set is generated from the information vector I(t i+1 )

[0132] For the format of the information vector I(t), several options are possible, depending on the type of sensed information. A general format that can be used to generate the Monte Carlo set is independent of sensor 117. This format is a matrix specifying the stochastic process of the arrival process for each pair of origin and destination floors.

[0133] Monte Carlo simulation is used to model the probabilities of different consequences in a process that may not be easily predictable due to the intervention of random variables. This is a technique for understanding the impact of risk and uncertainty in prediction and forecasting models. Some embodiments use Monte Carlo simulation to represent the probabilistic nature of future service requests in a deterministic manner using ensemble sets, where each ensemble set is deterministic but together represents the probabilities of future requests at different time instances.

[0134] Monte Carlo simulation can be a broad class of computational algorithms that rely on repeated random samples to obtain numerical results. The basic idea is to use randomness to solve problems that may be deterministic in principle. Monte Carlo simulation can be used to generate draws from probability distributions. Monte Carlo simulation provides decision-makers with a range of possible consequences and the probabilities that they may occur for any choice of action. Monte Carlo simulation can show extreme possibilities—the consequences of taking a big risk and the consequences of the most conservative decision—along with all the possible consequences for decisions in the middle ground. In addition, Monte Carlo simulation can perform risk analysis by building models of possible outcomes by replacing any factors that may have inherent uncertainty with ranges of values—a probability distribution. Monte Carlo simulation then calculates the results over and over again, each time using a different set of random values from the probability function. Depending on the amount of uncertainty and the ranges specified for them, Monte Carlo simulation may involve thousands or tens of thousands of recalculations before it is complete. Thus, Monte Carlo simulation can generate a distribution of possible consequence values.

[0135] In some embodiments, the control system 101 can perform a Monte Carlo simulation of the values determined for the checkerboard layout elements associated with the area served by the elevator group 109 for the polynomial 111 to generate multiple combinations of request times. The occurrence frequency of a particular time instance in the multiple combinations of request times is the probability of a future request for elevator service in a particular time interval, the probability having a value at the intersection of the particular time interval and the checkerboard layout elements associated with the area served by the elevator group 109 in the polynomial 111. In addition, the control system 101 can combine each of the multiple combinations of request times with the current request to generate multiple ensemble sets. In addition, the control system 101 can determine a schedule 107 for the elevator group 109 to optimize a performance metric for at least some of the passengers in all combinations of the ensemble sets. In an example embodiment, the control system 101 can use Monte Carlo simulation to implement the ensemble sets.

[0136] Figure 1FSchematic illustration of a method for generating continuations based on a passenger's polynomial 111 using Monte Carlo simulation according to some embodiments. Consider two possible future passengers, each with a different polynomial (note that at the current time, they may be on the same floor or different floors). For each possible future passenger, five samples are drawn from the polynomial, and if in the current sample the passenger is going to the elevator, he / she is included in the future arrival list 160 for the corresponding continuation. Note that in order to reduce the space (memory) cost, the drawn samples are filtered 194 to retain only the samples where the passenger is going to the elevator, and all other samples are discarded. Sample 2 of passenger 1 corresponds to the circled value 190 in the polynomial of passenger 1, and sample 2 of passenger 2 corresponds to the circled value 192 in the polynomial of passenger 2. Since both samples 191 and 193 have the elevator as the destination, they are included in the future arrival list 160 for continuation set 2. As Figure 1G shown, this future arrival list 160 will be combined with the current arrival stream to generate a complete set of continuation sets 170.

[0137] For example, the entry "(t2,p1)" in continuation set c2 indicates that passenger 1 will arrive at the elevator at time t2 and issue an elevator service request. If this happens to be the down-peak period, "p1" can be further specified as the arrival floor and the destination floor, where the arrival floor is the floor where passenger 1 is currently located, and the destination floor is 1 (i.e., the lobby). A more detailed example of forming the continuation set is given below. In a more general case, "p1" should be interpreted accordingly as the arrival floor and the destination floor of passenger 1, where the arrival floor is the floor where passenger 1 is currently located, and the destination floor can be determined by time period, etc. The most general method for setting the destination floor would be to randomly draw it from a uniform distribution over the candidate destination floors; when using this method, additional continuation sets should be formed to account for the randomness. In another embodiment, if the identity of the passenger is known and historical data on the floors the passenger has traveled to in the past is available, then the probability distribution of the target floor can be learned from this data, and when creating continuations, samples are drawn from it.

[0138] Figure 1G Exemplary representation of a continuation set 170 based on destination prediction according to an exemplary embodiment of the present disclosure.

[0139] In Figure 1G there are 25 continuation sets generated based on Monte Carlo simulation according to polynomial 111. Each continuation set 170 includes one or more current requests 150 and future requests 160 for the service of the elevators in elevator group 109 (as Figure 1FA combination generated as shown. The current request also includes the current passenger request 119, along with all those passenger requests that have been registered with the GES but have not yet been served. Note that for all 25 continuations, the current arrival stream 150 is exactly the same, while the future arrivals 160 are different for different sets of continuations depending on which samples are drawn from the multinomial distribution of future passengers.

[0140] Each in the current request 150 is associated with the current passenger request 119, and each in the future requests is associated with each time instance and destination indicated by the multinomial 111, where the destination corresponds to each area of the set of floors 129a to 129f served by the elevator group 109. Areas of the set of floors 129a to 129f that are not served by the elevator group 109 are discarded and not considered. Thus, a more precise set of continuations of future requests with high probability is obtained. The set of continuations is exhaustive, including all possible consequences of future requests. Such an exhaustive set of continuations enables the effective optimization of the AWT for all passengers. Each set of continuations in the set of continuations includes time information associated with the current request and future requests.

