AI-based elevator peak elevator taking flow dynamic scheduling system
By constructing a multi-level collaborative intelligent decision-making architecture, the problem of limited computing resources under edge deployment of elevator dispatching systems is solved, enabling high-precision prediction and dynamic scheduling of elevator demand, and improving the operating efficiency and service quality of elevator systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHONG TIE CHENG JIAN JI TUAN HUA DONG JIAN SHE YOU XIAN GONG SI
- Filing Date
- 2026-01-19
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies, when deployed at the edge of elevator dispatching systems, suffer from problems such as decreased model prediction accuracy, slow response during peak passenger flow, and uneven distribution of system resources due to severely limited computing resources.
A multi-level collaborative intelligent decision-making architecture is constructed, including an edge perception and feature extraction module, a cloud model collaborative optimization module, a dynamic scheduling decision module, and a system resource adaptive allocation module. Through real-time data collection by edge computing nodes, cloud model collaborative optimization, and dynamic scheduling of edge nodes, high-precision elevator demand prediction and dynamic scheduling are achieved.
Achieving high-precision elevator demand forecasting in resource-constrained edge environments improves the operational efficiency and service quality of elevator systems during peak hours, while reducing passenger waiting time, system energy consumption, and load balancing.
Smart Images
Figure CN122009928A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of elevator intelligent control technology, specifically relating to an AI-based elevator peak passenger flow dynamic scheduling system. Background Technology
[0002] In the fields of smart buildings and vertical transportation automation, the scheduling efficiency of elevator systems directly affects the travel experience of people within buildings and energy consumption. With the increasing prevalence of high-rise buildings and the deep integration of IoT technology, achieving intelligent and adaptive elevator group control has become a key direction for industry development.
[0003] Artificial intelligence-based elevator dispatching systems analyze historical passenger data, real-time passenger flow information, and building operating status to dynamically optimize elevator dispatching strategies, thereby reducing passenger waiting time, balancing system load, and improving overall operational efficiency. Existing technologies typically employ complex predictive models deployed in the cloud or on a central server to achieve dispatching decisions.
[0004] Existing technologies face a contradiction between model deployment and real-time performance. When models are lightweighted and deployed on edge devices to reduce latency, the accuracy of model predictions often drops significantly due to strict limitations in computing resources, making it difficult to accurately capture and respond to sudden fluctuations in passenger flow during peak periods.
[0005] This loss of accuracy is particularly prominent in typical peak scenarios such as morning and evening commuting in office buildings. It can easily lead to scheduling decision errors, uneven resource allocation problems such as some elevators being overcrowded while others are idle, and seriously reduce the service reliability and user experience of the system.
[0006] Therefore, there is an urgent need for an elevator traffic management solution that can maintain high-precision dynamic scheduling capabilities in resource-constrained edge environments. Summary of the Invention
[0007] The technical problem this invention aims to solve is to overcome the shortcomings of existing technologies in edge deployment scenarios for elevator dispatching systems, such as decreased model prediction accuracy, slow peak passenger flow response, and uneven system resource allocation due to severely limited computing resources. The purpose of this invention is to provide an AI-based dynamic scheduling system for peak elevator passenger flow. By constructing a multi-level collaborative intelligent decision-making architecture, it achieves high-precision passenger flow prediction and dynamic scheduling under the condition of limited resources on edge computing nodes.
[0008] To achieve the above objectives, the present invention provides an AI-based dynamic scheduling system for peak elevator passenger flow, which includes an edge perception and feature extraction module, a cloud-based model collaborative optimization module, a dynamic scheduling decision module, and a system resource adaptive allocation module.
[0009] The edge perception and feature extraction module is deployed on the edge computing nodes of each elevator group controller in the building. It is used to collect elevator operation status data, car load data, call request data of each floor, and queue statistics information obtained by visual sensors deployed in the waiting hall in real time.
[0010] This module further preprocesses and extracts features from the collected raw data to generate a multi-dimensional feature vector that includes time series features, spatial distribution features, and passenger flow intensity features.
