An elevator group control prediction partitioning method based on implicit representation of operation logs

Abstract

The application discloses an elevator group control prediction partitioning method based on running log implicit representation and belongs to the technical field of elevator group control systems. Structured running state data is extracted from the running log in the elevator group control system, fixed time length sliding windows are adopted for aggregate statistics, and the implicit passenger flow representation vector of each floor is constructed through a gated adaptive fusion mechanism. The time sequence evolution characteristics of the implicit passenger flow representation vector sequence are learned by using BiLSTM, and the space-time characteristics are generated in combination with the floor topological relationship. The space-time characteristics of multiple time windows are subjected to average pooling aggregation, and the enhanced space-time characteristics are obtained. The enhanced space-time characteristics are input into a multi-layer fully connected network to generate a partition boundary probability vector, and the boundary points with a probability exceeding a threshold value are selected as dynamic cutting points. The application realizes adaptive dynamic partitioning without real passenger flow labels, and significantly improves the response efficiency and robustness of the elevator group control system in a complex passenger flow scenario.

Description

Technical Field

[0001] This invention belongs to the technical field of elevator group control systems, specifically relating to an elevator group control prediction and partitioning method based on implicit representation of operation logs. Background Technology

[0002] Efficient passenger flow prediction and floor zoning are key to improving the service quality of elevator group control systems. However, existing technologies still face significant challenges in handling complex vertical traffic flows. First, detection methods based on physical sensors are prone to data privacy violations due to biometric data collection, and their deployment costs in large-scale buildings are extremely high. Second, mainstream prediction algorithms mostly employ traditional supervised learning frameworks, heavily relying on labeled passenger flow sequences, which is quite limiting and makes it difficult to handle unstructured implicit features in elevator operation logs. Simultaneously, existing technologies lack modeling of topological relationships between floors, failing to utilize graph structures to characterize passenger flow intensity and heterogeneous relationships between floors. Finally, zoning strategies in group control systems are generally static, unable to adapt to real-time passenger flow trends. More importantly, the lack of an effective alignment mechanism between the prediction logic and the execution state at the bottom of the elevator floor leads to deviations in the prediction data's guidance of physical execution.

[0003] For example, patent CN121516671A proposes a building elevator operation optimization method based on a dynamic allocation algorithm. Although this method can use a weighted combination of historical data and real-time monitoring data for congestion prediction, its scheduling is limited to path order optimization and does not involve a dynamic partitioning mechanism. Patent CN121626795A proposes an elevator group control energy-saving scheduling method based on load prediction. Although this method uses particle swarm optimization for optimization and introduces a correction value for passenger flow prediction, its prediction logic is essentially a shallow time-series statistical matching based on historical periodic data, completely ignoring the spatial topological relationship between floor nodes and the deep modeling of passenger flow intensity. Patent CN121097787A proposes a power storage and retrieval control method based on big data. Although this method mentions obtaining elevator cabinet control signals and operating status, its technical focus heavily relies on external video surveillance image recognition and high-cost real passenger flow statistical tags, and it also has privacy leakage issues.

[0004] Therefore, a predictive zoning method for elevator group control based on implicit representation of operation logs is proposed. Without relying on real passenger flow count labels, passenger flow distribution is predicted based on elevator group control operation log data. This solves the problems of existing technologies that highly rely on dedicated sensors or real passenger flow labels, and that the zoning strategy is static and fixed and cannot reflect the dynamic evolution of passenger flow. Summary of the Invention

[0005] This invention addresses the problems of existing technologies that heavily rely on dedicated sensors or real passenger flow tags, and whose zoning strategies are static and fixed, making it difficult to reflect the dynamic evolution of passenger flow. It proposes an elevator group control predictive zoning method based on implicit representations in operation logs.