[0141] For example, consider a building with eight floors and each floor (indicated by 1 to 8) has an elevator group. The starting time is t0 = 5:00:00 pm. The existing arrival stream is {(5:00:00 pm, 8, 1), (5:00:02 pm, 7, 1)}.

[0142] In the case of a quantization interval Δt = 0.01 s for elapsed time and a prediction duration T = 15 s, the future arrival information extracted from the destination predictor 141 may indicate that at the 7th floor, the probability that passenger 1 will arrive at the elevator within 9 s (checkerboard layout index for elapsed time is 900) is 80%, and the probability that the passenger will arrive at the elevator within 9.5 s (checkerboard layout index for elapsed time is 950) is 20%; at the 8th floor, the probability that passenger 2 will arrive at the elevator within 10 s (checkerboard layout index for elapsed time is 1000) is 20%, the probability that the passenger will arrive at the elevator within 10.2 s (checkerboard layout index for elapsed time is 1020) is 60%, and the probability that the passenger will arrive at the elevator within 10.6 s (checkerboard layout index for elapsed time is 1060) is 20%.

[0143] Furthermore, each set of continuations is deterministic in nature because the set of continuations represents the probabilities of future requests at different time instances. Thus, the control system 101 uses the set of continuations, via the scheduler 105, to determine the schedule 107 to control the elevator group 109 considering both the current service requests and future service requests.

[0144] Schedule

[0145] Figure 1H Schematic illustration of dispatching cars of elevator group 109 based on a continuation set according to an example embodiment of the present disclosure.

[0146] In some embodiments, the destination predictor 141 may work with the complete trajectories outside elevator group 109. In other words, the destination predictor 141 is used to predict the arrival time of reaching a destination using a particular elevator car of interest. In the most general case, once a passenger enters an elevator car (car 125 or car 127), his / her destination floor can be any one of the remaining floors in the requested direction (e.g., up or down) (among floor sets 129a to 129f). Additionally, the expected value of the waiting time can be obtained with respect to the uncertainty of the destination floors of the passengers who have not yet been picked up by the elevator cars. In some cases, during the up-peak or down-peak time periods, the destination floors of all passengers (from floor sets 129a to 129f) are uniquely specified. In response to receiving a request, the dispatcher 105 assigns an elevator car to each passenger in a manner that achieves a performance metric (e.g., minimizes the average waiting time (AWT) for all passengers).

[0147] Some embodiments are based on the recognition that minimizing the AWT for all passengers who request an elevator at the current time and during a future time interval can be conceived as a group elevator scheduling (GES) problem.

[0148] To facilitate the throughput of the GES system, consider H = {h1, h2,..., h J} representing the set of passengers arriving during the time interval, where passenger h i can be indicated by a tuple (τ i , o i , d i ), where τ i is the arrival time, o i is the arrival floor, and d i is the destination floor.

[0149] Some embodiments are based on the recognition that partitioning the set H into C subsets results in an exponentially growing number (i.e., C J number) of possible partitions, and a suitable combinatorial optimization method is needed to handle the scheduling problem. Additionally, the GES system has only limited access to arrival information at the current time instance t satisfying τ1 < t < τ J , and the GES system has information only about all service requests that have occurred up to the current time t and the states of the C cars in elevator group 109.

[0150] With this understanding, the purpose of the GES system is to optimize the scheduling 107 of the elevator group 109 by minimizing the AWT for all passengers requesting an elevator for a current time and a future time interval. To this end, the control system 101 can utilize information from a continuation set, including an exhaustive list of current and future requests, the list indicating information for each destination and corresponding time instance.

[0151] For a set H of passengers {h1, h2,..., h N} arriving during a time interval, a passenger h i can be indicated by a tuple (t i , o i , d i ), where t i is the arrival time (i.e., current time and future time), o i is the arrival floor, and d i is the destination floor. The assignment of N passengers to C cars in the group partitions the set H into C subsets H C , such that H = H1 ∪ H2 ∪ … ∪ H C , and H i ∩ H j is the empty set when i ≠ j

[0152] When all passengers in a set A are assigned to a car c, the waiting time for a passenger h in the set A assigned to the car c is W c (h|A). Similarly, when all passengers in a set H are assigned to a car c, the cumulative waiting time for all these passengers is W c (H|A). The sets H and A are not necessarily the same.

[0153] In general, the waiting time W c (H|A) depends on the predefined order in which the car c serves the passengers in the set H ∪ A. Most elevator systems use a fully collective policy, in which the car serves all requests in sequence in one direction and then reverses and answers all calls in the opposite direction. When the car is empty and stopped, the likely upward and downward directions are compared, and the direction that results in a shorter AWT is chosen. Other possible service orders that optimize the AWT are also possible. However, regardless of the method chosen, for a given combination of the sets H and A and the position of the car c, the resulting waiting time W c (H|A) can be fully determined.

[0154] For a given complete assignment, the total waiting time W(H) for all passengers in the set H can be expressed as follows:

[0155]

[0156] And the AWT of the passengers in set H is W(H) / N. Partition set H into C subsets, there are C N possible partitions. Using infinite computing resources and / or suitable combinatorial optimization methods, determine the optimal allocation.

[0157] Therefore, after obtaining the continuation set, the control system 101 optimizes the scheduling 107 of the elevator group 109 for serving the current requests from passengers and the future requests of people, so as to obtain elevator service at the time instances provided in the polynomial 111, in association with the area served by the elevator group 109.

[0158] Based on the optimization of the scheduling 107, control the elevator group 109. For example, Figure 1H It is shown that the arrival time ta is reduced and this time is divided into a time interval t1 for the passengers whose requests have been served, a time interval t2 for the passengers with unassigned allocations that have not been served, the current time tc, and a future time interval t3. The upward and downward signs of the solid line indicate the current requests, and the upward and downward signs of the dashed line indicate the future requests. In addition, the letters A and B represent the carriages (carriage 1 and carriage 2). During the time interval t1, the successive selection of the carriages is arranged as a decision tree. During the future time interval t3, the requests are fulfilled in the immediate allocation mode.

