ST-GCN and Transform integrated elevator group competition scheduling method and simulation system

Abstract

The invention discloses an elevator group competition scheduling method fusing ST-GCN and Transform and a simulation system, and belongs to the field of intelligent building traffic control. According to the method, firstly, an elevator system is constructed into a space-time diagram, passenger flow space-time characteristics are extracted through ST-GCN, and then the complex dependency relationship between the overall state of an elevator and passenger requests is captured through a self-attention mechanism of Transform. On the basis, a competition scheduling mechanism is adopted, each elevator is used as an intelligent agent to carry out assessment competition on requests of passengers based on a space-time dependency relationship, a central scheduler preferentially allocates tasks, and effective combination of centralized learning and distributed decision making is realized. A matched simulation system can simulate various typical passenger flow scenes, and support is provided for algorithm training and verification. According to the method, time-space correlation and time sequence dependence can be accurately captured, fine scheduling is achieved in combination with hierarchical decision, the elevator scheduling efficiency is remarkably improved, and the dynamic requirements of actual application scenes such as office buildings are effectively met.

Description

Technical Field

[0001] This invention belongs to the field of intelligent building traffic control technology, and in particular relates to an elevator group state representation method based on ST-GCN spatiotemporal graph convolutional network and Transformer, which combines hierarchical decision-making and competitive architecture to achieve efficient scheduling, and is applicable to intelligent elevator control systems in complex multi-car scenarios. Background Technology

[0002] Elevator group control systems are a core component of modern high-rise buildings, and the quality of their scheduling algorithms directly determines the transportation efficiency of the vertical transportation system, passenger waiting / riding time, and building energy consumption. As buildings become taller and passenger flow patterns become more complex, traditional group control scheduling methods have revealed many limitations.

[0003] In existing elevator scheduling methods, rule-based approaches (such as minimum waiting time and first-come-first-served) statically set scheduling strategies, failing to perceive the overall passenger flow status and dynamic changes. This leads to inflexible performance under complex and variable passenger flows (such as morning and evening rush hours and lunchtime peaks), resulting in low overall system efficiency. While classic optimization algorithms (such as genetic algorithms and particle swarm optimization) can find near-optimal solutions, their computational time is long, making it difficult to meet the stringent real-time requirements of elevator scheduling (typically requiring decisions to be made within hundreds of milliseconds).

[0004] In recent years, relevant research attempts to apply deep learning to elevator dispatching. For example, a method learns and predicts elevator dispatching behavior through a Q-learning network, and models the elevator demand based on ST-GCN combined with a generative adversarial network. However, similar methods still have significant defects: (1) The spatial competition relationship cannot be effectively modeled. An elevator group is a typical spatio-temporal system, and multiple global elevators run between the same group of floors. They are spatially related to each other and compete for resources (floor calls). Existing methods still treat local elevators as independent individuals or only perform simple aggregation, and cannot explicitly model the complex dynamic competition and cooperation relationship between elevators; (2) The long-range temporal dependence capturing ability is weak: elevator dispatching decisions depend on long-term historical state sequences. For example, an elevator responded to a high-level call a few minutes ago, which will affect its ability to respond to new calls. Traditional RNN or LSTM models, Q-learning methods, and existing work based on ST-GCN combined with a generative adversarial network to predict elevator behavior all have the problem of gradient disappearance / explosion, making it difficult to effectively capture this long-range dependence, resulting in short-sighted decision-making; (3) The decision delay of the core dispatching process is high: the central controller needs to calculate the global optimal allocation in real time, and needs to re-traverse all elevator and call combinations when a new call is added, which is difficult to adapt to high-peak high-frequency call scenarios. Meanwhile, elevators passively execute allocation instructions and lack the ability to make autonomous judgments, and cannot adjust decisions based on their own operating states (such as current targets, loads, and directions); (4) The model generalization and simulation verification are disconnected: most research models are trained and tested in a fixed, simplified simulation environment, lacking a high-fidelity simulation platform that can truly reflect complex passenger flow and is tightly coupled with the algorithm. This leads to a sharp decline in performance when the algorithm is deployed in reality due to environmental differences. Meanwhile, simulation and dispatching decision-making processes are separated, and cannot provide closed-loop feedback for online learning and continuous optimization of the algorithm. SUMMARY

[0005] Problem solving: The present application aims to overcome the defects of the prior art, such as "separation of local features and global optimization, disconnection of short-term dispatching and long-term behavior", and provides an elevator dispatching method combining ST-GCN and Transformer. The local spatio-temporal features are extracted by ST-GCN, the long-term behavior model is modeled by combining the global attention mechanism of Transformer, and the response efficiency and global optimization ability of elevator dispatching are improved by combining the hierarchical decision-making and competition architecture, realizing the integrated representation of elevator group state and optimization of dispatching decision-making. At the same time, in order to simulate elevator behavior and verify the performance of the algorithm, the present application realizes an elevator dispatching system in a master-slave mode. The dispatching algorithm is deployed on the server side, and the basic behavior model of the elevator is implemented and deployed on the client side, which can simulate different elevator behaviors and verify different application scenarios. TECHNICAL SCHEME

[0006] To achieve the above objectives, the present invention provides the following technical solution: an elevator group competition scheduling method and simulation system that integrates STGCN and Transformer.

[0007] First, we introduce a hierarchical competitive scheduling scheme for elevator groups that integrates ST-GCN spatiotemporal graph convolution and Transformer, which mainly includes the following steps:

[0008] Step 101: Construct a dynamic spatiotemporal graph data model for the elevator group: Collect real-time elevator data, including elevator status data and call point (passenger call) status information, and then construct elevator nodes. Call point node The polygonal graph network relationships are then used to construct a dynamic spatiotemporal graph data model of the elevator group. It mainly includes a set of nodes, feature vectors, and an adjacency matrix;

[0009] Step 102, ST-GCN extracts local spatiotemporal features: the graph sequence generated in step 101 through a fixed period is... The node set, feature vector, and adjacency matrix are input into the ST-GCN network, respectively. After processing through the embedding layer and the spatiotemporal convolutional layer, the output is a local spatiotemporal feature tensor. ;

[0010] Step 103, Transformer models global long-term behavior: The spatiotemporal feature tensor generated in step 102 is used to model the global long-term behavior. The input is fed into a Transformer network, where it undergoes feature transformation to obtain tokens. These tokens are then fed into a multi-head self-attention layer to obtain global associations within the tokens. Finally, the output layer provides a global state representation of the elevator group. ;

[0011] Step 104, Hierarchical competitive scheduling decision based on fused features: Elevator intelligent agent The target floor is selected based on its own global state representation (first floor), and then competitive decision-making among multiple elevator agents is carried out when there is a new call (second floor).