[0006] A further objective of this invention is to transform the zoning problem into a constrained continuous boundary prediction problem and adopt a penalty weighting mechanism based on passenger flow correlation strength, so that the zoning results can adaptively aggregate high-frequency visiting floors while satisfying the physical continuity of the floors, thereby solving the problem that traditional static zoning is difficult to adapt to dynamic flow fields and realizing the strategy drift and adaptive adjustment of the group control system under different passenger flow patterns.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: an elevator group control prediction and partitioning method based on implicit representation of operation logs, comprising: S1. Collect the CAN bus operation logs of the elevator group control system to construct a hidden passenger flow representation; S2. A bidirectional long short-term memory network is used to mine the temporal features of the passenger flow representation vector and extract the floor-level passenger flow evolution representation with temporal coherence. S3. A sparse directed matrix based on the inter-floor relationship was constructed to extract the passenger flow coupling strength features across floors. S4. Based on the temporal characteristics of passenger flow on each floor and the correlation strength between floors, predict several dynamic segmentation points, divide the entire floor into multiple logically continuous intervals, and use the partitioning results for elevator group control and dispatching.

[0008] Preferably, the CAN bus operation log includes at least timestamp information and data fields; The timestamp information is used to align CAN frames from different sources to a unified time base; The original CAN frame is parsed into a structured operating status, including elevator number, current floor, called floor, target floor, door status, operating status, and load-related quantities; Event sequences are constructed based on structured operational status, including elevator call, station stop, and door opening / closing events.

[0009] Preferably, in step S1, a fixed-duration sliding window is used to aggregate operational events. By statistically analyzing the number of floor calls, the number of door reopens, and the cumulative change in load, an implicit passenger flow representation for the corresponding time window is constructed. By transforming elevator physical actions into traffic flow proxy indicators, the system can achieve seamless perception of passenger flow trends without the need for additional sensing equipment.

[0010] Preferably, when constructing the implicit passenger flow representation for the corresponding time moment, the operational event indicators statistically analyzed within each time window include: The number of elevator call events or the number of valid calls at each floor within a unit time window represents the call density. The number of stops within a unit time window represents the stop density; The number of times the door reopens within a unit time window represents the door opening frequency; and The load change, obtained by the cumulative change of load-related quantities within a unit time window, characterizes the load change features. The frequency of passage events between the starting floor and the target floor within a unit time window represents the passage frequency between different floor nodes.

[0011] As a preferred approach, after aggregating and statistically analyzing operational events, a load change mapping is established to characterize the intensity of passenger flow exchange within the time window of that floor; and an event frequency mapping is established to characterize the traffic activity of that floor. A gated adaptive fusion mechanism is introduced, which uses a learnable scaling factor to nonlinearly couple load change characteristics with event frequency characteristics to generate the implicit passenger flow representation vector.

[0012] As a preferred option, in step S2, the implicit passenger flow vector sequence is input into the BiLSTM model for deep modeling, extracting the long-term dependence features of floor-level passenger flow in the time dimension, and capturing the periodic patterns and temporal evolution trends of passenger flow distribution in vertical transportation.

[0013] As a preferred option, in step S3, a sparse directed matrix of passenger flow relationships between floors is constructed, and information propagation is performed on the graph structure defined by the sparse directed matrix. Through GAT, floor nodes dynamically calculate the attention weights of neighboring nodes based on the injected temporal features, and aggregate spatial coupling information through Top-K dynamic masking, finally outputting the spatiotemporal feature vector S.

[0014] Preferably, in step S4, a state-aware correction operator for system operating capacity is introduced during the partition boundary generation stage; the system operating capacity is determined based on the current number of available elevators and is used to constrain the final number of partitions; when it is detected that some elevators have stopped running, the number of partitions is reduced accordingly, and the coverage of each partition is adjusted.

[0015] Preferably, the spatiotemporal feature vector S is aggregated to generate the final enhanced spatiotemporal feature H; The enhanced spatiotemporal features H are input into a multilayer fully connected network. The abstract spatiotemporal semantic features are reduced in dimensionality and mapped to the floor decision space through nonlinear transformation, providing a pre-feature representation for the generation of boundary probabilities. The Sigmoid activation function is used to normalize the mapped features, generating a partition boundary probability vector B, and a corresponding partitioning scheme is generated based on the number of partitions.