[0159] The AWT over all the continuation sets is calculated for each temporary allocation of the current passenger requests (in this case, either a request for carriage 1 or a request for carriage 2), and then the carriage with the shortest AWT is selected to allocate the current passenger requests at the current time tc. In other words, the scheduler 105 compares how long all the carriages available at the current time point will be waited for by the set of existing passengers and possible future passengers. Multiple continuation sets ensure that this calculation takes into account not only one possible future realization of the passenger arrival stream, but also more possible future realizations of the passenger arrival stream. Therefore, the carriages (carriage 1 and carriage 2) of the elevator group 109 are scheduled according to the optimization of the scheduling 107.

[0160] Figure 2A The block diagram of the neural network 103 according to an example embodiment of the present disclosure is illustrated. The destination predictor 141 can be implemented by the neural network 103 to predict the future requests of one or more passengers. In some embodiments, the neural network 103 can be implemented according to a model-based method, where the model-based method has difficulties in processing long sequences of past observed data (e.g., partial trajectories 113), because when encountering long sequences, the early information tends to dissipate.

[0161] Alternatively, in some other embodiments, neural network 103 may be based on deep learning methods. Such embodiments are based on the recognition that deep learning-based methods are capable of overcoming natural language problems and also have the ability to adapt to non-linguistic applications such as destination / trajectory prediction. Examples of deep learning-based neural networks include recurrent neural networks (RNNs) that work with variable-length time series data, and the (corresponding to the time series data) trajectory can be predicted via the RNN to predict future requests. Examples of RNNs may include, but are not limited to, long short-term memory (LSTM)-based neural networks, bidirectional long short-term memory (BiLSTM)-based neural networks, and the like.

[0162] In Figure 2A , the neural network 103 is an LSTM-based neural network, which includes an input gate, a forget gate, and an output gate. Each of these gates corresponds to a "standard" neuron in a feedforward (or multi-layer) neural network, that is, they calculate the activation of a weighted sum (using an activation function). Additionally, it, ot, and ft represent the activations of the input gate, output gate, and forget gate at time step t, respectively. The three exit arrows from the storage primitive c to the three gates i, o, and f represent peephole connections. These peephole connections represent the contribution of the activation of the storage primitive c at time step t - 1, that is, c t-1 's contribution. In other words, the gates i, o, and f calculate their activations at time step t (i.e., i t , o t , and f t ), respectively), and also consider the activation of the storage primitive c at time step t - 1, that is, c t-1 . The single left-to-right arrow exiting the storage primitive c is not a peephole connection, but represents c t , and the "×" symbol represents element-wise multiplication between the inputs. Additionally, the LSTM-based neural network includes applying a differentiable function (such as the sigmoid (S-shaped) function) to the weighted sum.

[0163] In operation, the forget gate receives a partial trajectory 113 detected by the sensor 117. The forget gate decides what information to discard from the input data (partial trajectory 113), which can also be referred to as the primitive state. For example, the previous output and the new input can be passed through a sigmoid hidden layer. Each neuron in the hidden layer can calculate the weighted sum of the entries of the input (e.g., weighted with relevant parameters) and add a bias value to it. The resulting scalar can then be passed through a specific non-linear function named sigmoid, σ. The output of such a process can then be set by the neuron to the next step. The output of the forget gate is provided to the input gate.

[0164] In addition, an input gate determines what new information is to be stored in the primitive state. To this end, the previous output and the new input can be passed through a sigmoid hidden layer, and in some embodiments, the previous output and the new input can be separately passed through the hidden layer. The two different layers differ only in the type of non-linear function applied to the output of the individual computations applied to each neuron (e.g., one is sigmoid σ and the other is hyperbolic tangent). The state of the primitive can be updated based on the output of the input gate.

[0165] In addition, an output gate determines the output of the LSTM-based neural network based on the updated primitive state, the previous output, and the current input. Both the previous output and the new input can be passed through a sigmoid hidden layer. The result can be multiplied element-wise with the updated primitive state vector Ct, and the output is the resulting vector ht corresponding to polynomial 111. Since the LSTM-based neural network includes discarding information via a forget gate, polynomial 111 may not be accurately obtained at the output.

[0166] Figure 2B Block diagram illustrating a neural network 103 according to another example embodiment of the present disclosure.

[0167] Some embodiments are based on the recognition that the attention mechanism of the transformer architecture can model the dependencies of elements in order without considering the distance between the elements, and such advantages of the transformer architecture are not limited to speech-related applications, but also apply to extended destination prediction tasks.

[0168] Therefore, in such embodiments, the destination predictor 141 can be an attention-based destination predictor having an encoder-decoder structure. The attention-based destination predictor can be implemented using a neural network based on the transformer architecture. The destination predictor 141 can include a stack of N encoder 201 blocks and N decoder 203 blocks. In addition, in a manner using positional encoding and attention mechanism that can achieve parallelization in the transformer architecture, each stack of the N encoder 201 blocks and the N decoder 203 blocks includes a stack of self-attention and fully-connected feed-forward layer components.

[0169] As Figure 2B shown, initially, the input trajectory 205 passes through an embedding and positional encoding block 207, in which the input tokens can be transformed by an embedding layer into d modelA d-dimensional vector, where the tokens can be discrete elements. Additionally, order information about the elements of the input can be obtained using positional encoding and then combined with the embedding vectors to obtain a new vector via a summation operation. Additionally, the new vector can be passed into the encoder 201 via the multi-head attention block 209. The attention in the encoder 201 can be used as a means to reference other tokens in the input trace 205 when attempting to encode a particular token.