[0012] Step 105, Construct a master-slave elevator group simulation system: Deploy the elevator group competition scheduling algorithm on the server side and deploy the elevator group simulation on the client side;

[0013] In step 101 above, the node information in the dynamic spatiotemporal graph is first defined. The node set includes elevator nodes and call point nodes, which can be described as: Secondly, the static and dynamic structures of the elevator nodes and the call point nodes need to be provided;

[0014] In step 101 above, the elevator node set (M is the number of elevators), and each elevator node feature vector is: ;

[0015] wherein the static features ( is the elevator number, is the maximum load, is the rated speed)

[0016] In the formula, the dynamic features , ( is the current floor at the moment, is the moving direction (down / stopped / up), is the real-time load, ) is the door state (closed / opened), is the cumulative running distance.

[0017] In the above step 101, the call point set is wherein K is the total number of call points, represents the th call point ( ); each call point includes static features and dynamic features , which are the core attributes of the space-time graph nodes; the static features ( is the floor number, is the call direction (down / up));

[0018] In the formula, the dynamic features ( is the call time, is the waiting time at t, is the allocation state (unallocated / allocated), is the associated elevator ID).

[0019] In the above step 101, the dynamic space-time graph includes elevator and call point nodes, and also includes a set of edges, which are the connection relationships between multiple nodes between elevator nodes and call point nodes;

[0020] In the above step 101, the definition of the edge set is as follows: the adjacency matrix quantifies the spatial relationship dynamically, and the element represents the edge weight between node and node :

[0021] The relationship of the elevator-elevator edge ( ) can be described as: ​

[0022] in, These are the weighting coefficients. As the attenuation factor, This is an indicator function (1 for the same direction, 0 otherwise).

[0023] Elevator - Call Point Side ( , The relationship between them can be described as follows:

[0024] in, These are the weighting coefficients. As the attenuation factor, This is the maximum waiting time threshold.

[0025] Call point - Call point edge ( The relationship between them can be described as follows:

[0026] in, Assign a base weight, and assign a higher weight when the elements are in the same or adjacent layers and in the same direction.

[0027] In step 102 above, a fixed time step can be used. (e.g., 2 seconds) Collect data and construct a structure containing A graph sequence at each time step Each time step , This is the feature vector of all nodes (elevator nodes, call point nodes) at this moment. Let be the adjacency matrix at that moment.

[0028] In step 102 above, the embedding layer uses a heterogeneous feature mapping function. Transform the features of elevator nodes and call point nodes to a unified dimension. (e.g., 16-dimensional):

[0029] In step 102 above, stacking is performed in the spatiotemporal convolution blocks. Each convolutional module contains spatial graph convolutional layers and temporal convolutional layers;

[0030] Among them, the spatial graph convolutional layer uses an adjacency matrix. Spatial correlation modeling of embedded features is as follows:

[0031] in Degree matrix ( ), For spatial convolution weights, This is the activation function.

[0032] Where the time convolution layer adopts 1D convolution to capture the temporal changes of adjacent time steps:

[0033] Where, is the time convolution weight.

[0034] In the above step 102, the output local spatio-temporal feature tensor ( is the local feature dimension) after stacking.

[0035] In the above step 103, the feature conversion process reshapes into a sequence form and flattens it into , each "node-time step" being a token.

[0036] In the spatio-temporal position encoding process, an encoding vector PE is added to each token:

[0037] Where the spatial encoding contains node type (elevator / call point) and physical location (floor) information, and the temporal encoding uses the above step 103, the sine and cosine functions are:

[0038] In the above step 103, the function of the self-attention layer is to calculate the global correlation between tokens through the multi-head self-attention mechanism: MultiHead( ;

[0039] Where, , h is the number of attention heads, , , is the weight matrix.

[0040] In the above step 103, the output layer needs to perform global pooling on the Transformer output to obtain the global state representation of the elevator group ( is the global feature dimension).

[0041] In the above step 104, the first layer action (target floor decision) in the hierarchical action space is the selection of the next target floor by each elevator agent based on its own features (extracted from ), the total number of floors ( ), and the strategy is ;

[0042] In the above step 104, the function of the second layer action (dynamic response decision) is to respond to new calls Combined with call characteristics (from Extracting and generating response strategies The strategy is ;

[0043] In step 104 above, during the competition coordination process, the central scheduler is based on the global representation. Calculate competition priority ,choose The maximum elevator response time is achieved by updating policy parameters using the Actor-Critic algorithm (local agents maximize cumulative gains, while the central Critic is based on...). (Assess the value of the condition).

[0044] In step 105 above, a simulation mechanism for elevator groups is provided. This mechanism is deployed on the client side of the simulation system, and its core includes a building simulation module, an elevator behavior simulation module, and a passenger flow generation module. The specific operation is as follows:

[0045] like Figure 2 As shown in step 201, the elevator simulation client needs to configure basic data such as the elevator environment.

[0046] like Figure 2 As shown in step 202, the building is simulated. Building elements are defined, including building type, number of floors, and number of elevators. Building simulation is closely related to elevator simulation. The number of floors and the height of each floor affect the elevator's travel distance.

[0047] like Figure 2 As shown in step 203, the elevator behavior is simulated. The elevator behavior simulation module needs to implement the elevator's operation process, including the elevator's current position, the upward and downward processes, and the opening and closing of the doors.

[0048] like Figure 2 As shown in step 204, passenger flow is simulated. Passenger flow models are constructed, such as the upward passenger flow pattern during the morning peak and the downward passenger flow pattern during the evening peak. Different passenger flow patterns can be used for algorithm verification and analysis in different scenarios.

[0049] like Figure 2 As shown in steps 205 and 210, both the elevator simulation client and the algorithm server need to establish a network connection channel.

[0050] The advantages of this invention, specifically regarding the elevator group control and dispatching algorithm based on the DQN machine learning mechanism, can be summarized as follows:

[0051] The application significantly enhances adaptability to dynamic traffic patterns. By constructing a space-time graph model, the elevator, floor and time are associated as a unified space-time graph convolution network, which accurately depicts the multi-dimensional dependence of traffic flow. The space-time graph convolution network can extract spatial proximity and temporal continuity features, so that the model can not only perceive the instantaneous influence of calls on surrounding elevators, but also learn long-term rules such as peak tidal flow, thereby overcoming the poor adaptability of traditional methods to complex scenes.

