Elevator passenger flow prediction method and system based on spatio-temporal graph neural network

Abstract

The application discloses an elevator passenger flow prediction method and system based on a space-time graph neural network, and relates to the technical field of intelligent elevators.The method comprises the following steps: constructing graph nodes of each floor of a building and passenger flow conduction relationships between floors to obtain a space correlation feature vector; encoding time dimension variation rules of passenger flow through the space correlation feature vector to obtain a time evolution feature vector; and fusing visual perception data and building scene data through the time evolution feature vector to predict the number of passengers waiting for an elevator every minute of each floor, the distribution proportion of target floors of each floor, and the proportion of heavy objects carried by each floor.The application realizes minute-level accurate passenger flow prediction through a space-time graph neural network, drives an elevator system to actively and globally perform intelligent scheduling, and thus significantly improves transport capacity during peak periods and effectively reduces energy consumption during off-peak periods.

Description

Technical Field

[0001] This invention relates to the field of intelligent elevator technology, and in particular to a method and system for predicting elevator passenger flow based on spatiotemporal graph neural networks. Background Technology

[0002] Elevator systems are the core of vertical transportation in modern high-rise buildings, and the efficiency of their scheduling algorithms directly determines passenger waiting experience, building operational efficiency, and energy consumption levels. Traditional elevator group control systems mostly employ rule-based scheduling strategies or optimization methods based on classical mathematical models. However, with increasingly complex building structures and more dynamic and diverse pedestrian flow patterns, these traditional methods primarily rely on real-time signals triggered by floor selection buttons inside the elevator car and call buttons outside the lobby for scheduling decisions, representing a typical "request-response" model. This model cannot predict the waiting demand on each floor in the next few minutes, leading to delayed responses to sudden passenger flow events such as meeting dismissals or lunch breaks, easily causing instantaneous congestion and long waiting times for passengers. Summary of the Invention

[0003] The main objective of this invention is to provide a method and system for predicting passenger flow in elevators based on spatiotemporal graph neural networks, aiming to solve the technical problem of scheduling delay caused by the lack of forward-looking perception in elevator group control scenarios without power grid or requiring extreme energy saving.

[0004] To achieve the above objectives, this invention proposes a method for predicting elevator passenger flow based on a spatiotemporal graph neural network, comprising: Construct graph nodes for each floor of the building and the passenger flow transmission relationship between floors to obtain spatial correlation feature vectors; By using spatial correlation feature vectors, the temporal dimension change pattern of people flow is encoded to obtain temporal evolution feature vectors; By fusing visual perception data and building scene data using time evolution feature vectors, the system can predict the number of people waiting for elevators on each floor per minute, the proportion of target floors on each floor, and the proportion of heavy objects carried on each floor within a time window.

[0005] Furthermore, each floor's waiting area is considered an independent node in the graph. The node features include three types of real-time sensing data: the number of people waiting for the elevator, the percentage of people carrying heavy items, and the current passenger flow density. The real-time sensing data includes the number of people waiting for the elevator, the position of people, and the status of carrying heavy items. The real-time sensing data is collected by cameras and millimeter-wave radar deployed on each floor.

[0006] Furthermore, constructing the graph nodes for each floor of the building and the passenger flow transmission relationships between floors includes: Construct weighted edges between graph nodes to form a spatial graph, wherein the weighted edges include physical association edges and passenger flow transmission edges; The spatial graph is encoded using a graph convolutional network to obtain the spatial association feature vector.

[0007] Furthermore, the weighted edges include physical association edges and passenger flow transmission edges; Physical association edges are weighted according to the adjacent relationship of floors, with adjacent floors having a higher weight than cross-floor edges, and cross-floor edges having a lower weight as the floor difference increases; passenger flow transmission edges are weighted based on the percentage of passenger flow from the first floor to the second floor according to historical data, with a higher percentage resulting in a higher weight.

[0008] Furthermore, the step of encoding the temporal dimension change pattern of pedestrian flow through spatial correlation feature vectors to obtain temporal evolution feature vectors specifically includes: Obtain the time series data of the spatial correlation feature vector; The time series data is encoded using a recurrent neural network to obtain the temporal evolution feature vector of pedestrian flow on each floor.