[0016] Preferably, in step S4, a constrained joint loss function Loss is constructed. Loss consists of two losses: the balancing loss Loss and the equilibrium loss Loss. bal Loss with topology conflict topo ; The decision parameters are continuously corrected through gradient backpropagation to minimize the total loss. Based on the converged probability vector, the dynamic split point where the probability peak exceeds the preset threshold is selected as the physical partition boundary, and the dynamic partition structure is output.

[0017] The beneficial effects of this invention are as follows: By using the multi-source operation logs inherent in the elevator group control system as the basis for perception, and by constructing a mapping operator from physical actions to latent passenger flow variables, passenger flow modeling is achieved without the need for additional passenger flow detection equipment and real passenger flow labels. This avoids expensive hardware deployment and maintenance costs, and at the same time solves the privacy leakage risk brought about by visual perception solutions from the source. By introducing a temporal feature injection mechanism and a graph attention network, the system can transform discrete elevator operation events into structured passenger flow representations with spatiotemporal coupling. This helps the system accurately capture the traffic burst patterns and dynamic attraction between floors during specific time periods within the building, significantly improving the feature expression depth of the group control system for complex passenger flow changes. By transforming the zoning problem into a constrained continuous boundary prediction problem and adopting a penalty weight mechanism based on the strength of passenger flow association, the zoning results can adaptively aggregate high-frequency visiting floors while satisfying the physical continuity of the floors. This mechanism effectively solves the problem that traditional static zoning is difficult to adapt to dynamic flow fields, and facilitates the strategy drift and adaptive adjustment of the group control system under different passenger flow patterns. By introducing a state-aware correction operator, the node output under abnormal operating conditions can be intervened in real time, which effectively avoids the defect of the traditional prediction model being out of touch with the physical system state. This helps to ensure that the generated passenger flow distribution data has the physical effectiveness to directly guide elevator dispatching, especially for twin elevators, thereby improving the robustness of the group control system under extreme operating conditions. Attached Figure Description

[0018] Figure 1 This is a flowchart of a predictive partitioning method for elevator group control based on implicit representations in runtime logs.

[0019] Figure 2 This is a schematic diagram of the overall architecture of time-space grid construction in one embodiment of the present invention, including sparse matrix creation, time series modeling unit and spatial floor relationship modeling unit.

[0020] Figure 3 This is a schematic diagram illustrating the process of determining the partition boundary based on floor passenger flow characteristics and inter-floor relationships in one embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and beneficial effects of the present invention clearer, the specific embodiments of the present invention will be described below in conjunction with the accompanying drawings.

[0022] Example 1. This example provides a method for learning and predicting passenger flow representation in elevator group control based on multi-source operating status and spatiotemporal coupling modeling.

[0023] This method, without requiring additional hardware sensing devices or privacy compliance constraints, deeply mines multi-source heterogeneous data from the CAN bus operation logs inherent in the elevator group control system. This data includes elevator number, current floor, called floor, target floor, door operator status, and load changes. By constructing a latent passenger flow representation, the sequence of elevator physical actions is transformed into a multi-dimensional passenger flow situation vector. In the fusion phase, an adaptive mechanism based on learnable weights is introduced, and initialization is performed using prior distributions such as load changes and door control duration, achieving accurate capture of latent passenger flow variables for each floor.

[0024] In the temporal dimension, a Bidirectional Long Short-Term Memory (BiLSTM) network is employed to mine temporal features from the passenger flow representation vector, extracting a floor-level passenger flow evolution representation with temporal coherence. In the spatial dimension, this invention abandons the traditional static partitioning criteria and constructs a sparse directed matrix based on the inter-floor relationships. This matrix eliminates physically weakly correlated spatial noise through physical topological constraints and a dynamic Top-K masking mechanism, and utilizes a Graph Attention Network (GAT) to extract cross-floor passenger flow coupling strength features, thereby characterizing the flow attraction between different floors.