[0052] The global optimization and individual competition decision are innovatively integrated. The upper ST-GCN is responsible for learning the overall state of the system and achieving global collaborative optimization, and the lower Transformer allows each elevator to compete for task allocation based on its own state through the attention mechanism. This hierarchical architecture decomposes the macro efficiency target into micro agent competition behavior, improving the overall throughput of the system while ensuring the real-time and rationality of the dispatching decision, and achieving the balance between the overall and local interests.

[0053] The Transformer mechanism is used to improve the explainability of sequence modeling and decision-making. When modeling the local sequence of a single elevator, the Transformer can effectively capture the long-range dependencies between its historical running states, such as the relationship between consecutive stops and current load. Its multi-head self-attention mechanism not only calculates the task competition weight of each elevator, but also reveals the basis for decision-making, for example, the model can show that the weight of an elevator is reduced due to the mismatch of running direction, enhancing the transparency and credibility of the dispatching process.

[0054] The simulation system in the application provides a high-fidelity and efficient simulation environment for algorithm development. A master-slave architecture is used to separate centralized dispatching decision and distributed elevator running simulation. The master module focuses on running complex dispatching models, and the slave module simulates the accurate physical behavior of multiple elevators in parallel, so that the running details of large-scale elevator groups can be truly restored, and the simulation efficiency is guaranteed, providing an indispensable reliable platform for algorithm verification.

[0055] The application combines STGCN and Transformer deep learning networks for time series behavior learning to obtain the optimal dispatching scheme. Compared with traditional first-in-first-out dispatching algorithms and traditional deep learning methods, the proposed algorithm performs best in terms of average waiting time, average boarding time and other core service indicators. The average waiting time is improved by 26.6%, and the empty load rate is improved by 12.5%, with obvious improvement in technical indicators. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the examples of the present application or the prior art, the following will briefly introduce the drawings required to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0057] Figure 1 For the present application, the flow chart of the elevator group competition scheduling method fusing STGCN and Transformer

[0058] Figure 2 For the present application, the algorithm server module diagram and the elevator simulation client module diagram

[0059] Figure 3 For the present application, the elevator state-behavior transition state diagram in the elevator simulation module

[0060] Figure 4 For the present application, the elevator execution behavior activity diagram in the elevator simulation module

[0061] Figure 5 For the present application, the performance comparison diagram of different dispatching algorithms

[0062] Figure 6 For the present application, the algorithm performance comparison diagram under different modes Specific implementation

[0063] In order to enable the persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.

[0064] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can cover the order implementation other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to which steps or units are clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0065] Embodiment 1: A local spatio-temporal feature extraction module for an elevator group control scheduling model

[0066] This embodiment takes a 30-story commercial office building (configured with 10 elevators and 31 call points) as a scenario to illustrate the construction process of dynamic spatio-temporal graph in detail, providing input data support for the ST-GCN local spatio-temporal feature extraction module.

[0067] Step one: Node set definition

[0068] Node set , wherein:

[0069] Elevator node set ( 10 elevators), and each node feature vector is :

[0070] Static features : (elevator number, such as E1 corresponding to ID = 1); people (uniform rated capacity); (uniform rated speed).

[0071] Dynamic features : (the floor at time t, such as E1 at t = 0 on the 5th floor, ); (-1 = down, 0 = stop, 1 = up, such as E1 up ); (current load, such as E1 currently carrying 6 people, ); (0 = closed, 1 = open, such as E1 stopping at ); (cumulative running distance, unit m, such as E1 from 1st floor to 5th floor ).

[0072] Call point node set (K = 31 call points, 1st floor lobby + 1 each from 2nd to 30th floor), and each node feature vector is :

[0073] Static features : (call point floor, such as C5 corresponding to ); (-1 = down call, 1 = up call, 1st floor only ).

[0074] Dynamic features :

[0075] (call occurrence time, e.g. C5 called at t=10s, ); (waiting time at t, e.g. t=15s );

[0076] (elevator assigned, 1 = assigned, e.g. C5 assigned to E1 when );

[0077] (associated elevator ID, e.g. C5 associated with E1 ).

[0078] Step two: Edge set definition

[0079] Edge set defined by adjacency matrix Dynamic quantification, element defined as follows:

[0080] Elevator-elevator edge ( ):

[0081]

[0082] weight coefficient , decay factor ;

[0083] Example: Edge weight of E1 (pos=5, dir=1) and E2 (pos=7, dir=1):

[0084]

[0085] Elevator-call point edge ( ):

[0086]

[0087] weight coefficient , decay factor , maximum waiting threshold ;

[0088] Example: Edge weight of E1 (pos=5) and C7 :

[0089]

[0090] Call point-call point edge ( ):

[0091]

[0092] base weight ;

[0093] Example: C5 and C6 Edge weights:

[0094]

[0095] Step three: graph sequence construction

[0096] Fixed time step Collect data, construct graph sequence of steps , where:

[0097] Each time step :

[0098] Real-time features (static features remain unchanged, dynamic features update with t) of 10 elevators and 31 call points;

[0099] Dynamic update with elevator position and call state (e.g. change of floor difference caused by elevator movement, real-time adjustment of edge weight).

[0100] The final input tensor dimension is:

[0101] Elevator node features (10 time steps 10 elevators 9-dimensional features);

[0102] Call point features (10 time steps 31 elevators 9-dimensional features);

[0103] Adjacency matrix sequence (10 time steps 41 nodes 41 nodes);

[0104] This embodiment defines the node static / dynamic features and dynamic association rules of the edge in detail, constructs a space-time graph sequence that fits the elevator group control scene, provides a structured input for the subsequent ST-GCN module to extract local space-time features, and ensures the accuracy and real-time performance of feature extraction.

[0105] Embodiment 2: ST-GCN local space-time feature extraction module

[0106] This embodiment takes the results of dynamic space-time graph construction, and constructs a 30-story office building scene (10 elevators, 31 call points, The specific processing procedures of the embedding layer and the spatio-temporal convolution block are described in detail, and finally the local spatio-temporal feature tensor is output.

[0107] Step one: uniform mapping of heterogeneous features in the embedding layer

[0108] Through the heterogeneous feature mapping function Convert the original features of the elevator nodes and the call point nodes to the uniform embedding dimension The specific implementation is as follows:

[0109] Definition of the mapping function: for the elevator node features Respectively mapped by a dedicated fully connected layer,

[0110]

[0111] wherein, is the elevator node feature mapping layer (input 9 dimensions, output 16 dimensions), is the call point node feature mapping layer (input 9 layers, output 16 dimensions), is an activation function.

[0112] Output result: feature matrix after embedding , containing the uniform dimension features of all nodes at the moment, eliminating the differences between the heterogeneous features of the elevator and the call point.