[0011] The cloud-based model collaborative optimization module runs on a central server. It receives feature vectors and historical scheduling results data uploaded by multiple edge nodes and builds a passenger flow prediction model based on a deep neural network.
[0012] The model continuously optimizes its network parameters through periodic incremental learning, and then distributes the optimized model parameters to each edge node in a differential update manner.
[0013] The dynamic scheduling decision module is integrated into the edge node. It loads a lightweight prediction model delivered from the cloud, predicts the elevator demand within a specific future time window based on real-time feature vectors, and generates a preliminary elevator dispatch plan according to the preset scheduling objective function.
[0014] The system resource adaptive allocation module monitors the computing load and memory usage of edge nodes in real time. When the load exceeds a set threshold, it dynamically adjusts the number of dimensions for feature extraction or the complexity level of the prediction model to ensure the real-time nature of scheduling decisions.
[0015] Furthermore, the feature extraction process performed by the edge perception and feature extraction module specifically includes the following steps: First, the raw sensor data is filtered to remove noise and outliers. Secondly, the passenger flow count sequence with 5-minute intervals is extracted from the time dimension, the call frequency distribution of each floor is extracted from the spatial dimension, and the change rate of passenger flow in the current period relative to the same period in history is extracted from the intensity dimension. Finally, the above features are standardized and combined into a feature vector with fixed dimensions.
[0016] Furthermore, the deep neural network used in the cloud-based model collaborative optimization module is a hybrid architecture of long short-term memory network and self-attention mechanism.
[0017] The network takes multi-dimensional feature vectors uploaded from the edge as input, captures the temporal dependencies through three layers of long short-term memory network units, and then uses a self-attention mechanism layer to weight and fuse the importance of different features, finally outputting the predicted value of elevator demand for each floor in the next 15 minutes.
[0018] The model training process uses mean squared error as the loss function and performs an incremental learning round every 24 hours using the latest collected data. The update of model parameters is compressed and then sent to the edge.
[0019] Furthermore, the scheduling objective function used by the dynamic scheduling decision module is a multi-objective optimization function, which simultaneously considers three optimization objectives: minimizing the average waiting time of passengers, minimizing the total energy consumption of the elevator system, and maximizing the load balance of each car.
[0020] This module uses an improved genetic algorithm to iteratively optimize the initial ladder scheme. The population size of the genetic algorithm is set to 100, the maximum number of iterations is 50 generations, the crossover probability is set to 0.85, and the mutation probability is set to 0.02.
[0021] Furthermore, the system resource adaptive allocation module dynamically adjusts the system operating mode based on the real-time utilization rate of the central processing unit of the edge nodes.
[0022] When the CPU utilization rate is greater than 80% for 10 consecutive seconds, the module automatically reduces the dimension of the feature vector from 128 to 64 and switches to a lightweight prediction model with lower computational complexity. When the occupancy rate drops below 60% and remains below that level for 30 seconds, the system reverts to the computation mode for high-dimensional features and the complete model.
[0023] Furthermore, the system also includes a safety and fault-tolerant control submodule, which continuously monitors the operating status of the elevator group and the execution results of scheduling instructions.
[0024] If an elevator fails to respond to dispatch instructions for two consecutive scheduling cycles, or if the actual load of the car deviates significantly from the predicted demand by more than 25%, the abnormal handling process will be triggered immediately, putting the elevator into manual mode and reassigning the call tasks to the floors it serves.
[0025] Furthermore, the entire system operates within a hierarchical time-scale coordination framework, where cloud-based model optimization is performed on a daily basis, edge passenger flow prediction is performed on a 5-minute basis, dynamic scheduling decisions are generated on a 15-second basis, and resource allocation and security monitoring are performed in real time at the second level.