[0170] The multi-head attention block 209 can consist of an array of scaled dot-product attention components, where the attention function can be calculated based on the data matrix X. Three different matrices learned during training can be used to project the data matrix, and projection matrices Q, K, and V representing queries, keys, and values respectively are obtained.

[0171] In some embodiments, the attention can then be calculated as:

[0172]

[0173] where d k is the number of columns of the K matrix.

[0174] In some embodiments, the multi-head attention mechanism involves using different learned projections to calculate multiple attention functions in order to obtain improved representation performance.

[0175] Additionally, the addition and normalization block 211 in the encoder 201 can include a residual connection along with layer normalization (RLN), which can connect the input of the multi-head attention block 209 to the output of the multi-head attention block 209, add them, and then normalize. The resulting elements can further pass through the feed-forward sub-layer 213 and are then followed, on the path leaving the encoder 201, by another addition and normalization block 215 including an RLN connection. The addition and normalization block 215 can connect the input of the feed-forward layer 213 to the output of the feed-forward layer 213, add them, and then normalize.

[0176] Additionally, the target output trace 217 can, before entering the masked multi-head attention block 221, pass through an embedding and positional encoding block 219, which can be similar to the embedding and positional encoding block 207. The masked multi-head attention block 221 components can be similar to the multi-head attention block 209 of the encoder 201, except that the output trace 217 is masked such that the masked multi-head attention block 221 can only reference previous output trace positions. Additionally, an addition and normalization block 223 including a residual connection along with layer normalization (RLN) can connect the input of the masked multi-head attention block 221 to the output of the masked multi-head attention block 221, add them, and then normalize.

[0177] Following the addition and normalization block 223, there may be another multi-head attention block 225, which may be similar to the multi-head attention block 209 of the encoder 201, except that the multi-head attention block 225 uses the information from the encoder output as an additional input for calculating attention.

[0178] In addition, another addition and normalization block 227 may connect the input of the multi-head attention block 225 to the output of the multi-head attention block 225, add them, and then normalize. The resulting elements may further pass through the feed-forward sub-layer 229, followed by another addition and normalization block 231 on its path out of the decoder 203. The addition and normalization block 231 may connect the input of the feed-forward layer 229 to the output of the feed-forward layer 229, add them, and then normalize.

[0179] Outside the decoder 203, there may be a linear layer and a softmax block 233, which may obtain the output of the decoder 203 and create a logical vector over all possible output elements. In addition, the linear layer and the softmax block 233 may perform a softmax operation to transform the values in the logical vector to output a polynomial 111, which indicates the probability including the predicted arrival information. From the polynomial 111, the maximum value and its index may be extracted to find the predicted destination, or to generate multiple realizations (such as a continuation of an existing arrival stream) of the arrival time for reaching a specific one or more floors in the set of floors 129a to 129f using an elevator car (e.g., car 125 or car 127) as one or more destinations. In an example embodiment, when training the destination predictor 141 as a transformer-based destination predictor, the KL divergence loss function may be minimized over multiple epochs, and a greedy decoding scheme may be used as a method for finding the output destination.

[0180] Thus, unlike RNNs that process sequences in a sequential manner, which may prevent parallelization during training, the transformer architecture uses positional encoding and attention modules for parallelization, thereby enabling a more efficient training process. Due to parallelization, neural networks based on the transformer architecture are configured to process data at a high processing speed and are a preferred neural network architecture.

[0181] Training

[0182] In an example embodiment, the transformer-based destination predictor 141 may include N = 4 encoder / decoder blocks, along with a 4-level stacked long short-term memory (LSTM) / bidirectional long short-term memory (BiLSTM) model.

[0183] To train the destination predictor 141 based on the transformer architecture, different embodiments use real and / or simulated indoor movement. For example, SimTread simulation software can be used to generate movement trajectories in an indoor environment based on a created floor architecture. Additionally or alternatively, in one embodiment, a neural network 103 initially trained based on simulated data is further trained online during its operation based on actual measurements of partial trajectories and the final destination. When the training dataset contains a small number of trajectories, for a rested prediction model, it can lead to the leave-one-out cross-validation (LOOCV) technique for obtaining average performance.

[0184] In addition, for a given total of S samples (trajectories), S iterations can be performed, where in each iteration, one of the S trajectories can be used for model testing and the remaining S - 1 trajectories can be used for model training. In each iteration, an array of correct destination probabilities can be obtained by increasing the percentage of the query test trajectory. Such an array of probabilities can be averaged over all S iterations to obtain the average destination prediction performance for a given model.

[0185] In some embodiments, to obtain a continuation set, the destination predictor 141 can be trained on the training trajectories. Let be the ξ-th training trajectory, where is a tuple with (correspondingly, ) that represents the checkerboard layout index for the position (correspondingly, elapsed time) of the κ-th record for trajectory S (ξ) , and the last tuple represents the checkerboard layout index for the position and elapsed time at the destination, ξ = 1,..., N train . The set of training trajectories {S (ξ) ; ξ = 1,..., N train} can be obtained from the historical records (i.e., historical trajectories) extracted from the sensors 117. The trained destination predictor 141 can use the generated continuation set to determine the schedule 107.

[0186] The control system 101 can further include a prediction period or prediction duration T, where T represents the length of the time interval for potential future arrivals for generating continuations. T can be chosen smaller or larger depending on the availability of computing resources. On each floor in the set of floors 129a to 129f, let the sensors 117 monitor all pedestrian movements up to the current time t, and let the processor 137 identify a set of candidate future arrival partial trajectories (e.g., partial trajectory 113) within the time interval [t, t + T].