[0113] Step two: spatio-temporal correlation modeling of the spatio-temporal convolution block

[0114] Stack 2 spatio-temporal convolution modules (each containing a spatial graph convolution layer and a temporal convolution layer) to gradually extract local spatio-temporal features:

[0115] The first convolution module includes a spatial graph convolution layer and a temporal convolution layer.

[0116] Spatial graph convolution layer: based on the adjacency matrix Model the spatial correlation of nodes, and the formula is as follows:

[0117]

[0118] wherein, the degree matrix , (the degree of the node );

[0119] The spatial convolution weight (output dimension 32);

[0120] Example: at t=5s, the embedding features of are converted to the spatial correlation features of through this layer, strengthening the correlation between the elevator and the adjacent call point.

[0121] Temporal convolution layer: 1D convolution kernel (kernel_size=3) is used to capture the time sequence dependence of the next 3 time steps:

[0122]

[0123] where the temporal convolution weight (maintain dimension 32);

[0124] Output , fusion of spatial features at t-1, t, t+1 time, capture the time sequence correlation of elevator running trend and call change.

[0125] The second convolution module

[0126] Spatial graph convolution layer: the input is the time convolution output of the first module, and the formula is as follows:

[0127]

[0128] where the spatial convolution weight (output dimension 64).

[0129] Temporal convolution layer: also use 1D convolution with kernel_size=3, formula as follows:

[0130]

[0131] where the temporal convolution weight (maintain dimension 64), output .

[0132] Step three: local spatio-temporal feature output

[0133] After 2 spatio-temporal convolution blocks, integrate the features of T=10 time steps, output local spatio-temporal feature tensor

[0134] where the dimension meaning is: 10 time steps 64-dimensional local features. This tensor integrates the static attributes, dynamic state, spatial correlation and time sequence change rule of each node, providing accurate feature support for subsequent hierarchical decision module (such as local observation of elevator i).

[0135] This embodiment realizes the uniformity of heterogeneous features through embedding layer, and combines spatio-temporal convolution block to capture spatial correlation and time sequence dependence step by step, ensuring that the extracted local spatio-temporal features can accurately reflect the dynamic running state of the elevator group control system, laying a foundation for improving the accuracy of scheduling decision.

[0136] Embodiment 3 Global feature fusion module based on Transformer

[0137] This embodiment is based on the local spatio-temporal feature tensor output by the aforementioned ST-GCN Take a 30-story office building scenario (10 elevators + 31 call points, N = 41, T = 10, ) as an example to illustrate the complete process of feature conversion, spatio-temporal position encoding, self-attention calculation and global state output in detail.

[0138] Step 1: Feature conversion - reconstruction in sequence form

[0139] The local spatio-temporal feature tensor output by the ST-GCN (time step node number local feature dimension) is adjusted to a sequence form, as follows:

[0140] Dimension arrangement: exchange the time step and node number dimensions to obtain sequence features , where

[0141] The first dimension is the node (10 elevators + 31 call points);

[0142] The second dimension is the time step (every 2 seconds, a total of 20 seconds);

[0143] The third dimension is the local feature dimension .

[0144] Combine the "node-time step" into an independent , which is expanded to , where:

[0145] The total number is (every corresponds to the feature of "the nth node at the t time step");

[0146] Feature dimension (maintain the same as the local feature dimension).

[0147] Step 2, spatio-temporal position encoding - position information embedding Add a spatio-temporal position encoding vector

[0148] to each to strengthen the prior information of the node physical position and time step, as follows:

[0149]

[0150] ​where the semicolon denotes vector concatenation, each part has a dimension of 32, and the total dimension is 64. Match.

[0151] 1. Spatial encoding : contains node type and physical location information,

[0152] If node n is an elevator ( ): where is the elevator number, is the floor at time t, mapped to a 32-dimensional vector through an embedding layer;

[0153] If the node is a call point ( ): where, is the call point floor, is the call direction, also mapped to a 32-dimensional vector.

[0154] 2. Time encoding : generated using a cosine function to capture the periodic changes in time steps:

[0155]

[0156] where, (16 pairs of sine and cosine functions due to a time encoding dimension of 32), t is the time step index (0 ).

[0157] Example: when t = 5, the 0th dimension (2k = 0): ; the 1st dimension (2k + 1 = 1):

[0158]

[0159] Encoded sequence: the sequence input to the Transformer is , which integrates local features and spatio-temporal location information.

[0160] Step three: self-attention layer-global correlation modeling

[0161] Use 2-layer Transformer encoding layer, each layer contains multi-head self-attention mechanism, to capture the global correlation between 410 tokens:

[0162] 1. Multi-head self-attention definition:

[0163] Let the number of attention heads be , and the feature dimension of each head be , the calculation is as follows:

[0164]

[0165] where, is the output projection matrix, and the dimension after multi-head concatenation is kept as 64.

[0166] 2. Single-head attention calculation:

[0167] Each head calculates the attention weight by querying (Q), key (K), and value (V) matrices:

[0168] , ,

[0169]

[0170] where, , , is the weight matrix of the i-th head, which captures the dependency between tokens by scaled dot-product attention (e.g., the association between the "5th floor calling point" and the "3rd floor elevator" at t = 3).

[0171] Step 4: Output layer-global state representation

[0172] After 2-layer Transformer encoding, the output sequence is , and the global state of the elevator group is extracted by global pooling:

[0173] 1. Pooling method

[0174] The features of all tokens are averaged, and the formula is as follows:

[0175] where, (global feature dimension)

[0176] 2. Physical meaning

[0177] Global state integrates the local spatiotemporal features and global associations of 10 elevators and 31 calling points within 20 seconds, which can be directly used for macro decision-making of elevator group control (e.g., global allocation strategy of elevator resources during peak hours).

[0178] This embodiment converts local features into Token sequences through sequence remodeling, strengthens prior information by combining spatiotemporal position encoding, and captures global associations using multi-head self-attention. The final output of the global state representation provides global perspective feature support for the overall decision of the elevator group control system, which makes up for the limitations of local features only focusing on individual associations.

[0179] Embodiment 4: Hierarchical decision and competition coordination module

[0180] ​This embodiment takes a 30-story office building scenario (10 elevators, 31 call points, total number of floors ) as an example, based on the aforementioned global state representation , the specific implementation process of hierarchical decision-making (target floor decision + dynamic response decision) and competition coordination mechanism is described in detail.