[0009] Furthermore, in the step of encoding the time series data of the spatial correlation feature vector based on the recurrent neural network, a fine-grained sliding time window is used to process the time series data, wherein the step size of the sliding time window is t, the historical time window is 2×t minutes, and the prediction time window is t minutes.

[0010] Furthermore, the fusion of visual perception data and architectural scene data through time-evolutionary feature vectors specifically includes: The quality of the real-time sensing data is evaluated, and the evaluation results are obtained; Based on the evaluation results, the weight ratio of the time evolution feature vector and the architectural scene data during fusion is dynamically adjusted. Specifically, when the quality of the perceived data is lower than a preset threshold, the weight ratio of the building scene data is increased. The building scene data includes: building operation schedule, meeting schedule and holiday information.

[0011] Furthermore, this method also includes: determining the flow status of the prediction window based on the calculated number of people waiting for the elevator on each floor, wherein the flow status includes: peak state, off-peak state or low-peak state; and obtaining the passenger flow direction trend of each floor based on the calculated distribution ratio of the target floors, wherein the direction trend includes upward or downward.

[0012] Furthermore, this method also includes: executing corresponding scheduling strategies based on passenger flow status; wherein the scheduling strategies include: in peak conditions, performing elevator coupling scheduling by floor zone; in off-peak conditions, coupling scheduling of elevators in pairs; and in low-peak conditions, placing some elevators in a dormant state.

[0013] This invention also proposes an elevator passenger flow prediction system based on a spatiotemporal graph neural network, comprising a processor and a memory. The memory stores a computer program. When the processor executes the computer program to implement the elevator passenger flow prediction method based on the spatiotemporal graph neural network, it does so by constructing and running a spatiotemporal graph neural network model. The model adopts the following four-layer sequential architecture: The spatial graph coding layer is used to construct the graph nodes of each floor of the building and the passenger flow transmission relationship between floors in order to obtain spatial association feature vectors. The time series coding layer is used to encode the temporal dimension changes of people flow by spatially correlated feature vectors, so as to obtain time evolution feature vectors. The multi-source fusion layer is used to fuse visual perception data and architectural scene data through time-evolved feature vectors; The predictive output layer is used to predict the number of people waiting for the elevator on each floor every minute within a time window, the distribution ratio of the target floor on each floor, and the proportion of heavy objects carried on each floor.

[0014] This invention uses a spatiotemporal graph neural network to achieve minute-level accurate passenger flow prediction, driving the elevator system to perform proactive and global intelligent scheduling, thereby significantly improving transport capacity during peak hours and effectively reducing energy consumption during off-peak hours. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the elevator passenger flow prediction method based on spatiotemporal graph neural network of the present invention. Figure 2 This is a schematic diagram of the architecture of the spatiotemporal graph neural network model provided by the elevator passenger flow prediction system based on spatiotemporal graph neural network of the present invention.

[0018] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of the present invention and are not intended to limit the present invention.

[0020] To better understand the technical solution of the present invention, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, Figure 1 This is a flowchart illustrating the elevator passenger flow prediction method based on spatiotemporal graph neural network provided by the present invention.

[0022] This invention proposes a method for predicting elevator passenger flow based on a spatiotemporal graph neural network, comprising: S10, construct the graph nodes of each floor of the building and the passenger flow transmission relationship between floors to obtain the spatial association feature vector; S20 uses spatial correlation feature vectors to encode the temporal dimension change pattern of people flow in order to obtain temporal evolution feature vectors. S30 uses time-evolution feature vectors to fuse visual perception data and building scene data to predict the number of people waiting for elevators on each floor per minute, the distribution ratio of target floors on each floor, and the proportion of heavy objects carried on each floor within a time window.

[0023] The system in this embodiment is deployed in a 30-story office building equipped with 8 group-controlled elevators. The system aims to achieve proactive and intelligent scheduling of the elevator group control system by accurately predicting future short-term passenger flow.