[0025] At the decision-making level, this invention transforms the dynamic partitioning problem into a continuous boundary prediction problem under spatiotemporal constraints. By predicting several dynamic dividing points, the entire floor is divided into multiple logically continuous intervals. The determination of partition boundaries no longer relies on fixed criteria, but is jointly optimized based on the temporal characteristics of passenger flow on each floor and the correlation strength between floors. Specifically, a "correlation strength-potential energy penalty" mechanism is introduced in the partitioning process: when the passenger flow frequency or correlation strength between two floors exceeds a preset threshold, the algorithm automatically increases the cost of assigning them to different partitions. This method ensures that the partitioning results, while satisfying physical continuity, can adaptively balance the consistency of passenger flow distribution and the collaborative scheduling relationship between floors, significantly improving the response efficiency and energy consumption balance capability of the group control system in complex passenger flow scenarios.

[0026] Please refer to the overall flowchart of the method of this invention. Figure 1 and Figure 2 First, in the data preprocessing stage, the CAN bus operation log of the elevator group control system is acquired. Through timestamp alignment and structured parsing, multi-dimensional information including elevator load, door operator status, floor changes, and operating status is extracted.

[0027] The CAN bus log includes at least timestamp information and data fields; the timestamp information is used to align CAN frames from different sources to a unified time base. The raw CAN frames are parsed into structured operating states, including elevator number, current floor, target floor, door status, operating status, and load-related quantities. An event sequence is constructed based on the structured operating states, including elevator call, stop, and door open / close events.

[0028] Subsequently, a sequence of implicit passenger flow representation vectors is constructed based on multidimensional information. The detailed construction process includes using a sliding window with a preset step size to aggregate and statistically analyze running events.

[0029] A fixed-duration sliding window is used to aggregate multi-source events. By statistically analyzing the number of calls at each floor, the number of stops, the frequency of door openings, and the cumulative changes in load, an implicit passenger flow representation for the corresponding time window is constructed. By converting elevator physical actions into flow proxy indicators, the system can achieve seamless perception of passenger flow trends without the need for additional sensing equipment.

[0030] The operational event metrics counted within each time window include: the number of elevator call events or the number of valid floor calls within a unit time window, used to characterize call density; the number of stops within a unit time window, used to characterize stop density; the number of door reopens within a unit time window, used to characterize door opening frequency; the load change amount obtained from the cumulative changes of load-related quantities within a unit time window, used to characterize load change characteristics; and the frequency of passage events between the starting floor and the target floor within a unit time window, used to characterize the passage frequency between different floor nodes.

[0031] On the one hand, a load change mapping is established by calculating the load difference ∆L before and after the elevator stops at a floor, and combining this with the elevator's rated load L. rated Normalization is performed to characterize the instantaneous passenger flow exchange intensity f on that floor. .

[0032] On the other hand, an event frequency mapping is established, and the response frequency (call density + docking density) of stations within a unit window is statistically analyzed. i The cumulative duration d of the gantry crane being open i and the frequency of reopening the door n i This is used to characterize the traffic activity level of that floor. Finally, a gated adaptive fusion mechanism is adopted, which uses a learnable scaling factor to nonlinearly couple the aforementioned load variation characteristics with event frequency characteristics to generate a hidden passenger flow representation vector V. i , .

[0033] in, For vector concatenation operators; and These are learnable scaling parameters for load characteristics and frequency characteristics, respectively, used to adjust the contribution of different dimensions. Their initial values ​​can be guided by prior knowledge to help the model converge quickly. This is the gated weight matrix, used to achieve nonlinear coupling of heterogeneous features; b This is the bias term. This mechanism enables the generation of latent vectors that can deeply characterize the latent variables of floor passenger flow without the need for physical sensors.

[0034] Trainable scaling factors are used and applied to the representation vector through gating units; by introducing prior knowledge such as load changes to set initial weights, the model can quickly lock in key passenger flow features in the early stage of training and avoid feature noise interference in the process of heterogeneous data fusion.

[0035] After completing the implicit passenger flow representation, this invention enters the spatiotemporal deep modeling stage, which aims to generate high-order enhanced features H by capturing the temporal evolution pattern and spatial coupling relationship of passenger flow.