[0181] Step one: the first layer of action-target floor decision (macro-scheduling)

[0182] Each elevator agent independently selects the next target floor based on its own characteristics to achieve global routing planning:

[0183] 1. Elevator individual feature extraction:

[0184] The local features of elevator i are extracted from the global state , including the real-time position, load, associated call state and global passenger flow trend of elevator i (such as the features of E1, the features of E2).

[0185] 2. Definition of strategy function:

[0186] The target floor strategy of elevator i is realized through a multi-layer perception (MLP):

[0187] where, is a 3-layer MLP (input 64-dimensional, output 30-dimensional, corresponding to 30 floors) exclusive to elevator i, and the output layer is converted into a discrete probability distribution (P) by softmax is the target floor.

[0188] 3. Decision execution:

[0189] Example: during the morning rush hour, E1 is currently located on the 1st floor (P ), , shows that calls are concentrated on floors 20-30, the probability of outputting the 25th floor is the highest (0.32), and E1 selects the 25th floor as the next target floor.

[0190] Step two: the second layer of action-dynamic response decision (micro-control)

[0191] When a new call appears during the execution of the first layer of action, whether to respond is decided based on the call characteristics 1. Call feature extraction:

[0192]

[0193] ​​​When global state extracts new call features , including call floor (e.g. 8th floor), direction (upward), waiting time (3 seconds), and surrounding elevator distribution.

[0194] 2. Strategy function definition:

[0195] Elevator i's response strategy to new call is a binary decision:

[0196] where, is the feature concatenation ( concatenated with elevator i's is 128-dimensional),, is a 2-layer MLP (output 1-dimensional ),, is a sigmoid function (output probability value, indicates response ,, indicates ignore ).

[0197] 3. Trigger and execution:

[0198] Trigger condition: E1 is going from 1st floor to 25th floor (upward), a new upward call appears at 8th floor ( ) on the way, and 8th floor is located between 1st and 25th floors.

[0199] Decision example: shows that there are many people waiting at 8th floor, shows that E1's current load is low (5 / 13 people ),, outputs a response probability of 0.85, and E1 decides to respond to the call.

[0200] Step three: competition coordination mechanism - central dispatcher arbitration

[0201] When multiple elevators have response conditions for the same call, the central dispatcher calculates the priority based on the global state and selects the optimal elevator to respond:

[0202] 1. Priority calculation:

[0203] For new call , all elevators that meet the response conditions (e.g. E1, E3, E5) enter the competition pool, and the central dispatcher calculates the priority of each elevator through MLP: ;

[0204] where, is the central coordination network (input 128-dimensional, output 1-dimensional priority score), which considers the distance between the elevator and the call point, load, and running direction matching degree (e.g. E1 is closer to 8th floor and has low load, ).

[0205] 2. Strategy update (Actor-Critic):

[0206] Local agent (elevator): maximize cumulative rewards (e.g. call response timeliness, passenger waiting time reduction) through Actor network, update and parameters;

[0207] Central Critic: evaluate the state value of the current decision based on the global state , guide the Actor network optimization through the time series score error (TD-error), and ensure the global optimal income (e.g. avoid multiple elevators running empty, balance the load).

[0208] 3. Execution results:

[0209] The central dispatcher selects E1 with the highest priority to respond to the call from the 8th floor, while notifying E3 and E5 to ignore the call and continue executing the original target floor plan.

[0210] This embodiment realizes fine scheduling through hierarchical decision-making, solves multi-elevator resource conflicts through a competition and coordination mechanism, retains the decision-making flexibility of individual elevators, and ensures the overall optimality of the system through global state and central arbitration. In the morning peak scenario, the average waiting time can be reduced by 38%, and the elevator empty load rate can be reduced by 25%, which is significantly better than the traditional centralized scheduling scheme.

[0211] Embodiment 5 Master-slave elevator group simulation system

[0212] The embodiment discloses a master-slave elevator group simulation system, which is suitable for a 30-story commercial office building scenario (1st floor is a lobby, 2nd-30th floors are office areas), supports blind layer constraints, multiple passenger flow modes (including evening peak double peak distribution), and simulation verification of ST-GCN+Transformer hierarchical decision-making algorithm, realizes collaborative simulation of "global scheduling + local execution" through master-slave architecture, and ensures that the performance of the algorithm under real scene constraints is reproducible and quantifiable.

[0213] I. Master-slave elevator simulation system - system model

[0214] As shown in Figure 2 , the algorithm server function module of the elevator group scheduling and simulation system based on multiple agents includes:

[0215] ​System state setting module: as steps 201 and 206, the elevator group scheduling and simulation system includes an algorithm server end and an elevator behavior simulation client end. Both ends need to set the running environment of the system elevator. The algorithm server end needs to select the elevator through the stored building and elevator state information. The elevator simulation client end needs to simulate the behavior through the stored building, elevator state and other information, and generate passenger information to request the elevator. This is the basic running environment of the elevator group scheduling method and simulation system based on multi-agent of the embodiment of the application. Specifically, the client end and the server end of the elevator group scheduling and simulation system can input the relevant system parameters through the interface. The building information includes: number of floors, number of elevators, floor height, blind layer information. The elevator parameters include: elevator number, maximum speed, door opening time, door closing time, maximum load (number of people), etc. After setting the data, the elevator group scheduling and simulation system can be run.

[0216] System state sensing module: in step 207, the scheduling algorithm server and the elevator client read the state information in a fixed time period on the basis of establishing a network communication connection.

[0217] In this embodiment, the typical read data items are: sender (identified by the ID number of the elevator), receiver (number of the server), current floor of the elevator, current floor of the elevator, current running direction of the elevator (0: stop, 1: up, -1: down), door control signal currently read by the elevator (including door opening completion and door closing completion), state of the elevator (the state describes whether the elevator is abnormal, overloaded, full, abnormal, etc.). The data is packaged by the elevator simulation client through real-time data collection and transmitted to the algorithm server end.

[0218] Data preprocessing and space-time state construction module: in step 208, after the algorithm server reads the state data, it completes data cleaning and preprocessing, and after normalization and coding operation, it constructs a space-time graph structure based on the elevator node, floor node and the connection relationship between the two. At the same time, the historical 10-step time sequence state data is extracted, the preprocessed features are fused with the space-time graph and the time sequence data, and the input data required by the ST-GCN+Transformer model (time-space sequence input containing multi-dimensional features) is formed, which provides accurate state support for hierarchical decision-making.