[0026] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves high-precision elevator demand prediction at resource-constrained edge nodes by constructing a model optimization mechanism that coordinates edge and cloud environments, overcoming the inherent defect of insufficient accuracy in lightweight models. By introducing a system resource adaptive allocation module, the real-time performance and stability of the scheduling system under high load conditions are ensured. By adopting a dynamic scheduling decision-making method with multi-objective optimization, the overall operating efficiency and service quality of the elevator system during peak hours are improved, and the comprehensive optimization of passenger waiting time, system energy consumption and load balance is achieved. Attached Figure Description
[0027] Figure 1 This is a schematic diagram of the overall technical architecture of the AI-based elevator peak passenger flow dynamic scheduling system proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the edge-cloud collaborative model optimization mechanism in this invention; Figure 3 This is a logical flowchart of the edge perception and feature extraction module in this invention; Figure 4 This is a schematic diagram of the multi-objective optimization framework of the dynamic scheduling decision module in this invention; Figure 5 This is a schematic diagram of the operating mode switching logic of the system resource adaptive allocation module in this invention. Detailed Implementation
[0028] Please refer to the attached document. Figures 1 to 5 This embodiment details the specific implementation architecture of an AI-based elevator peak passenger flow dynamic scheduling system. The system consists of an edge perception and feature extraction module, a cloud-based model collaborative optimization module, a dynamic scheduling decision module, and a system resource adaptive allocation module. Each module achieves high-precision prediction and dynamic scheduling of peak passenger flow through a hierarchical collaborative mechanism.
[0029] The edge perception and feature extraction module is deployed on the edge computing nodes of each elevator group controller in the building.
[0030] This module collects elevator operating status data in real time through the data interface connected to the elevator control system, including the current position of the car, direction of travel, door status, and speed information; and obtains real-time load data through weighing sensors installed at the bottom of the car. Call request data is collected through input records on the call panels on each floor; and queue statistics are collected through visual sensors deployed on the top of the elevator lobby.
[0031] The visual sensor uses an infrared depth camera to capture images of the waiting area at a rate of 5 frames per second, and executes a head detection algorithm through an embedded image processing unit to output a queue count that is updated every 5 seconds.
[0032] This module preprocesses and extracts features from the collected raw data. Please refer to the appendix for the specific process. Figure 3 .
[0033] The preprocessing stage first applies Kalman filtering to the sensor data to eliminate random noise, and then uses the 3σ criterion to remove outliers from the load and call data. The feature extraction stage generates feature vectors from three dimensions: In terms of time, the passenger flow count sequence for each floor is statistically analyzed at 5-minute intervals, and its mean, variance, and first difference are calculated. In terms of spatial dimension, the distribution of call frequency on each floor is statistically analyzed, and the entropy value is calculated to characterize the uniformity of distribution; In terms of intensity, the rate of change of passenger flow in the current period relative to the passenger flow in the same period of the past 7 days is calculated. The formula for calculating the rate of change is: ; This represents the passenger flow count for the current time period. This represents the average passenger flow for the same period over the past 7 days. The passenger flow count for the current period is the average passenger flow for the same period in history.
[0034] Finally, the 24-dimensional data, consisting of 12 temporal features, 8 spatial features, and 4 intensity features, were standardized using Z-score and combined into a fixed-dimensional feature vector.
[0035] The cloud-based model collaborative optimization module runs on a central server, which is equipped with two 16-core CPUs and four computing accelerator cards.
[0036] This module receives 24-dimensional feature vectors and corresponding historical scheduling results uploaded from at least 50 edge nodes and stores them in a distributed time-series database.
[0037] The model employs a hybrid architecture combining a long short-term memory network and a self-attention mechanism. Please refer to the appendix for the specific structure. Figure 2 .
[0038] The network input layer receives a 24-dimensional feature vector, which is processed by three layers of long short-term memory network units. Each layer has 128 hidden units, and the activation function is the hyperbolic tangent function, which is used to capture time dependencies. The output of the Long Short-Term Memory Network is connected to the Self-Attention Mechanism Layer, which contains 8 attention heads. Each head generates a 16-dimensional attention weight vector to weight and fuse the importance of different features. Finally, the predicted sequence of elevator demand for each floor in the next 15 minutes is output through the fully connected layer.
[0039] The model training uses mean squared error as the loss function, and the optimizer uses the adaptive moment estimation algorithm. The initial learning rate is set to 0.001.