[0187] Let denote the ζ-th such partial trajectory, where is a tuple having (correspondingly, ) that represents the checkerboard layout index for the position (correspondingly, elapsed time) of the ι-th record for the trajectory , where ζ = 1,..., N . Based on the available partial trajectory information, the last tuple test can be further away from or closer to the tuple of the checkerboard layout index of the position and elapsed time at the final destination.

[0188] In addition, represent the start time of as represent the predicted arrival time as Then for a given input partial trajectory the actual output of the destination predictor 141 is a probability distribution P (ζ) (ω) over the set Ω of all candidate tuples of the checkerboard layout index of the position and elapsed time. Let Ω′ denote a subset of Ω that represents the tuples of the checkerboard layout index of the position and candidate elapsed time corresponding to the elevator doors on the elevator car on the corresponding floor. In addition, extract the set of probabilities {P (ζ) (ω); ω ∈ Ω′} of the subset Ω′ to generate a continuation set. Based on the prediction for the trajectory , then, by means of Monte Carlo simulation, different continuations of the arrival stream can be formed using the probability information in {P (ζ) (ω); ω ∈ Ω′}. Then, by considering the prediction information of all the available partial trajectories 113 a complete continuation set and further multiple continuation sets are obtained.

[0189] In some embodiments, the neural network 103 (LSTM-based neural network, neural network based on the transformer architecture, etc.) in the control system 101 can be trained online. In some other embodiments, the neural network 103 (LSTM-based neural network, neural network based on the transformer architecture, etc.) in the control system 101 can be trained offline (pre-trained). In some other embodiments, the neural network 103 (LSTM-based neural network, neural network based on the transformer architecture, etc.) in the control system can be partially offline trained and partially online trained.

[0190] Figure 2CIllustrate the layout of a floor in a building according to an embodiment of the present disclosure, where multiple trajectories of multiple pedestrians are used for training the destination predictor 141, and a partial trajectory of one pedestrian is used for testing the destination predictor 141.

[0191] In some embodiments, the partial trajectory 113 may include a trajectory that only contains location information, where each trajectory may be represented by a sequence of indices of a checkerboard layout. The indices of the checkerboard layout may be obtained by discretizing the range of position coordinates, where the range of position coordinates may represent the longitude and latitude associated with the checkerboard layout elements corresponding to the floor. In another embodiment, the partial trajectory 113 may include a trajectory that contains both location information and timing information. In addition, each trajectory may be represented by a sequence of tuples of indices of a checkerboard layout, where each tuple may contain the indices of the checkerboard layout obtained by discretizing the range of position coordinates and the prediction time period. The prediction time period may be the maximum time considered by the control system 101 for future requests of passengers.

[0192] As Figure 2C shown, the coordinates of the positions on floor 235 are discretized into the checkerboard layout indices 237, thus simplifying the representation of the trajectories. The continuous space of floor 235 is discretized into a 50×50 checkerboard layout, and each trajectory is converted into a sequence of checkerboard layout locations used as the input to the transformer architecture. Generally speaking, this data can be used to mimic the movement of a single user throughout the floor during a period of a day, where the visited destinations may represent the most visited real-life indoor destinations, such as stairways, elevators, and lounges. Each training trajectory 239 (i.e., historical trajectory) may have its own starting point 241 and final destination 243, and each test trajectory 245 (i.e., new trajectory) may have its own starting point 247, but the test trajectory 245 does not yet have a final destination 249, and the final destination 249 will be predicted by the destination predictor 141 of the control system 101. The destination predictor 141 trained using the training trajectories 239 can accurately predict the possible destinations of a person and the arrival time of the person at the corresponding destinations based on the partial trajectory of the person.

[0193] In an example embodiment, in a multi-story building, each floor may have its own layout and destination predictor 141. In another embodiment, the destination predictor 141 may also be constructed across multiple floors, as long as part of the trajectory 113 is outside of the elevator carriages (e.g., carriage 125 and carriage 127). Additionally, when generating a continuation set from the polynomial 111 indicating predicted arrival information of potential passengers, a parallel assistant taking into account the time of day may be incorporated. For example, during the down-peak period on a weekday, the destination floor of a typical passenger is considered to be the lobby rather than all the remaining floors with equal probability.

[0194] Furthermore, training the destination predictor 141 to output a predictive estimate in the form of a polynomial may have several advantages. First, the polynomial may replace the separate prediction of the probability of arriving at the elevator request area with a joint prediction of arriving at any destination (where the elevator request area is one of the destinations). Thus, when two conditions are met, the probability of arriving at the elevator request area is determined relative to other destinations.

[0195] The two conditions may include that the list of destinations is exhaustive and that the sum of all the values in the polynomial equals "1". The first requirement is satisfied by spatial quantization, and the second requirement is implemented during the training of the neural network 103 including the destination predictor 141. In this way, the extended destination prediction of the probability of arriving at the elevator request area can be simplified.

[0196] Figure 2D Plot 251 illustrating a comparison of the probabilities of destination predictions determined by a neural network based on a transformer architecture, a neural network based on long short-term memory (LSTM), and a neural network based on bidirectional long short-term memory (BiLSTM) according to an example embodiment of the present disclosure.

[0197] Plot 251 includes the percentage of observed test trajectories plotted on the X-axis and the probability of the correct destination plotted on the Y-axis. Plot 251 considers the median performance, along with the behavior of nearby quantiles. Plot 251 shows the 40th, 50th, and 60th percentiles of the destination probability to measure the variability of the test trajectory complexity. When performing the leave-one-out cross-validation (LOOCV) method, test trajectories of different lengths and similarities to the training trajectories may be encountered. Additionally, the destination probabilities obtained on some partitions may be different, and a robust destination prediction model should be able to distinguish different trajectories and accordingly alter its prediction, as can be seen with a greater spread in the predicted probabilities.