[0219] The agent decision module (ST-GCN+Transformer hierarchical decision engine) is a core module of the application and realizes fine scheduling by incorporating the hierarchical decision idea. The module loads the trained ST-GCN+Transformer model, receives the above-mentioned spatiotemporal sequence input data, first captures the spatiotemporal correlation features of elevator scheduling through the ST-GCN layer, and then accurately models the temporal dependence relationship through the 2-layer Transformer encoding layer (4 attention heads, 128-dimensional hidden layer). Based on the fusion feature, the hierarchical action priority score is output: the first layer is the "core target floor" decision score, which selects the optimal target floor; the second layer is the "in-flight decision" score, which judges whether the new call at the passing floor is on the way. The final optimal action is determined by combining the scores of the two layers , which is converted into specific elevator control signals (including target floor instructions and in-flight response strategies) and issued to the corresponding elevator controller for execution.

[0220] The data transmission module: as shown in steps 210 and 205, communication is needed between the client and the server to transmit passenger request instructions and state information reports. The specific data types transmitted are as follows:

[0221] (1) Registration information reporting: when the algorithm server starts, the elevator simulation end starts running and first sends registration information to the server for registration. The information sent includes the sender, receiver, device ID, device type, etc. After receiving the information, the algorithm server registers and feeds back.

[0222] (2) Heartbeat information reporting: after the elevator simulation client registers, the client will periodically report heartbeat signals to the algorithm server. The information includes: message sender, message receiver, timestamp, etc. After receiving, the algorithm server feeds back. If no heartbeat data is received within the period of time, the state of the elevator will be changed to offline.

[0223] (3) Passenger request reporting: when the elevator simulation client has a passenger making a dispatch request, the elevator client will send response data information, including departure floor, destination floor, departure door (before or after), arrival door (before or after), etc. After receiving, the algorithm server will dispatch the application according to the aforementioned DQN algorithm, obtain the assigned elevator ID number, and feed back the data to the elevator simulation client.

[0224] (4) Elevator state data reporting: the elevator simulation client periodically reports its state information, including elevator number, running direction, current floor, etc. to the algorithm server. The algorithm server modifies the data and updates the UI, allowing real-time viewing of the elevator state.

[0225] For example Figure 2As shown, the elevator simulation client functional module in a multi-agent-based elevator group scheduling and simulation system includes:

[0226] Building Module: The building module is the underlying core foundation of the elevator simulation system, responsible for modeling the vertical transportation environment of a real building and providing a virtual scene for elevator operation and passenger flow. Its core function requires defining the physical layout and static parameters of the elevator group system, such as the number of floors, functional zones, and floor heights. For a total number of floors... Buildings, using collections Let represent the set of all floors of the building, which can be formally described as relation 1, as follows:

[0227] Relation (1)

[0228] in Indicates the first Floors, such as the first floor, are represented as The second floor is represented as And so on, the first Layer is represented as .

[0229] In real life, elevators may also have blind floors. Blind floors are floors in a building that cannot be reached by the elevator due to early construction or planning reasons. Each elevator may have different blind floors, and the relationship between blind floors and elevators is one-to-one or one-to-many. Therefore, a matrix of 1s and 0s can be used to represent the correspondence between blind floors and elevator groups, as shown in Equation 2.

[0230] Relation (2)

[0231] in, This indicates the total number of floors in the building. This indicates the total number of elevators in the building. This indicates whether the i-th floor is a blind floor for the j-th floor elevator. A value of 1 indicates a blind floor, and a value of 0 indicates that it is not a blind floor. The value of can be expressed by the following relation (3):

[0232] Relation (3)

[0233] Several parameters of a building are defined as follows: (A collection of all floors) (Matrix of the blind layer) (Total number of floors in the building, range:) (Unit: layer) (The i-th floor, range:) (Unit: layer) (1 (1 (1 (1 (1 (1 (1 (1 (1

[0234]

[0235] Blind floor rule: (1 (1 (1

[0236] Elevator module: The relevant attribute definitions of the elevator equipment also need to be given. Specifically, they include: (1 (1 (1 (1 (1 (1 (1 (1 (1 (1 (1 (1 (1 (1 (1 (1 (1 (1 (1 (1

[0237] Passenger flow generation module: This module is responsible for simulating the generation of passenger elevator demand in real scenarios, providing passenger data for testing and verification of the simulation system. Individual passenger generation will generate virtual passengers with independent attributes according to the set flow characteristics, including calling elevator time, calling floor, calling direction (up / down), target floor, etc.

[0238] Preferably, the flow characteristics are divided into early peak, late peak, inter-floor mode, and mixed mode.

[0239] 1. Early peak mode (e.g. 7:30-9:00)

[0240] Passengers mainly start from the bottom floor and go to the middle and high floors, with few downward requests. The bottom floor calls are dense, and the elevator needs to send passengers from the bottom floor to the upper floors.

[0241] 1. Departure floor edge distribution , described as:

[0242]

[0243] Normalization processing: ensure , the proportion of departure from the bottom floor (s=1) , which shows the characteristics of departure from the first floor.

[0244] 2. Target floor conditional probability , described as:

[0245]

[0246] where, (targeted at middle and high floors, such as N=30 =24.2), =5 (dispersed system);

[0247] It means that passengers starting from the bottom floor mainly go to the 20-30th floor, and a small number go to the 10-19th floor.

[0248] Flow intensity: number of newly generated calls per unit time (minute) calls / minute;

[0249] 2. Late peak mode (e.g. 17:30-19:00)

[0250] Passengers mainly start from the middle and high floors and go to the bottom floor, with few upward requests. The middle and high floor calls are scattered, and the elevator needs to load passengers from the high floors to the lower floors.

[0251] 1. Departure floor edge distribution ( =0.8, after normalization )

[0252] Constraints: (less from bottom), (mainly from middle and high)

[0253] Target floor conditional probability , described as:

[0254]

[0255] Where: (the higher the departure floor, the stronger the bottom guidance, such as s=30, )

[0256] New call number per unit time: times / minute;

[0257] 3. Inter-floor mode (e.g. 10:00-12:00, 14:00-16:00)

[0258] No obvious one-way passenger flow, passengers move randomly between floors (e.g. meetings, pickups), up / down requests are balanced, and call distribution is scattered.

[0259] 1. Departure floor distribution , described as:

[0260] (uniform distribution, all floors have equal probability of departure)

[0261] Constraints: (N=30 is about 0.033), no concentrated departure floor.

[0262] Target floor conditional probability , described as:

[0263]

[0264] New call number per unit time can be set to: times / minute;

[0265] 4. Mixed mode (e.g. 8:30-10:00, 16:00-17:30)

[0266] Early peak to inter-floor mode transition, or inter-floor mode to late peak transition, with two-way passenger flow but one direction dominant.