[0024] In this embodiment, the spatiotemporal graph neural network model adopts a four-layer sequential architecture consisting of a spatial graph encoding layer, a time series encoding layer, a multi-source fusion layer, and a prediction output layer. The model's input integrates multi-dimensional perception data, specifically including: a) visual and radar perception data, collected in real-time at intervals t0 (e.g., 10 seconds) to track the number of people waiting for elevators on each floor, their positions, the status of heavy objects being carried, and pedestrian density; b) building scene data, updated at intervals t1 (t1>t0, e.g., 1 minute), including information such as weekdays / holidays, operating hours, and meeting schedules; c) historical data (stored for t2>=7 days) used for model training and optimization. The data period satisfies t2>t>t1>t0, where t is the predicted future time window. The model's output closely aligns with scheduling requirements, including: the number of people waiting for elevators on each floor per minute in the future time t (accuracy ±1 person), used to predict the load and allocate the roles of power users / generators; the distribution percentage of the target floors on each floor in the future time t (accuracy ±5%), used to predict the direction of movement and match coupling objects; the percentage of heavy objects carried on each floor in the future time t (accuracy ±3%), used to correct the load prediction; and the peak / off-peak / low-peak status determination G calculated based on X, used to trigger different global scheduling strategies (such as partition coupling, pairwise coupling, and hibernation).

[0025] S10, construct the graph nodes of each floor of the building and the passenger flow transmission relationship between floors to obtain the spatial association feature vector; In this embodiment, real-time sensing data is collected through a fusion sensing terminal combining visual and millimeter-wave radar deployed in the waiting areas of each floor. The visual sensor captures images of the waiting area and identifies the number of people and heavy objects using a target detection algorithm. The millimeter-wave radar generates point cloud data of personnel positions and movement trajectories to fill in visual blind spots. The data from both sensors undergoes spatiotemporal alignment and fusion processing within an edge computing unit, generating structured real-time sensing data for each floor with a period of t. This data includes at least the number of people waiting, personnel positioning distribution, heavy object carrying status indicators, and real-time passenger flow density calculated based on the waiting area area, where t is set to 10 seconds. Building scene data is automatically acquired through system interfaces with the building property management system and the enterprise calendar system, updated with a period of t. This building scene data includes building operation time phase attributes, preset meeting and event schedules, and their corresponding floor and time information. The historical data includes real-time sensing data from the same historical period and actual passenger flow statistics for the corresponding time period, and is used for model training and statistical calculation of passenger flow transmission weights.

[0026] In this embodiment, the quality of real-time perceived data is assessed before data transmission, and an assessment result including confidence index is generated to provide a basis for the dynamic weight adjustment of the subsequent fusion layer.

[0027] In this embodiment, the construction of the spatial graph structure first defines the waiting area of ​​each floor as an independent graph node, and constructs a node feature vector for each graph node; the node feature vector is directly mapped from the real-time sensing data of the corresponding floor, specifically including the following normalized scalar dimensions: the current number of people waiting for the elevator, the current proportion of people carrying heavy objects, and the current passenger flow density.

[0028] Furthermore, weighted directed edges are constructed between each pair of related graph nodes. These weighted directed edges include two types: physical connection edges and passenger flow transmission edges. Physical connection edges are constructed based on the physical adjacency between floors, and their weight values ​​are set according to the following rules: for adjacent floor nodes (floors with a floor number difference of 1), a basic weight value of Wa is set; for non-adjacent floor nodes, their weight value decreases linearly as the floor number difference increases, i.e., weight Wp = max(0, Wa-α). (|ij|-1)), where i and j are floor numbers, and α is a decreasing coefficient. This setting makes the correlation between floors that are closer in space stronger. The passenger flow transmission edge is constructed based on the cross-floor passenger flow migration pattern statistically analyzed in historical data. For any two nodes A and B, the weight value Wf(A, B) of the passenger flow transmission edge from A to B is proportional to the statistical proportion of passenger flow originating from floor A and destination floor B in historical data. The higher the proportion, the higher the weight. This proportion is normalized so that the sum of the weights of all passenger flow transmission edges originating from a single source node is a constant value. This setting makes frequent passenger flow migration paths stand out in the spatial graph. By integrating all the graph nodes and the weighted directed edges connecting them, a weighted directed spatial graph is finally generated. This graph structure completely encapsulates the static physical correlation and dynamic transmission pattern of passenger flow in the spatial dimension of the building, providing structured input for subsequent graph convolutional networks to encode spatial correlation features.