[0040] The model initiates an incremental learning process every 24 hours: The latest 30-day data is extracted from the time series database, and the training set and validation set are divided in an 8:2 ratio. After 50 rounds of training iterations, the parameter update amount is calculated. The update amount is then distributed to each edge node after Top-K sparsity compression (retaining the top 10% of the largest absolute value parameters).
[0041] The dynamic scheduling decision module is integrated into the edge node, which loads a lightweight prediction model delivered from the cloud. This model is a distilled version of the cloud model, retaining only a single-layer long short-term memory network unit and a 4-head self-attention mechanism, reducing the number of parameters to 20% of the full model.
[0042] Please refer to the attached document. Figure 4 The module performs a prediction task based on real-time feature vectors, and the prediction results are input into a multi-objective optimization framework to generate a preliminary ladder dispatch scheme. The scheduling objective function is defined as follows: ; , , These are the weighting coefficients. Average passenger waiting time This represents the total energy consumption of the elevator system. To ensure load balance in each car.
[0043] The sum of three factors is calculated by weighting the reciprocal of the average passenger waiting time, the reciprocal of the total energy consumption of the elevator system, and the load balance of each car.
[0044] The weight coefficients are determined using the analytic hierarchy process (AHP), with typical values of 0.5, 0.3, and 0.2. This module employs an improved genetic algorithm for iterative optimization. The initial population size is 100, and each individual is encoded as a sequence of 10 elevators. The selection operation adopts a tournament selection strategy with a tournament size of 5. The crossover operation adopts a two-point crossover with a crossover probability of 0.85. The mutation operation uses uniform mutation with a mutation probability of 0.02; the maximum number of iterations is 50 generations, and 10 elite individuals are retained in each generation.
[0045] After optimization, the optimal elevator dispatch command is output to the elevator control system.
[0046] The system resource adaptive allocation module monitors the computing resource status of edge nodes in real time.
[0047] This module collects CPU usage and memory usage every second through the operating system interface and maintains a sliding window of length 30 to record historical data.
[0048] Please refer to the attached document. Figure 5 When the CPU utilization rate exceeds the 80% threshold for 10 consecutive sampling points, the module immediately triggers the degradation mode. The feature vector dimension was compressed from 24 dimensions to 12 dimensions, retaining only core features such as the mean of the time dimension, the entropy value of the spatial dimension, and the rate of change of the intensity dimension; at the same time, the prediction model was switched to a linear regression model, which fits the mapping relationship between features and elevator demand through the least squares method.
[0049] When the CPU utilization rate is less than 60% for 30 consecutive sampling points, the system automatically recovers the full-dimensional features and lightweight deep learning model.
[0050] The module also monitors memory usage. When it exceeds the 85% threshold, it activates a data cache cleanup mechanism, prioritizing the deletion of historical feature data older than 1 hour.
[0051] The safety and fault tolerance control submodule runs as an independent process on the edge node.
[0052] This module obtains real-time operational feedback from each elevator through the elevator status bus, and monitors indicators including command response status and the deviation between the actual load of the car and the predicted demand.
[0053] When an elevator fails to respond to a dispatch command for two consecutive dispatch cycles (30 seconds), or if the load deviation is greater than 25%, a Level 3 emergency response will be immediately triggered. A Level 1 response switches the elevator to manual mode and sends an alarm to the maintenance terminal. The Level 2 response recalculates the load distribution scheme for the remaining elevators; A Level 3 response initiates the connection of the backup elevator to the dispatch sequence. All abnormal events are logged in local non-volatile memory and uploaded to the cloud analysis platform via a secure channel.
[0054] The system operates within a hierarchical time-scale coordination framework: The cloud-based model optimization cycle is 24 hours, and the incremental learning process starts at 2:00 AM every day. The edge passenger flow prediction cycle is 5 minutes, and feature extraction and model inference are performed at every 5-minute interval. The dynamic scheduling decision cycle is 15 seconds, and a new round of elevator dispatching schemes is generated every 15 seconds; resource allocation and safety monitoring operate in real time with a cycle of 1 second.