[0198] In addition, Plot 251 shows that a neural network based on a transformer architecture can obtain improved destination predictions early in the development of a trajectory, unlike LSTM-based neural networks and BiLSTM-based neural networks that require more observations to make accurate predictions. Additionally, the variation in the correct class probabilities for the transformer architecture neural network shows improved accuracy compared to the baseline method.

[0199] Figure 2E Plot 253 illustrating a comparison of the probabilities of destination predictions determined by a neural network based on a transformer architecture, an LSTM-based neural network, and a BiLSTM-based neural network according to an exemplary embodiment of the present disclosure.

[0200] Plot 253 corresponds to the probability of destination prediction when the true test trajectory destination probability is a function of the percentage of the observed test trajectory. Plot 253 includes the percentage of the observed test trajectory plotted on the X-axis and the probability of the correct destination plotted on the Y-axis. Plot 253 considers one partition of the LOOCV method as shown Figure 2E In this case, the test trajectory begins to overlap with some of the training trajectories early in its development. This overlap provides helpful information about the destination and can be recognized and utilized by the prediction model. The ideal model would detect the early overlap and account for it by increasing the true destination probability around the 20% mark of the test trajectory.

[0201] Therefore, these results show that a neural network based on a transformer architecture is able to learn patterns in this simplified, low-data environment (as evidenced by its increased destination probability), unlike LSTM-based neural networks and BiLSTM-based neural networks that cannot correctly detect the destination early in the development of a movement trajectory.

[0202] Figure 3 Illustrating steps of a method implemented for controlling the movement of elevator group 109 according to some embodiments of the present disclosure. The method may be performed by components of control system 101.

[0203] The method starts at step 301. At step 303, a current request from a person for the service provided by elevator group 109 can be obtained. The current request can be obtained from the person via a first input interface.

[0204] At step 305, a partial trajectory of the movement of the person can be obtained. The partial trajectory of the person moving on a floor (e.g., floor 129a) including the area served by elevator group 109 can be obtained from a second input interface 133, where the second input interface 133 can include various motion sensors.

[0205] In step 307, after receiving a partial trajectory of the person's movement, the neural network 103 trained to perform an extended destination prediction for the person having the movement can be executed based on the partial trajectory to generate a polynomial (e.g., polynomial 111) for the extended destination prediction of the person.

[0206] In step 309, the scheduling 107 of the elevator group 109 serving the current request and the future requests from the person can be optimized. The scheduling 107 can be optimized by providing time instances in the polynomial, the probability of the corresponding values determined by the checkerboard arrangement elements associated with the area served by the elevator group 109 in the polynomial, and providing elevator service.

[0207] In step 311, the elevator group 109 can be controlled according to the scheduling 107. The scheduling 107 ensures minimization of the average waiting time (AWT) for all passengers. The method ends at step 313.

[0208] Figure 4 Illustrate the steps of a method implemented for Monte Carlo simulation according to some embodiments of the present disclosure, the method for determining a scheduling for controlling the movement of an elevator group 109. The method can be executed by components of the control system 101.

[0209] The method starts from step 401. In step 403, the polynomial 111 generated by the neural network 103 can be obtained. The polynomial 111 has at least two dimensions, the at least two dimensions including a first dimension of the destination of the person and a second dimension of the time interval for the person to reach the destination of the first dimension, wherein the floors are quantized to form checkerboard arrangement elements of the destination of the first dimension, and at least one checkerboard arrangement element is associated with the area served by the elevator group 109.

[0210] In step 405, the values of the polynomial 111 can be filtered based on the destinations associated with the area served by the elevator group 109. In this way, other destinations indicated in the polynomial 111 that are not related to the area served by the elevator group 109 can be discarded to provide an accurate destination prediction.

[0211] In step 407, a Monte Carlo simulation of the filtered values of the polynomial 111 can be performed to generate multiple combinations of the time of the request. The occurrence frequency of a specific time instance among the multiple combinations of the time of the request is the probability of a future request for elevator service at the specific time instance, and the probability has a value at the intersection of the specific time instance and the checkerboard arrangement elements associated with the area served by the elevator group 109 in the polynomial.

[0212] In an alternative embodiment, this step is omitted because the samples from the previous step already include the arrival times. Some other embodiments select this step based on how the sampling is done. One method picks the basis elements of the polynomial according to the probability equal to the value in the basis element. Since each basis element corresponds to a combination of destination and time, the time component will already be present in the sample.

[0213] In step 409, each of the plurality of combinations of the requested times can be combined with the current request to generate a plurality of continuation sets. In some embodiments, the continuation sets are formed by filtering the destinations by only considering the areas served by elevator group 109.

[0214] In step 411, a schedule for elevator group 109 (e.g., schedule 107) can be determined, where schedule 107 is determined to optimize a performance metric for at least some of the passengers in all combinations of the continuation sets.

[0215] In step 413, elevator group 109 can be controlled according to schedule 107. Schedule 107 ensures minimization of the average wait time (AWT) for all passengers. The method ends at step 415.

[0216] In an example embodiment, motion sensors are installed on each floor of a building. For example, motion sensors can be installed in all cabins, corridors, cafeterias, etc. The motion sensors can be used to quantify each floor into a checkerboard layout or checkerboard layout elements. Thus, each checkerboard layout element corresponds to a motion sensor at a specific area installed on the floor. The checkerboard layout elements can be regular or irregular in shape. The motion sensors can always track the movement of one or more passengers.