[0029] Here, α is set by comprehensively considering the total number of floors, functional zoning, and typical passenger flow patterns. For most high-rise buildings (e.g., 20-50 floors), after extensive simulation testing and verification with real data, the typical effective value range for α is 0.05 to 0.20. Within this range: smaller values ​​(e.g., α=0.05) are suitable for buildings with mixed floor functions and frequent cross-floor passenger flow (e.g., open-plan office floors), emphasizing broad spatial connections; larger values ​​(e.g., α=0.15) are suitable for buildings with clear floor functional zoning and passenger flow mainly concentrated between adjacent floors (e.g., hotels, apartments), emphasizing close local connections.

[0030] In this embodiment, for the 30-story office building, after optimization, α=0.10 is selected as the base value, and the base weight Wa of adjacent floors is set to 0.5.

[0031] In another embodiment, α can be obtained based on a statistical fitting method using historical data. Specifically, by analyzing historical passenger flow data, the correlation strength between any two non-adjacent floor passenger flow change sequences is calculated as an observed value, and an optimization algorithm is used to minimize the overall error between the predicted physical correlation weight value Wp(α) and the observed correlation strength, thereby solving for the optimal value of α.

[0032] The core objective of this step is to address spatial correlation prediction issues such as whether the concentration of people in lower areas foreshadows future growth in people in higher areas. Subsequently, a graph convolutional network is used to encode the weighted spatial graph. By aggregating neighbor node information, spatial correlation feature vectors for each floor are output, thereby capturing the pattern that "increased pedestrian flow on a certain floor will drive future changes in pedestrian flow on its related floors (physically adjacent or frequently visited destinations)".

[0033] S20 uses spatial correlation feature vectors to encode the temporal dimension change pattern of people flow in order to obtain temporal evolution feature vectors. In this embodiment, a graph convolutional network is used to encode the constructed weighted directed spatial graph to extract the spatial association features of each floor node. The specific encoding process is as follows: The structural information of the spatial graph is abstracted into a weighted adjacency matrix A. The value of the matrix element A{i, j} corresponds to the comprehensive weight of the directed edge from node i to node j, where i and j are the index identifiers of the graph nodes. This comprehensive weight is obtained by weighted summation of the weights of the physical association edges and the passenger flow transmission edges. Here, i represents the floor number corresponding to a source node or the currently focused node, and j represents the floor number corresponding to a target node or another node associated with node i.

[0034] Furthermore, the feature vectors of each node are stacked to form a node feature matrix X. The graph convolutional network uses multiple graph convolutional layers for feature propagation and transformation; the basic operation of each layer is defined as follows: ,in, =A+I, where I is the identity matrix, representing adding a self-connect to each node to preserve its own characteristics; for The degree matrix, It is the inverse square root of the degree matrix, used to normalize the adjacency matrix to avoid excessive amplification of features by nodes with high degree (such as one floor connecting multiple floors) during aggregation, so as to make the weighting of information of each floor more equitable; Let be the feature matrix of the nodes in the k-th layer. ; σ is the trainable parameter weight matrix of the k-th layer; σ is the non-linear activation function. Through this operation, the new features of each node aggregate the feature information of all its first-order neighboring nodes and are weighted and averaged according to the edge weights, thus realizing the transmission and diffusion of information in the graph structure. The network stacks two of the above graph convolutional layers, so that the final feature vector of each node can aggregate the node information in its two-hop neighborhood, thereby capturing a wider range and multi-step passenger flow transmission impact. After being encoded by this graph convolutional network, each floor node is mapped to a fixed-dimensional, high-order spatial association feature vector. This vector not only encodes the real-time perception data of the node itself, but also aggregates the state information of related floors through the graph structure, quantitatively expressing the potential impact of changes in passenger flow on a certain floor on the future passenger flow of its related floors due to physical proximity or historical passenger flow migration patterns.