[0055] The modules employ a timestamp-based data synchronization mechanism, achieving millisecond-level clock synchronization accuracy.
[0056] This embodiment achieves high-precision prediction and real-time decision-making for a dynamic elevator traffic scheduling system under the condition of limited edge computing resources through the fine coordination of the above modules.
[0057] Deployment tests of the system in typical office building scenarios show that the average waiting time for passengers during peak hours is reduced, the total energy consumption of the system is reduced, and the load balance of the elevator car is improved.
[0058] This embodiment provides an alternative implementation scheme for an AI-based elevator peak passenger flow dynamic scheduling system, focusing on optimizing the system architecture for super high-rise building scenarios.
[0059] Please refer to the attached document. Figure 1 This embodiment enhances the edge perception and feature extraction module based on the above embodiment by adding a multispectral vision sensor to the top of each elevator car to detect the real-time personnel density inside the car.
[0060] The sensor uses near-infrared and visible light dual-band imaging, and uses a convolutional neural network to segment the outline of the standing area, and combines the detection of key points in the skeleton to count the actual number of people taking the elevator.
[0061] The feature extraction dimension has been expanded to 32 dimensions, and four new features inside the car have been added, including the rate of change of car density and the entropy value of the heat map of personnel distribution.
[0062] The cloud-based model collaborative optimization module employs a hierarchical modeling strategy, establishing dedicated prediction models for different floor intervals. The building is divided into intervals of 20 floors, with each interval deploying an independent Long Short-Term Memory (LSTM) network-self-attention hybrid model. Model training data is stored in isolation for each floor interval, and incremental learning updates the models for each interval in parallel. Incremental encoding is used in the parameter distribution phase, transmitting only the parameter difference sequences between adjacent versions, reducing bandwidth usage by 40%.
[0063] The dynamic scheduling decision module introduces a flexible target weight mechanism, dynamically adjusting the objective function coefficients based on real-time passenger flow intensity. When the system detects that the passenger flow intensity on a certain floor exceeds 90% of the historical peak, it automatically increases the weight of the average passenger waiting time to 0.7; when the overall system energy consumption exceeds a set threshold, the energy consumption weight is increased to 0.5. The genetic algorithm population size adjusts linearly with the number of elevators, expanding by 15 individuals for each additional elevator, with a maximum population size limit of 200.
[0064] The system resource adaptive allocation module adds network bandwidth monitoring functionality. When the communication latency between edge nodes and the cloud exceeds 500 milliseconds, it automatically enables local caching of model parameters, allowing for continuous operation for up to 72 hours. The degradation mode adds a simplified operating state. When the CPU utilization exceeds 90% for 30 seconds, only the 8-dimensional core features are retained, and a moving average algorithm is used for demand forecasting.
[0065] The safety and fault-tolerant control submodule adds a collaborative diagnostic function for elevator groups. When multiple elevators malfunction simultaneously, a swarm intelligence decision-making algorithm is activated to re-divide the service floor intervals. In real-world testing on a super high-rise building, the system demonstrates improved peak-hour scheduling efficiency compared to the aforementioned embodiment, with abnormal state recovery time reduced to within 45 seconds in a 120-story building scenario.
[0066] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. An AI-based dynamic scheduling system for peak elevator passenger flow, characterized in that, include: The edge perception and feature extraction module is deployed on the edge computing nodes of each elevator group controller in the building. It is used to collect elevator operation status data, car load data, call request data of each floor, and queue statistics information obtained by visual sensors deployed in the waiting hall in real time. It also preprocesses and extracts features from the collected raw data to generate multi-dimensional feature vectors including time series features, spatial distribution features, and passenger flow intensity features. The cloud-based model collaborative optimization module runs on the central server. It receives feature vectors and historical scheduling results data uploaded by multiple edge nodes, builds a passenger flow prediction model based on a deep neural network, continuously optimizes its network parameters through periodic incremental learning, and distributes the optimized model parameters to each edge node in a differential update manner. The dynamic scheduling decision module, integrated into the edge node, is used to load a lightweight prediction model delivered from the cloud. Based on real-time feature vectors, it predicts the elevator demand within a specific future time window and generates a preliminary elevator dispatch plan according to the preset scheduling objective function. The system resource adaptive allocation module is used to monitor the computing load and memory usage of edge nodes in real time. When the load exceeds a set threshold, it dynamically adjusts the number of dimensions for feature extraction or the complexity level of the prediction model.