[0036] First, the hidden passenger flow vector sequence Input BiLSTM. BiLSTM extracts long-term dependency features of passenger flow in the time dimension through the collaborative action of forward and backward hidden layers. This step ensures that the feature vector can perceive the evolution trend of passenger flow before and after the current moment.

[0037] The BiLSTM model is used to perform deep modeling of the time window sequence; and the temporal feature representation of passenger flow changes at each floor level is extracted, enabling the system to capture the periodic patterns and temporal evolution trends of passenger flow distribution in vertical transportation.

[0038] In order to accurately characterize the physical relationships between floors and reduce background noise, this invention extracts deep spatial relationships by mapping temporal features to physical topology.

[0039] First, an original traffic frequency matrix is ​​constructed based on historical operational data. Using threshold Generate sparse directed matrices , .

[0040] Subsequently, the temporal features output by BiLSTM are... As the initial feature of a node, in a sparse directed matrix Information propagation occurs on the defined graph structure. Using GAT, nodes (floors) dynamically calculate the attention weights of their neighbors based on injected temporal features, and spatial coupling information is aggregated through a Top-K dynamic mask, ultimately outputting a spatiotemporal feature vector S. .

[0041] in, It is a non-linear activation function; This represents the set of neighboring nodes that are connected to node i under the constraints of the sparse matrix. This represents the attention weight coefficient between node i and node j. Its calculation is constrained by the sparse matrix structure and only participates in the calculation when there is a connection relationship. W V is the characteristic linear transformation matrix; j This represents the temporal features of node j extracted by BiLSTM.

[0042] Finally, the spatiotemporal features of different time windows are deeply aggregated through an average pooling layer to generate the final enhanced spatiotemporal feature representation H, which contains deep spatiotemporal evolution information.

[0043] Please refer to the schematic diagram of the partition boundary determination process of this invention. Figure 3 This stage aims to transform the enhanced spatiotemporal features H into a physically feasible dynamic partitioning scheme. By introducing topological constraint loss, it ensures that the decision results take into account both the flow balance of partitions and the continuity of intervals.

[0044] The floor zoning is transformed into a boundary location prediction problem. Multiple continuous intervals are generated by predicting dynamic dividing points. The zoning results are constrained by a penalty weight based on the strength of passenger flow association, so that closely related floors tend to be assigned to the same zoning, which helps to optimize the dispatch efficiency of the group control system and reduce passenger transfer costs.

[0045] First, the enhanced spatiotemporal features H output from Stage 2 are input into a multilayer fully connected network. This process uses nonlinear transformation to reduce the dimensionality of the abstract spatiotemporal semantic features and map them to the floor decision space, providing a preliminary feature representation for the generation of boundary probabilities.

[0046] To accurately capture the floor segmentation location, the system uses the Sigmoid activation function to normalize the mapped features, generating a partition boundary probability vector B. ;in, This represents the probability that a partition boundary exists between layer i and layer i+1. A state-aware correction operator for system operational capability is introduced at this stage to constrain the final number of partitions. This step transforms the partitioning problem into a continuous probability estimation problem, facilitating end-to-end optimization via gradient descent.

[0047] The core of this invention lies in constructing a constrained joint loss function, Loss, such as... Figure 3 As shown, the loss consists of two losses: the balancing loss and the loss itself. bal After initial partitioning of floors based on boundary probability B, the load within each partition is weighted and aggregated, and its dispersion is calculated. The goal is to force the model to find the split point that most evenly distributes capacity across partitions; and to penalize loss due to topology conflicts. topo When the model attempts to access routes with strong correlation frequencies (i.e., When establishing partition boundaries between relatively large floor pairs, this item will generate a high penalty, implementing a hard constraint on the partitioning results, thereby guiding the model to avoid partitioning decisions that disrupt passenger flow continuity.

[0048] The decision parameters are continuously adjusted through gradient backpropagation to minimize the total loss. Finally, based on the converged probability vector, the system selects the points where the probability peak exceeds a preset threshold as the physical partition boundary and outputs the dynamic partition structure.