[0035] In this embodiment, the time-series data of the output spatial correlation feature vectors of each floor are encoded based on a recurrent neural network to extract the evolutionary pattern of pedestrian flow status on each floor in the time dimension, thereby obtaining the time evolution feature vector of pedestrian flow on each floor. Specifically, the time-series data is organized as follows: a fixed time step t (e.g., 5 minutes) is used as the step size of the sliding window to continuously extract spatial correlation feature vectors within historical time periods; the length of each input sequence is 2×t, where each element in the sequence is a matrix composed of spatial correlation feature vectors of all floors at a given time; through this sliding window mechanism, continuous, equally spaced time-series samples reflecting recent historical states are constructed.

[0036] This step employs a fine-grained sliding time window strategy, the core purpose of which is to address the problem of predicting tidal fluctuations in elevator passenger flow (such as morning and evening peak hours). Specifically, the prediction window slides in steps t (e.g., 5 minutes), and the historical window takes 2×t minutes of data to cover recent changes, predicting the results for the next t minutes, thereby meeting the scheduling layer's need for advance prediction.

[0037] Furthermore, a gated recurrent unit network (GRU) is employed as the core encoder. The GRU, a core module of the recurrent neural network, transforms iterative multiplication operations into cumulative operations, resolving gradient explosion or vanishing problems. Through its unique update and reset gate mechanisms, this network adaptively memorizes long-term dependency information and filters short-term noise, effectively capturing periodic, trend-based, and sudden change patterns in pedestrian flow data. For each floor node, the evolution of its spatial correlation feature vector over time is independently input into a shared-parameter GRU network for processing. This network iterates step-by-step, updating its hidden state over time. The final hidden state of each time step represents the initial temporal feature of that floor. Further, to enhance the model's ability to focus on key time-period features, a temporal attention mechanism is introduced. This mechanism calculates the attention weight distribution of the hidden states at all time steps in the sequence. For time steps similar to typical peak-period patterns or exhibiting drastic changes, the model automatically assigns higher attention weights. By weighted summing of the hidden states at all time steps, a richer temporal evolution feature vector incorporating global temporal information is generated. This vector not only encodes the recent historical changes in passenger flow on this floor, but also highlights the most critical temporal pattern features for future prediction through an attention mechanism. This step, by combining a gated recurrent unit with an attention mechanism, enables the model to accurately model the complex temporal dynamics of elevator passenger flow, such as tidal and event-driven characteristics. It effectively solves the technical problem that relying solely on spatial features or simple temporal statistics cannot predict inflection points in passenger flow trends (such as peak arrivals or meeting departures), providing a feature representation with strong temporal discriminative power for final multi-source fusion and accurate prediction.

[0038] S30 uses time-evolution feature vectors to fuse visual perception data and building scene data to predict the number of people waiting for elevators on each floor per minute, the distribution ratio of target floors on each floor, and the proportion of heavy objects carried on each floor within a time window.

[0039] In this embodiment, the generated temporal evolution feature vector and the acquired architectural scene data are adaptively weighted and fused to generate a more robust fused feature vector. The specific fusion process is as follows: the quality of the real-time sensing data is quantitatively evaluated to generate a data quality evaluation result; the evaluation is calculated based on the integrity of the real-time sensing signal, the signal-to-noise ratio, and the consistency index among multiple sensors. When any index is lower than a preset threshold, the sensing data is determined to be of abnormal quality. Simultaneously, the architectural scene data is converted into a scene feature vector aligned with the dimension of the temporal evolution feature vector through an embedding layer; the scene feature vector encodes the scene attributes and event information of the current and near-current time periods. The core of the fusion process is the use of a learnable attention-weighted fusion module. This module receives temporal evolution feature vectors and scene feature vectors as input, and its attention weight generation mechanism is dynamically coupled with the data quality assessment results. Specifically, when the perceived data quality assessment result is normal, the attention mechanism assigns a higher dominant weight (e.g., 0.7-0.9) to the temporal evolution feature vector, making the fusion primarily based on real-time perception-driven temporal prediction. When the perceived data quality assessment result is abnormal, the weight ratio of the scene feature vector is automatically increased (e.g., increased to 0.4-0.6), making the fusion rely more on the prior rules and schedule knowledge carried by the building scene data, and outputs a unified fusion feature vector for each floor through weighted summation. This step addresses the critical technical challenge of drastically degrading prediction model performance in actual elevator deployment environments when sensor data becomes invalid or loses credibility due to temporary sensor obstruction or drastic changes in lighting. By introducing a dynamic weighted fusion mechanism linked to data quality, this step transforms the system from a fragile data-driven system into a robust system with knowledge assistance and fault mitigation capabilities, ensuring the continuity and reliability of pedestrian flow prediction outputs in diverse and non-ideal real-world operating environments.