2. The AI-based elevator peak passenger flow dynamic scheduling system according to claim 1, characterized in that, The feature extraction process performed by the edge perception and feature extraction module includes: filtering and denoising the raw sensor data and removing outliers; Extract passenger flow count sequences at 5-minute intervals from the time dimension; The call frequency distribution of each floor is extracted from the spatial dimension; the rate of change of passenger flow in the current period relative to the same period in history is extracted from the intensity dimension; the above features are standardized and combined into a feature vector with fixed dimensions.
3. The AI-based elevator peak passenger flow dynamic scheduling system according to claim 1, characterized in that, The cloud-based model collaborative optimization module uses a hybrid architecture of long short-term memory network and self-attention mechanism in its deep neural network. The deep neural network takes multi-dimensional feature vectors uploaded from the edge as input, and after capturing time dependencies through three layers of long short-term memory network units, it uses a self-attention mechanism layer to weight and fuse the importance of different features, and finally outputs the predicted value of elevator demand for each floor in the next 15 minutes.
4. The AI-based elevator peak passenger flow dynamic scheduling system according to claim 3, characterized in that, The training process of the cloud-based model collaborative optimization module uses mean squared error as the loss function and performs an incremental learning round every 24 hours using the latest collected data. The updated model parameters are compressed and then sent to the edge.
5. The AI-based elevator peak passenger flow dynamic scheduling system according to claim 1, characterized in that, The scheduling objective function based on the dynamic scheduling decision module is a multi-objective optimization function, which simultaneously considers three optimization objectives: minimizing the average waiting time of passengers, minimizing the total energy consumption of the elevator system, and maximizing the load balance of each car.
6. The AI-based elevator peak passenger flow dynamic scheduling system according to claim 5, characterized in that, The dynamic scheduling decision module uses an improved genetic algorithm to iteratively optimize the initial ladder dispatching scheme.
7. The AI-based elevator peak passenger flow dynamic scheduling system according to claim 1, characterized in that, The system resource adaptive allocation module dynamically adjusts the system operation mode based on the real-time CPU utilization of the edge nodes. When the CPU utilization remains above 80% for 10 seconds, the module automatically reduces the dimension of the feature vector from 128 to 64 and switches to a lightweight prediction model with lower computational complexity. When the utilization drops below 60% and remains below 60% for 30 seconds, the system reverts to the computation mode of high-dimensional features and the complete model.
8. The AI-based elevator peak passenger flow dynamic scheduling system according to claim 1, characterized in that, The system also includes a safety and fault-tolerant control submodule, which is used to continuously monitor the operating status of the elevator group and the execution results of scheduling instructions; If an elevator fails to respond to dispatch instructions for two consecutive scheduling cycles, or if the actual load of the car deviates significantly from the predicted demand by more than 25%, the abnormal handling process will be triggered immediately, putting the elevator into manual mode and reassigning the call tasks to the floors it serves.
9. The AI-based elevator peak passenger flow dynamic scheduling system according to claim 1, characterized in that, The system operates within a hierarchical time-scale coordination framework; cloud-based model optimization is performed on a daily basis; edge passenger flow prediction is performed on a 5-minute basis; dynamic scheduling decisions are generated on a 15-second basis; and resource allocation and security monitoring are performed in real time on a second-by-second basis.
10. The AI-based elevator peak passenger flow dynamic scheduling system according to claim 2, characterized in that, The process of extracting the rate of change of passenger flow in the current period relative to the historical same period from the intensity dimension is as follows: obtain the passenger flow count value in the current period and the average passenger flow in the historical same period; calculate the difference between the passenger flow count value in the current period and the average passenger flow in the historical same period; divide the difference by the average passenger flow in the historical same period and multiply by 100% to obtain the rate of change.