[0049] The design of the aforementioned joint loss function essentially transforms the dynamic partitioning decision in the elevator group control system into a continuous mathematical problem that can be optimized end-to-end. Its core lies in constructing a constrained multi-objective optimization framework, driving a neural network to automatically search for the optimal partition boundary under competitive constraints through gradient backpropagation. The balancing loss aims to force each partition to tend towards equal capacity load, avoiding an imbalance where some partitions are overloaded while others are idle. Its calculation process includes three logical stages. First, in the soft partitioning stage, the system performs initial soft partitioning on all floors based on the currently output boundary probability vector B, dividing the floor sequence into several continuous candidate intervals, each representing a logical partition. Second, in the load aggregation stage, for each candidate partition, the system performs a weighted summation of the enhanced spatiotemporal features of each floor generated in the preceding spatiotemporal modeling stage (which contain the predicted passenger flow exchange intensity of that floor), thereby obtaining the total capacity demand scalar for that partition. Finally, in the dispersion evaluation stage, the system calculates the variance or standard deviation of the load of all partitions and directly uses this dispersion value as the output value of the total loss function. When the loss term tends to be minimized, the passenger flow load distribution in each partition tends to be uniform, which means that the model is forced to be guided to those partition locations that can achieve load balancing.

[0050] Topological conflict penalty aims to protect strong connectivity between floors and prevent zoning decisions from severing adjacent floors with high-frequency passenger flow interactions. The construction of this penalty term strictly relies on a sparse directed matrix, where elements A... sparse(i,j) quantitatively characterizes the historical frequency of passage between layer i and layer j; a higher value indicates a stronger association. The penalty term is calculated using a product coupling mechanism; specifically, the boundary probability B is... i Corresponding correlation strength A sparse Multiply the values ​​at (i,j). For all candidate boundary positions spanning the entire floor sequence, sum the products; the sum is the Loss. topo The mathematical essence of this loss term is the construction of an asymmetric rigid constraint. When the splitting probability of a candidate boundary position is high, and the corresponding floor pairs have strong access correlations, this loss term will produce a large penalty value; conversely, if the splitting occurs at the gap between floors with extremely weak correlation, the penalty approaches zero. Through this mechanism, the system is forced to reduce the boundary probability between strongly correlated floor pairs during the optimization process, thereby ensuring that frequently visited floors tend to be assigned to the same logical partition.

[0051] The number of fully connected layers, activation function type, sparse matrix filtering threshold, and weight coefficients in the joint loss function listed in this embodiment are merely illustrative examples of preferred embodiments of the present invention. In practical applications, those skilled in the art can adjust the above parameters, network depth, and communication interface protocol according to the physical floor height of the elevator system, building traffic flow density, and real-time performance requirements.

[0052] In another preferred embodiment, spatiotemporal modeling and traffic prediction adopt an offline training and periodic update mode, that is, historical log data is asynchronously accumulated in the background to evolve model parameters, while in actual operation, prediction and partitioning results are output in real time based on the current perception data through an online inference mechanism, thereby significantly reducing real-time computing overhead while ensuring the system's adaptive capability.

[0053] In another preferred embodiment, a state-aware correction operator for system operating capacity is introduced during the partition boundary generation stage; the system operating capacity is determined based on the number of currently available elevators and is used to constrain the final number of partitions; when it is detected that some elevators have exited operation, the number of partitions is reduced accordingly, and the coverage of each partition is adjusted.

[0054] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predictive partitioning of elevator group control based on implicit representations in operation logs, characterized in that, include: S1. Collect the CAN bus operation logs of the elevator group control system to construct a hidden passenger flow representation; S2. A bidirectional long short-term memory network is used to mine the temporal features of the passenger flow representation vector and extract the floor-level passenger flow evolution representation with temporal coherence. S3. A sparse directed matrix based on the inter-floor relationship was constructed to extract the passenger flow coupling strength features across floors. S4. Based on the temporal characteristics of passenger flow on each floor and the correlation strength between floors, predict several dynamic segmentation points, divide the entire floor into multiple logically continuous intervals, and use the partitioning results for elevator group control and dispatching.