[0040] Based on the generated fused feature vector, a multi-task output network is used for calculation and mapping to generate quantitative prediction parameters that can be directly used for elevator scheduling within the prediction time window. These parameters include: the number of people waiting for elevators on each floor per minute, the proportion of target floors on each floor, and the proportion of heavy objects carried on each floor. The specific implementation of this step is as follows: First, the fused feature vector is input into a fully connected neural network with a shared bottom layer for feature deepening and decoupling. The output layer of this network is divided into three parallel specific task branches. The first task branch is responsible for calculating the number of people waiting for elevators on each floor per minute. It uses a linear activation function combined with non-negativity constraints to map the features into a non-negative real matrix representing the predicted number of people on each floor for each minute within the next t minutes. This output directly quantifies the spatiotemporal distribution of the future short-term load. The second task branch is responsible for calculating the distribution ratio of target floors for each floor. It first maps the features to a target preference score matrix for each floor using a fully connected layer. Then, it normalizes each row of this matrix using a normalized exponential function, resulting in a probability distribution matrix. The elements of this matrix represent the predicted passenger flow ratio from a given departure floor to each target floor, with the sum of the ratios for all target floors being 1. This output accurately depicts the spatial flow of passengers. The third task branch is responsible for calculating the proportion of passengers carrying heavy items on each floor. It maps the features to the predicted percentage of passengers carrying heavy items on each floor within the prediction window using a fully connected layer with a sigmoid activation function. This value is a scalar between 0 and 1.

[0041] Furthermore, the three branch tasks employ a multi-task joint training strategy, sharing underlying feature representations and optimizing them through a composite loss function. This composite loss function integrates the mean squared error of predicted number of people, the cross-entropy loss of passenger flow distribution ratio, and the binary cross-entropy loss of heavy object ratio. This step accurately decodes the high-dimensional abstract fusion features into operational parameters with clear physical meaning and precision requirements that the scheduling system can directly understand and execute, achieving end-to-end mapping from data perception to control commands. Through multi-task joint learning, the model shares the inherent correlation information between passenger flow status, direction, and composition during optimization, significantly improving the consistency and overall accuracy of various predictions. This solves the technical problem of isolated prediction indicators and poor coordination in traditional methods, providing a complete and reliable decision-making basis for subsequent refined and differentiated elevator scheduling strategies based on prediction.

[0042] In this embodiment, the elevator passenger flow prediction method based on spatiotemporal graph neural network further includes: determining the passenger flow status of the prediction window based on the calculated number of people waiting on each floor, wherein the passenger flow status includes: peak state, off-peak state, or low-peak state; obtaining the passenger flow direction trend of each floor based on the calculated target floor distribution ratio, wherein the direction trend includes upward or downward. A corresponding scheduling strategy is executed based on the passenger flow status; wherein the scheduling strategy includes: in the peak state, elevator coupling scheduling is performed by floor zone; in the off-peak state, elevators are coupled and scheduled in pairs; in the low-peak state, some elevators are placed in a dormant state.