[0217] Let the period of interest for predicting future arrivals be 10 seconds. The motion sensors can be configured such that when any of the motion sensors determines that there are more than a threshold amount of passengers near the elevator request area on the floor, the motion sensors can trigger the destination predictor 141 based on the neural network 103. The destination predictor 141 can then retrieve the partial trajectories 113 of all passengers within the past 10 seconds. The partial trajectory 113 of each passenger can be determined individually, or in combination with other passengers traveling or moving along the same route. Based on the retrieved data, the destination predictor 141 predicts the arrival information of potential future passengers in the form of a polynomial 111. The polynomial 111 can then be used by the control system 101 to determine the continuation sets, based on which the scheduler 105 determines the schedule 107.

[0218] In another example embodiment, the control system 101 may be configured to be triggered for each employee of a company arriving in the area served by the elevator. To this end, each employee in the building may be provided with an electronic identity card that can be read by motion sensors installed on each floor to identify the identity of the corresponding employee. The electronic identity card may be, for example, but not limited to, a radio frequency identification (RFID) card / tag. The control system 101 may include a database in which the identification information of all employees can be stored. The information may include the names of the employees, information about the floors on which the employees work, the historical trajectories of the employees, the work shift timings of the employees, etc.

[0219] Embodiments of the present disclosure include a control system 101 that controls the movement of an elevator group 109 based on a current request for service provided by one or more passengers to the elevator group 109 and a partial trajectory 113 of the movement of the one or more passengers on floors including one or more areas served by the elevator group 109. In a preferred embodiment, the control system 101 includes a neural network based on a transformer architecture, and the neural network is trained based on the partial trajectory 113 for extended destination prediction. Specifically, the control system 101 executes a neural network 103 based on the partial trajectory of each passenger to generate a polynomial for the extended destination prediction of the one or more passengers. The polynomial includes the possible destinations of each passenger and the possible time instances at which each passenger arrives at the destination. The polynomial is determined by quantifying the floors into checkerboard arrangement elements forming possible destinations and further quantifying the prediction time period into time instances. The prediction time period is the maximum time that the control system 101 considers for future requests of the one or more passengers.

[0220] Embodiments of the present disclosure perform a Monte Carlo simulation of the values of the polynomials determined for the checkerboard arrangement elements associated with the area served by the elevator group 109 to generate multiple combinations of request times. The occurrence frequency of a particular time instance in the multiple combinations of request times is the probability of a future request for elevator service at the particular time instance, and the probability has the value at the intersection of the particular time instance and the checkerboard arrangement elements associated with the area served by the elevator group 109 in the polynomial. Embodiments of the present disclosure combine each of the multiple combinations of request times with the current request to generate multiple continuation sets, and further determine a schedule 107 of the elevator group 109 to optimize a performance metric (e.g., average waiting time) for at least some of the passengers in all combinations of the continuation sets.

[0221] In addition, the various methods or processes outlined herein can be encoded as software executable on one or more processors employing any one of a variety of operating systems or platforms. Additionally, such software can be written using several suitable programming languages and / or programming or scripting tools, and can also be compiled into executable machine language code or intermediate code that is executed on a framework or virtual machine. Generally, in various embodiments, the functionality of program modules can be combined or distributed as desired.

[0222] Moreover, embodiments of the present disclosure can be implemented as a method, examples of which have been provided. The acts performed as part of the method can be ordered in any suitable way. Accordingly, embodiments can be constructed in which the acts are performed in an order different from that illustrated (which can include performing some acts simultaneously, even though shown as sequential in illustrative embodiments). Additionally, the use of ordinal terms (such as "first," "second") in the claims to modify claim elements does not in and of itself imply any precedence, priority, or order of one claim element over another claim element, or the temporal order of acts of a method, but is merely used as a label to distinguish one claim element having a certain name from another element having the same name (but using an ordinal term) to distinguish claim elements.

[0223] Although the present disclosure has been described with reference to certain exemplary embodiments, it is to be understood that various other changes and modifications can be made within the spirit and scope of the present disclosure. Accordingly, all such variations and modifications that fall within the true spirit and scope of the present disclosure are aspects of the appended claims.

Claims

1. A control system for controlling the movement of an elevator group, the control system comprising: A first input interface configured to receive a current request from a passenger for the service provided by the elevator group; A second input interface configured to receive a partial trajectory of the movement of a person moving on the floors served by the elevator group; And A processor configured to: Submit the partial trajectory to a neural network trained for extended destination prediction to generate a polynomial having at least two dimensions, the at least two dimensions including a first dimension of the destination of the person and a second dimension of the time interval for the person to reach the destination of the first dimension, wherein the floors are quantized to form checkerboard layout elements of the destination of the first dimension, and wherein at least one checkerboard layout element corresponds to a service area associated with the request for the service provided by the elevator group; Optimize the scheduling of the elevator group for serving the current request from the passenger and the future requests of the person to obtain elevator service at each time interval provided in the polynomial, with the probability of the corresponding value determined for the checkerboard layout element associated with the service area in the polynomial; And Control the elevator group according to the scheduling, Wherein the partial trajectory includes trajectory data containing only position information, wherein each trajectory is represented by a sequence of checkerboard layout indices, wherein the checkerboard layout indices are obtained by discretizing the range of position coordinates, and wherein the range of position coordinates represents the longitude and latitude associated with the checkerboard layout element.

2. A control system for controlling the movement of an elevator group, the control system comprising: A first input interface configured to receive a current request from a passenger for the service provided by the elevator group; A second input interface configured to receive a partial trajectory of the movement of a person moving on the floors served by the elevator group; And A processor configured to: Submit the partial trajectory to a neural network trained for extended destination prediction to generate a polynomial having at least two dimensions, the at least two dimensions including a first dimension of the destination of the person and a second dimension of the time interval for the person to reach the destination of the first dimension, wherein the floors are quantized to form checkerboard layout elements of the destination of the first dimension, and wherein at least one checkerboard layout element corresponds to a service area associated with the request for the service provided by the elevator group; Optimize the scheduling of the elevator group for serving the current request from the passenger and the future requests of the person to obtain elevator service at each time interval provided in the polynomial, with the probability of the corresponding value determined for the checkerboard layout element associated with the service area in the polynomial; And Control the elevator group according to the scheduling, Wherein the time interval of the second dimension quantizes a prediction time period, and wherein the prediction time period is the maximum time considered by the control system for the future requests of the person. The partial trajectory includes trajectory data that includes both position information and timing information, where each trajectory is represented by a sequence of tuples indexed by a checkerboard layout, and where each tuple contains a checkerboard layout index obtained by discretizing the range of position coordinates and the prediction time period.