[0267] Set the early peak / late peak weight as , the inter-floor mode weight as , and the weight of different modes as:

[0268] ① Early peak mode - inter-floor mode transition: ​ =8:30, (linear decay);

[0269] ② Inter-floor pattern—Evening peak transition: ( =16:00, (linear growth);

[0270] The traffic generation model uses a weighted combination of joint probabilities: ;

[0271] Number of new calls generated per unit time: The frequency per minute gradually decreases or increases with the transition;

[0272] Elevator Execution Behavior Logic: The core execution process of an elevator is crucial for verifying optimal scheduling. Finite state automata can be used to describe elevator behavior. For example... Figure 3 As shown, the core control algorithm needs to consider both internal and external call signals. After initialization, external call signals are processed first. When a passenger enters the elevator, their internal call signal (arrival floor information) is added to the up / down request list, the target floor is selected, and the up / down movement continues. The elevator has seven state behaviors: Unregistered, Offline, Idle, Stopped, Moving, Open Door, Closed Door, Passenger Entering (LAND_IN), and Passenger Descending (LAND_OFF). The elevator's state behaviors and their switching logic are defined as follows:

[0273] ① Unregistered: After initialization, the elevator simulation client and algorithm server enter an unregistered state. The client sends registration information to the server, and after the server confirms, both parties switch to offline status.

[0274] ② Offline: The elevator periodically sends heartbeat data (every 5-10 minutes). After successfully receiving the heartbeat data, the elevator status is updated to idle.

[0275] ③ Idle: The elevator has no internal or external call signals, is stationary at a floor, and its direction of travel is 0. It switches to moving mode when a new call signal is received, and switches to stopped mode when performing maintenance or other operations.

[0276] ④ Move: The elevator runs in the preset direction in response to a call signal. Upon reaching the target floor, it decelerates and switches to a stop; in case of emergency, it immediately switches to a stop and maintains its original direction.

[0277] ⑤ STOP: The elevator stops at a floor, there is an unprocessed signal, and the original running direction is maintained. Switch to the open door state in response to the signal of the current floor; switch to the moving state in response to other floor signals; and switch to the idle state after all signals are processed.

[0278] ⑥ OPEN_DOOR: The elevator door starts to open to full opening. After the door is fully opened, it is automatically switched to the upper passenger state; if it is triggered by mistake, it is directly switched to the closed door state.

[0279] ⑦ LAND_IN: The process of passengers entering the car. After completion, it is switched to the closed door state; new external call signals during the process do not affect the current state.

[0280] ⑧ LAND_OFF: The process of passengers leaving the car. After completion, it is switched to the closed door state; if passengers need to be loaded at the same time, it is switched to the upper passenger state.

[0281] ⑨ CLOSE_DOOR: The elevator door starts to close to full closing. After the door is closed, it is switched to the moving state if there is an unprocessed signal, and it is switched to the idle state if there is no signal; and it is switched to the open door state if an obstacle is encountered.

[0282] Simulation process of algorithm server and elevator client

[0283] As shown in Figure 4 , the interaction process between the dispatching algorithm server and the elevator client is as follows:

[0284] In step 401, first, the elevator is initialized, and the logical time of the simulation is updated synchronously (the logical time of the simulation is set by a timer, starting from 00:00:00), and the logical time of the simulation is associated with the time of the generated passengers;

[0285] In step 402, after the relevant data is initialized, the elevator simulation client sets the state of the elevator to offline, and sends registration information to the algorithm server. If the algorithm server feedbacks confirmation information, the state of the elevator is modified to running. Otherwise, the elevator will always be in an offline state. After the elevator sends the registration information and succeeds, the elevator will be in an idle state and wait for dispatching.

[0286] In step 403, the logical time of the simulation client is constantly updated at this time, and the passenger with the earliest request time is selected from the generated passenger information, and it is judged whether the request time is the same as the current system logical time. If the times are equal, it indicates that a virtual passenger call request is generated at this time. And send the information to the dispatching algorithm server through the "passenger request reporting" data packet. After receiving it, the algorithm server parses the request and performs scheduling calculation according to the same building and elevator state information as the client to obtain the optimal elevator, and feeds back the result to the elevator simulation client.

[0287] In step 404, at this moment, if the elevator simulation client has been fed back the dispatched elevator result (elevator number), the elevator simulation client first checks whether the requested floor is the current floor. If it is the current floor, the door opening operation is performed, otherwise, if the requested floor is not the current floor, it needs to go up / down to the request floor.

[0288] In steps 405-408, if the requested floor is different from the current floor of the elevator, the elevator needs to run to the request floor, which needs to go through the acceleration-constant speed-deceleration process. If the floor distance is short, it may go through the acceleration-deceleration process, which is related to the acceleration and deceleration of the elevator and the floor height of the elevator.

[0289] In steps 409-410, when the elevator arrives at the passenger request floor, it needs to strictly go through the door opening-passenger boarding and alighting-door closing process. This part needs to consider the number of passengers boarding and alighting and the load limit. At the same time, when initializing the elevator attributes, the opening time (open_door_time) and closing time (close_door_time) of the elevator need to be set, and the time required for passengers to board and alight also needs to be set. In order to simulate the behavior of the elevator as realistically as possible.

[0290] The parameter configuration table of the elevator simulation system is as follows:

[0291]

[0292] Embodiment effect

[0293] After applying the method described in the embodiment, the elevator group control system shows the following beneficial effects:

[0294] The algorithm accurately captures the dynamic spatiotemporal dependence of traffic flow through spatiotemporal graph convolution, significantly improving the adaptability to complex patterns such as peak tides.

[0295] The hierarchical competition mechanism integrates the global optimization of ST-GCN and the individual competition decision of Transformer, balancing system efficiency and real-time response capability.

[0296] The dependence relationship of elevators and call points is plotted by space-time graph, combined with the multi-head attention mechanism of Transformer, precise sequence modeling is realized, passengers can better understand the reasons for dispatching elevators, and the transparency and explainability of the dispatching decision process are enhanced.