[0043] Specifically, the system automatically determines the overall passenger flow status during the predicted time window by comparing the average or maximum number of passengers waiting for elevators on each floor per minute with preset thresholds: if the average exceeds the first preset threshold, it is considered a peak period; if the average is below the second preset threshold, it is considered a low-peak period; and if it falls between the two, it is considered a non-peak period. Simultaneously, based on the distribution ratio of target floors on each floor, the system dynamically obtains the passenger flow trend for that floor by calculating the difference between the total percentage of passengers heading to higher floors and the total percentage heading to lower floors; if the difference is greater than zero, the trend is upward, and vice versa. Based on this, the system automatically triggers and executes pre-configured differentiated scheduling strategies according to the determined passenger flow status: When in peak conditions, a zone coupling scheduling strategy is activated, which divides all elevators into two or more independent operating zones according to the service floors, and the elevators in each zone are coupled in coordination to maximize vertical transportation throughput and cope with concentrated large passenger flows; when in off-peak conditions, a pair coupling scheduling strategy is activated, which pairs elevators together and they cooperate throughout the entire building floor range to optimize response efficiency while ensuring basic service coverage; when in low-peak conditions, a hibernation strategy is activated, which, based on real-time demand prediction, puts some elevators into low-power hibernation or standby mode until the predicted passenger flow rebounds, and then automatically wakes them up, thereby achieving energy saving and consumption reduction. The technical effect of this step is that it establishes a closed-loop linkage between the refined predictive output of the front-end artificial intelligence model and the dynamic resource allocation of the back-end elevator group control system, realizing full-link intelligence from perception and prediction to decision-making and execution. Through state-driven adaptive strategy switching, the system can significantly reduce the average waiting time for passengers during peak hours and significantly reduce the building's ineffective energy consumption during low-demand periods, fundamentally solving the core technical problem that traditional fixed strategies cannot achieve a dynamic optimal balance between efficiency and energy efficiency.

[0044] In an embodiment of the present invention, a month-long field comparative test was conducted in a 30-story intelligent office building equipped with 8 group-controlled elevators. This verified the significant technical effect of the elevator passenger flow method based on spatiotemporal graph neural network compared to the traditional instant call system. The test showed that the system's prediction accuracy reached the design target. During the morning peak hours, the prediction error of the number of people waiting on the first floor in the next 5 minutes was within ±1 person, and the error in the distribution ratio of the target floor was less than ±5%. Based on the dynamic scheduling strategy driven by accurate prediction, the average waiting time during the morning peak hours was reduced from 45 seconds to 33 seconds, and the long waiting rate was reduced from 18% to 9%. In sudden scenarios such as the end of a meeting, the system can respond 5 minutes in advance, reducing the average waiting time of relevant passengers by 56% compared to the traditional system. During off-peak hours, the intelligent sleep strategy can reduce energy consumption by 42% within 2 hours and wake up in advance before the demand is predicted, without affecting service. When the simulated perception data is abnormal due to camera obstruction, the system can maintain scheduling stability through a dynamic fusion mechanism.

[0045] like Figure 2 As shown, Figure 2 This is a schematic diagram of the architecture of the spatiotemporal graph neural network model provided by the elevator passenger flow prediction system based on spatiotemporal graph neural network of the present invention.

[0046] refer to Figure 2 This invention also proposes an elevator passenger flow prediction system based on a spatiotemporal graph neural network, including a processor and a memory. The memory stores a computer program. When the processor executes the computer program to implement the elevator passenger flow prediction method based on the spatiotemporal graph neural network, it does so by constructing and running a spatiotemporal graph neural network model. The model adopts the following four-layer sequential architecture: Spatial graph coding layer 10 is used to construct the graph nodes of each floor of the building and the passenger flow transmission relationship between floors in order to obtain spatial association feature vectors; The time series coding layer 20 is used to encode the temporal dimension change pattern of people flow by spatial correlation feature vectors in order to obtain time evolution feature vectors. Multi-source fusion layer 30 is used to fuse visual perception data and architectural scene data through time-evolved feature vectors; The prediction output layer 40 is used to predict the number of people waiting for the elevator on each floor per minute, the distribution ratio of the target floor on each floor, and the proportion of heavy objects carried on each floor within the predicted time window.

[0047] The above description is only a part of the embodiments of the present invention and does not limit the patent scope of the present invention. All equivalent structural transformations made under the technical concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for predicting elevator passenger flow based on a spatiotemporal graph neural network, characterized in that, include: Construct graph nodes for each floor of the building and the passenger flow transmission relationship between floors to obtain spatial correlation feature vectors; By using spatial correlation feature vectors, the temporal dimension change pattern of people flow is encoded to obtain temporal evolution feature vectors; By fusing visual perception data and building scene data using time evolution feature vectors, the system can predict the number of people waiting for elevators on each floor per minute, the proportion of target floors on each floor, and the proportion of heavy objects carried on each floor within a time window.