2. The elevator group control prediction and partitioning method based on implicit representation of operation logs according to claim 1, characterized in that, The CAN bus operation log includes at least timestamp information and data fields; The timestamp information is used to align CAN frames from different sources to a unified time base; The original CAN frame is parsed into a structured operating status, including elevator number, current floor, called floor, target floor, door status, operating status, and load-related quantities; Event sequences are constructed based on structured operational status, including elevator call, station stop, and door opening / closing events.

3. The elevator group control prediction and partitioning method based on implicit representation of operation logs according to claim 1, characterized in that, In step S1, a fixed-duration sliding window is used to aggregate the running events. By statistically analyzing the number of calls at each station, the number of door reopens, and the cumulative change in load, an implicit passenger flow representation for the corresponding time window is constructed.

4. A method for predictive partitioning of elevator group control based on implicit representation of operation logs according to claim 1 or 3, characterized in that, When constructing the implicit passenger flow representation for the corresponding time moment, the operational event indicators statistically analyzed within each time window include: The number of elevator call events or the number of valid calls at each floor within a unit time window represents the call density. The number of stops within a unit time window represents the stop density; The number of times the door reopens within a unit time window represents the door opening frequency; and The load change, obtained by the cumulative change of load-related quantities within a unit time window, characterizes the load change features. The frequency of passage events between the starting floor and the target floor within a unit time window represents the passage frequency between different floor nodes.

5. A method for predictive partitioning of elevator group control based on implicit representation of operation logs according to claim 1 or 3, characterized in that, In step S1 After aggregating and statistically analyzing operational events, a load change mapping is established to characterize the intensity of passenger flow exchange within a time window on that floor; an event frequency mapping is also established to characterize the traffic activity on that floor. A gated adaptive fusion mechanism is introduced, which uses a learnable scaling factor to nonlinearly couple load change characteristics with event frequency characteristics to generate the implicit passenger flow representation vector.

6. The elevator group control prediction and partitioning method based on implicit representation of operation logs according to claim 1, characterized in that, In step S2, the implicit passenger flow vector sequence is input into the BiLSTM model for deep modeling to extract the long-term dependence features of floor-level passenger flow in the time dimension.

7. The elevator group control prediction partitioning method based on implicit representation of operation logs according to claim 1, characterized in that, In step S3, a sparse directed matrix of passenger flow relationships between floors is constructed, and information propagation is performed on the graph structure defined by the sparse directed matrix. Through the graph attention network, the floor nodes dynamically calculate the attention weights of neighboring nodes based on the injected temporal features, and aggregate spatial coupling information through Top-K dynamic masking, finally outputting the spatiotemporal feature vector S.

8. The elevator group control prediction and partitioning method based on implicit representation of operation logs according to claim 1, characterized in that, In step S4, a state-aware correction operator for system operating capability is introduced during the partition boundary generation stage; The system's operational capacity is determined based on the number of currently available elevators, which constrains the final number of partitions. When it is detected that some elevators have stopped operating, the number of partitions is reduced accordingly, and the coverage area of ​​each partition is adjusted.

9. A method for predictive partitioning of elevator group control based on implicit representation of operation logs according to claim 1 or 7, characterized in that, The spatiotemporal feature vector S is aggregated to generate the final enhanced spatiotemporal feature H; The enhanced spatiotemporal features H are input into a multilayer fully connected network. The abstract spatiotemporal semantic features are reduced in dimensionality and mapped to the floor decision space through nonlinear transformation, providing a pre-feature representation for the generation of boundary probabilities. The Sigmoid activation function is used to normalize the mapped features, generating a partition boundary probability vector B, and a corresponding partitioning scheme is generated based on the number of partitions.

10. The elevator group control prediction and partitioning method based on implicit representation of operation logs according to claim 1, characterized in that, In step S4, Construct a constrained joint loss function Loss, which consists of two losses: the balancing loss Loss. bal Loss with topology conflict topo ; The decision parameters are continuously corrected through gradient backpropagation to minimize the total loss. Based on the converged probability vector, the dynamic split point where the probability peak exceeds the preset threshold is selected as the physical partition boundary, and the dynamic partition structure is output.

Images