[0297] The test verification process takes a 30-story commercial office building (10 elevators) as a scenario, covering four typical passenger flow scenarios: morning peak, evening peak, inter-floor mode, and mixed mode. The passenger arrival interval follows an exponential distribution, and the destination selection matches the corresponding mode probability distribution. The test compares six algorithms: random scheduling (baseline), FCFS, zoned scheduling, traditional Q-learning, ST-GCN+GAN, and ST-GCN+Transformer. The parameters of ST-GCN+Transformer are set as follows: 50-dimensional input features, space-time graph containing elevator and floor nodes, 2 space-time convolution blocks, Transformer containing 2 encoding layers and 4 attention heads (hidden layer 128 dimensions), experience pool 10 4 , batch size 64, target network updated every 100 steps, learning rate 0.001, sequence input 10 steps. The mode test shows that, as shown in the accompanying Figure 5 , random scheduling performs the worst (overall time 89.4 seconds in the morning peak), FCFS and zoned scheduling are limited by the lack of global optimization and poor mode switching flexibility, traditional Q-learning is limited by the state space, ST-GCN+GAN lags behind in dynamic scene timing capture, and ST-GCN+Transformer is optimal in each mode, with an overall time of 44.8 seconds in the morning peak (12.8% lower than ST-GCN+GAN) and 41.7 seconds in the mixed mode (49.9% of random scheduling). As shown in the accompanying Figure 6 , overall, ST-GCN+Transformer has a passenger waiting time of 18.2 seconds and an overall time of 41.4 seconds, with a significant reduction compared to other algorithms, and an elevator empty rate of 12.5% and an energy consumption coefficient of 0.72, both of which are the lowest, achieving triple optimization of passenger waiting experience, overall travel experience, and operating cost.

[0298] This embodiment realizes the decoupling of global scheduling and local simulation through master-slave architecture, ensuring the global optimization capability of hierarchical decision algorithms, and restoring the real running scenario of elevator group control through detailed physical simulation and blind layer constraints, providing a reliable simulation platform for practicality verification of algorithms.

Claims

1. An elevator group hierarchical competition scheduling method fusing a ST-GCN spatio-temporal graph convolution and a Transformer, characterized in that, Comprising the following steps: (1) Constructing a dynamic space-time graph data model of the elevator group: defining a node set and an edge set, the node set including elevator nodes and call point nodes, the elevator node features including static features (ID, maximum load, door opening and closing time, maximum speed) and dynamic features (current floor, running direction, real-time load, door state, cumulative running distance), the call point node features including static features (floor number, call direction) and dynamic features (call time, waiting time, allocation state, associated elevator ID); the edge set dynamically quantifies spatial correlation through an adjacency matrix, including elevator-elevator edges (weight inversely proportional to floor difference, proportional to same-direction running coefficient), elevator-call point edges (weight inversely proportional to distance, proportional to waiting time), and call point-call point edges (same floor / adjacent floor same direction assigned high weight); (2) Spatio-temporal graph sequence generation: collect data at fixed time steps, form a graph sequence containing time steps , each containing a node feature matrix and an adjacency matrix; (3) Extract local spatio-temporal features by ST-GCN graph convolution: map the heterogeneous features of elevator nodes and call point nodes to a unified dimension through the embedding layer, and then output the local spatio-temporal feature tensor through the spatio-temporal convolution block composed of 2-3 "spatial graph convolution layer + time convolution layer" (shaped as , is the total number of nodes, is the feature dimension); (4) Modeling global long-term behavior using Transformer: Reshape into Transformer input format For each "node-time step" token, a spatiotemporal location code is added (spatial code includes node type + physical location, time code is sine and cosine timestamps). After calculating the correlation between tokens through a self-attention layer, a global fusion feature is output. Pooling yields the global state representation of the elevator group. ; (5) Hierarchical competition scheduling decision based on fusion features: the global state is characterized Input decision layer, output hierarchical action, the first layer is the next target floor selection of the car, and the second layer is the dynamic response decision of the new call on the way. Through the coordination of multi-agent competition architecture, the local agent generates action based on the central dispatcher generates action based on evaluate system value and allocate calls; (6) Constructing a master-slave type elevator group simulation system: deploying a hierarchical scheduling algorithm that integrates graph convolution and Transformer on the server side, and deploying an electrical simulator on the client side for simulation verification.

2. The method of claim 1, wherein, The weight calculation formula of the elevator-elevator side in step (1) is: wherein is a weight coefficient, is a decay factor, is an indicator function; and the weight calculation formula of the elevator-call point side is wherein is a weight coefficient, is a decay factor, is a maximum waiting time threshold value; The relationship of call point - call point edge (CPE) can be described as: wherein, wherein, is the base weight.

3. The method of claim 1, wherein, The space graph convolution layer in step (2) passes through an adjacency matrix The space correlation modeling of the embedded features is The time convolution layer adopts to capture the temporal changes of adjacent time steps.

4. The method of claim 1, wherein, The embedding layer in step (3) is mapped by a heterogeneous feature mapping function The feature dimension is unified, and the spatial graph convolution layer is based on an adjacency matrix The spatial correlation feature is calculated, and the formula is: , wherein is a degree matrix, is a spatial convolution weight, is an activation function; the time convolution layer is mapped by The time sequence change is captured by convolution, and the formula is: , wherein is a time convolution weight.

5. The method of claim 1, wherein, The time coding of the space-time position coding in step (4) satisfies: The self-attention layer adopts a multi-head self-attention mechanism, and the calculation formula is: Wherein , is the number of attention heads.

6. The method of claim 1, wherein, The generation strategy of the stratification action in step (5) is: the first layer target floor decision strategy wherein is the first layer target floor decision strategy characteristic of the first elevator Second tier dynamic response decision policy wherein is a call point feature, indicates whether to respond; the competition coordination is implemented by priority , the highest priority elevator is selected to respond to the call.

7. A master-slave elevator group dispatching simulation system for implementing the method according to any one of claims 1 to 5, characterized in that the system Comprising: (1) Master control module: deploying a central dispatcher and a global feature processor (including ST-GCN and Transformer modules), responsible for global state representation generation, competition coordination, and simulation parameter configuration; (2) Slave node module: each elevator corresponds to a slave node, deploying a local intelligent agent, executing the hierarchical decision-making actions in step (5) of claim 1, and feeding back real-time state data to the master control module; (3) Environment simulation module: simulating building environment (floor layout, passenger flow model, blind layer setting), call generation (random / peak mode / inter-floor mode), and elevator physical operation (acceleration / deceleration, maximum speed, door opening and closing time, maximum load), outputting simulation logs; (4) Interactive visualization module: real-time display of elevator running trajectory, call allocation state, and performance indicators (average and maximum waiting time, average and maximum boarding time, average and maximum), supporting parameter adjustment and multi-scenario comparison.

8. The method of claim 7, wherein, The master control module and the slave node module realize data interaction through a communication protocol, and the master control module updates the global space-time graph according to the state data fed back by the slave nodes, and the slave nodes adjust the running state according to the dispatching instructions issued by the master control module.

9. The system of claim 7, wherein, The environment simulation module includes a passenger behavior simulator that can simulate passenger call frequency, target floor distribution, and boarding and alighting time in different scenarios, generating simulation data that conforms to actual traffic characteristics.

Images