3. The control system according to claim 2, wherein timing information of a trajectory is extracted by considering relative time with respect to elapsed time, the elapsed time being relative to their start timestamps.

4. The control system according to claim 3, wherein each trajectory is represented by a sequence of symbols, and wherein each symbol is mapped from a tuple indexing a checkerboard arrangement indicating the position information and the timing information.

5. A control system for controlling the movement of an elevator group, the control system comprising: A first input interface configured to receive a current request from a passenger for service provided by the elevator group; A second input interface configured to receive a partial trajectory of the movement of a person moving on a floor served by the elevator group; And A processor configured to: Submit the partial trajectory to a neural network trained for extended destination prediction to generate a polynomial having at least two dimensions, the at least two dimensions including a first dimension of the destination of the person and a second dimension of the time interval for the person to reach the destination of the first dimension, where the floors are quantized to form checkerboard layout elements of the destination of the first dimension, and where at least one checkerboard layout element corresponds to a service area associated with a request for service provided by the elevator group; Optimize the scheduling of the elevator group for serving the current request from the passenger and future requests of the person to obtain elevator service at each time interval provided in the polynomial, with the probability of the corresponding value determined for the checkerboard layout element associated with the service area in the polynomial; And Control the elevator group according to the scheduling, Where the neural network has a transformer architecture, Where the transformer architecture uses positional encoding and attention mechanisms that enable parallelization.

6. A method for controlling the movement of an elevator of an elevator group, the method comprising: Obtain a current request from a passenger for service provided by the elevator group; Obtain a partial trajectory of the movement of a person moving on a floor including the area served by the elevator group; After receiving the partial trajectory of the movement of the person, according to the partial trajectory, execute a neural network trained for extended destination prediction of the person with the movement to generate a polynomial for the extended destination prediction of the person, where the polynomial has at least two dimensions, the at least two dimensions including a first dimension of the destination of the person and a second dimension of the time interval for the person to reach the destination of the first dimension, where the floors are quantized to form checkerboard layout elements of the destination of the first dimension, and where at least one checkerboard layout element is associated with the area served by the elevator group; Optimize the scheduling of the elevator group for serving the current request from the passenger and future requests of the person to obtain elevator service at each time interval provided in the polynomial, with the probability of the corresponding value determined for the checkerboard layout element associated with the area served by the elevator group in the polynomial; And Control the elevator group according to the scheduling, The partial trajectory includes trajectory data containing only position information, where each trajectory is represented by a sequence of indices in a checkerboard arrangement, the checkerboard arrangement indices being obtained by discretizing the range of position coordinates, and the range of position coordinates representing the longitude and latitude associated with the checkerboard arrangement elements.

7. A method for controlling the movement of an elevator in an elevator group, the method comprising: Obtain a current request from a passenger for the service provided by the elevator group; Obtain a partial trajectory of the movement of a person moving on a floor in an area served by the elevator group; After receiving the partial trajectory of the movement of the person, according to the partial trajectory, execute a neural network trained for extended destination prediction of the person having the movement to generate a polynomial for the extended destination prediction of the person, where the polynomial has at least two dimensions, the at least two dimensions including a first dimension of the destination of the person and a second dimension of the time interval for the person to reach the destination in the first dimension, where the floors are quantified to form checkerboard arrangement elements of the destination in the first dimension, and where at least one checkerboard arrangement element is associated with the area served by the elevator group; Optimize the scheduling of the elevator group for serving the current request from the passenger and the future requests of the person to obtain elevator service at each time interval provided in the polynomial, with the probability of the corresponding value determined for the checkerboard arrangement element associated with the area served by the elevator group in the polynomial; and Control the elevator group according to the scheduling, where the time interval in the second dimension quantifies a prediction time period, and where the prediction time period is the maximum time considered for the future requests of the person, The partial trajectory includes trajectory data containing both position information and timing information, where each trajectory is represented by a sequence of tuples of checkerboard arrangement indices, and where each tuple contains checkerboard arrangement indices obtained by discretizing the range of position coordinates and the prediction time period.

8. A method for controlling the movement of an elevator in an elevator group, the method comprising: Obtain a current request from a passenger for the service provided by the elevator group; Obtain a partial trajectory of the movement of a person moving on a floor in an area served by the elevator group; After receiving the partial trajectory of the movement of the person, according to the partial trajectory, execute a neural network trained for extended destination prediction of the person having the movement to generate a polynomial for the extended destination prediction of the person, where the polynomial has at least two dimensions, the at least two dimensions including a first dimension of the destination of the person and a second dimension of the time interval for the person to reach the destination in the first dimension, where the floors are quantified to form checkerboard arrangement elements of the destination in the first dimension, and where at least one checkerboard arrangement element is associated with the area served by the elevator group; Optimize the scheduling of the elevator group for serving the current request from the passenger and the future requests of the person to obtain elevator service at each time interval provided in the polynomial, with the probability of the corresponding value determined for the checkerboard arrangement element associated with the area served by the elevator group in the polynomial; and Control the elevator group according to the scheduling, wherein the neural network has a transformer architecture, wherein the transformer architecture uses positional encoding and attention mechanisms that enable parallelization.

Images