2. The method according to claim 1, characterized in that, The graph nodes are independent nodes for each floor's waiting area. The node features include three types of real-time sensing data: the number of people waiting for the elevator, the percentage of people carrying heavy items, and the current passenger flow density. The real-time sensing data includes the number of people waiting for the elevator, the position of people, and the status of carrying heavy items. The real-time sensing data is collected by cameras and millimeter-wave radar deployed on each floor.

3. The method according to claim 1, characterized in that, Constructing the graphical nodes of each floor of the building and the passenger flow transmission relationship between floors includes: Construct weighted edges between graph nodes to form a spatial graph, wherein the weighted edges include physical association edges and passenger flow transmission edges; The spatial graph is encoded using a graph convolutional network to obtain the spatial association feature vector.

4. The method according to claim 3, characterized in that, The weighted edges include physical association edges and passenger flow transmission edges; Physical association edges are weighted according to the adjacent relationship of floors, with adjacent floors having a higher weight than cross-floor edges, and cross-floor edges having a lower weight as the floor difference increases; passenger flow transmission edges are weighted based on the percentage of passenger flow from the first floor to the second floor according to historical data, with a higher percentage resulting in a higher weight.

5. The method according to claim 1, characterized in that, The method of encoding the temporal dimension change pattern of pedestrian flow through spatial correlation feature vectors to obtain temporal evolution feature vectors specifically includes: Obtain the time series data of the spatial correlation feature vector; The time series data is encoded using a recurrent neural network to obtain the temporal evolution feature vector of pedestrian flow on each floor.

6. The method according to claim 5, characterized in that, In the step of encoding the time series data of the spatial correlation feature vector based on the recurrent neural network, a fine-grained sliding time window is used to process the time series data, wherein the step size of the sliding time window is t, the historical time window is 2×t minutes, and the prediction time window is t minutes.

7. The method according to claim 1, characterized in that, The process of fusing visual perception data and architectural scene data through time-evolved feature vectors specifically includes: The quality of the real-time sensing data is evaluated, and the evaluation results are obtained; Based on the evaluation results, the weight ratio of the time evolution feature vector and the architectural scene data during fusion is dynamically adjusted. Specifically, when the quality of the perceived data is lower than a preset threshold, the weight ratio of the building scene data is increased. The building scene data includes: building operation schedule, meeting schedule and holiday information.

8. The method according to claim 1, characterized in that, Also includes: Based on the calculated number of people waiting for the elevator on each floor, the flow status of the prediction window is determined, including peak, off-peak, or low-peak conditions. Based on the calculated percentage of target floor distribution, the passenger flow direction trend for each floor is obtained, including upward or downward flow.

9. The method according to claim 8, characterized in that, Also includes: The scheduling strategy is executed according to the flow of people; wherein the scheduling strategy includes: in the peak period, the elevators are coupled and scheduled by floor; in the off-peak period, the elevators are coupled and scheduled in pairs; and in the low-peak period, some elevators are put into a dormant state.

10. A passenger flow prediction system for elevators based on spatiotemporal graph neural networks, characterized in that, The system includes a processor and a memory, the memory storing a computer program. When the processor executes the computer program to implement the method as described in any one of claims 1 to 9, it does so by constructing and running a spacetime graph neural network model, the model employing the following four-layer sequential architecture: The spatial graph coding layer is used to construct the graph nodes of each floor of the building and the passenger flow transmission relationship between floors in order to obtain spatial association feature vectors. The time series coding layer is used to encode the temporal dimension changes of people flow by spatially correlated feature vectors, so as to obtain time evolution feature vectors. The multi-source fusion layer is used to fuse visual perception data and architectural scene data through time-evolved feature vectors; The predictive output layer is used to predict the number of people waiting for the elevator on each floor every minute within a time window, the distribution ratio of the target floor on each floor, and the proportion of heavy objects carried on each floor.

Images