Intelligent scheduling method and device for scattered human resources

Abstract

The application discloses a kind of scattered human resources intelligent scheduling method and equipment, it is related to human resources scheduling technical field, by scattered human resources intelligent scheduling method, it can realize real-time sensing, dynamic mapping, depth feature extraction, efficient candidate screening, global optimization solution, conflict resolution, scheme execution and the whole-link intelligent scheduling of continuous self-learning, significantly improve the response speed of scheduling, global cost, matching coverage, skill accuracy and concurrent scale, with dynamic environment self-adaptation and continuous self-optimization function, effectively solve the widely distributed scattered tasks and service providers, traditional manual scheduling mode cannot realize efficient matching, resulting in a large number of resources idle and task delay coexist, and simply rely on manually set rules, cannot adapt to complex scene change Problem, while further improving the efficiency, adaptability and expansibility of scattered human resources scheduling, can also reduce cost.

Description

Technical Field

[0001] This invention relates to the field of human resource scheduling technology, and in particular to a method and device for intelligent scheduling of scattered human resources. Background Technology

[0002] With the rapid development of the gig economy, flexible employment has become an important part of the labor market. The global gig economy market is growing at an average annual rate of over 20%. The scheduling and management of scattered human resources plays a key role in the gig economy. Through the scheduling and management of scattered human resources, real-time and accurate matching between massive tasks and massive service personnel can be achieved more efficiently, minimizing task response and completion time, reducing empty running and idle rates, increasing per capita output per unit time, and taking into account service quality and user experience. Ultimately, the core goal of maximizing platform operational efficiency and minimizing overall scheduling costs can be achieved.

[0003] However, the existing methods for scheduling and managing scattered human resources have many inconveniences. Scattered tasks and service providers are widely distributed, and traditional manual scheduling methods cannot achieve efficient matching, resulting in a large amount of idle resources and task delays. Furthermore, relying solely on manually set rules cannot adapt to changes in complex scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a method and device for intelligent scheduling of scattered human resources, in order to solve the many inconveniences of the existing scheduling and management methods for scattered human resources. Scattered tasks and service providers are widely distributed, and traditional manual scheduling methods cannot achieve efficient matching, resulting in a large amount of idle resources and task delays. Furthermore, relying solely on manually set rules cannot adapt to changes in complex scenarios.

[0005] To achieve the above objectives, this invention proposes an intelligent scheduling method for scattered human resources, characterized by the following steps: Real-time acquisition of multi-source heterogeneous data, and processing of the multi-source heterogeneous data; Based on the preprocessed multi-source heterogeneous data, a dynamic spatiotemporal graph is constructed, and the dynamic spatiotemporal graph is processed based on the spatiotemporal graph neural network to obtain a high-dimensional feature representation of each node; Based on the high-dimensional feature representation, the feature similarity between personnel nodes and task nodes is calculated, and the Top-K candidate personnel are selected for each task to form a candidate set. Based on the candidate set, a constrained combinatorial optimization model is constructed to minimize the global scheduling cost, and a preliminary supply and demand matching scheme is obtained by solving the model. The initial supply and demand matching scheme is subjected to conflict detection. If a conflict exists, it is resolved to obtain the final scheduling scheme and output it for execution.

[0006] Preferably, the real-time acquisition of multi-source heterogeneous data and the processing of the multi-source heterogeneous data specifically involves: performing data cleaning, normalization, and spatiotemporal alignment processing on the multi-source heterogeneous data. The multi-source heterogeneous data includes at least one of the following: task requirement information, service personnel status information, traffic condition data, and weather data.

[0007] Preferably, the dynamic spatiotemporal diagram is: .

[0008] in, This is the set of nodes at time step t, including personnel nodes and task nodes.

[0009] It is a set of edges, including spatially adjacent edges, temporally related edges, and semantically related edges.

[0010] It is an adjacency matrix.

[0011] Preferably, the spatial adjacent edge is determined based on whether the Euclidean distance between two spatial nodes is less than a preset distance threshold; The time-related edges are used to connect nodes of the same entity at adjacent time steps to capture temporal evolution features; The semantic association edge is determined based on whether the matching degree between personnel skills and task requirements exceeds a preset matching degree threshold.

[0012] Preferably, the process of processing the dynamic spatiotemporal graph based on the spatiotemporal graph neural network to obtain the high-dimensional feature representation of each node specifically involves: Spatial and temporal features are extracted from the dynamic spatiotemporal graph based on a spatiotemporal graph neural network, and the extracted multi-source features are fused through a multi-head attention mechanism to obtain a high-dimensional feature representation for each node.

[0013] Preferably, the spatial feature extraction employs a graph attention network, which aggregates neighborhood information by calculating the attention coefficients between nodes; The time feature extraction employs a gated temporal convolutional network, which extracts time series features through convolution operations and a gating mechanism. The multi-head attention mechanism is used to fuse spatial features, temporal features, traffic features, skill features, and weather features.

[0014] Preferably, the step of constructing a constrained combinatorial optimization model based on the candidate set, with the objective of minimizing the global scheduling cost, and solving for a preliminary supply-demand matching scheme specifically involves: For small-scale problems, mixed-integer linear programming (MILP) is used for exact solutions. For large-scale problems, a hybrid heuristic algorithm combining an improved Hungarian algorithm with ant colony optimization and simulated annealing is used to solve them.

[0015] Preferably, the step of performing conflict detection on the preliminary supply and demand matching scheme, and resolving conflicts if they exist, to obtain the final scheduling scheme and output the execution is specifically: resolving conflicts through priority preemption to obtain the final scheduling scheme; And / or, conflict resolution can be achieved through alternative personnel recommendation strategies to obtain the final scheduling scheme.

[0016] Preferably, it also includes: collecting actual execution result data, updating the parameters of the spatiotemporal graph neural network and the combined optimization model online, and realizing continuous automatic optimization.

[0017] In another aspect of the present invention, a device for intelligent scheduling of scattered human resources is proposed, comprising: Memory, used to store computer programs; The processor is used to execute the computer program stored in the memory to implement the steps in the method for intelligent scheduling of scattered human resources.

[0018] Compared with existing technologies, this invention has at least the following beneficial effects: Through the intelligent scheduling method for scattered human resources, it can achieve full-link intelligent scheduling, including real-time perception, dynamic mapping, deep feature extraction, efficient candidate screening, global optimization solution, conflict resolution, solution execution, and continuous self-learning. This significantly improves the scheduling response speed, global cost, matching coverage, skill accuracy, and concurrency scale. It also has the functions of dynamic environment adaptation and continuous self-optimization, effectively solving the problems that scattered tasks and service providers are widely distributed, and traditional manual scheduling methods cannot achieve efficient matching, resulting in a large amount of idle resources and task delays. Furthermore, relying solely on manually set rules cannot adapt to complex scene changes. While further improving the efficiency, adaptability, and scalability of scattered human resource scheduling, it can also reduce costs. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the intelligent scheduling method for scattered human resources in one embodiment of the present invention; Figure 2 This is a schematic diagram of the architecture of a scattered human resource intelligent scheduling system in one embodiment of the present invention; Figure 3 This is a schematic diagram of dynamic spatiotemporal graph construction in one embodiment of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will be described in detail below, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0021] Example 1 like Figure 1 As shown, this embodiment proposes an intelligent scheduling method for scattered human resources, including the following steps: Step 1: Collect multi-source heterogeneous data in real time, and perform data cleaning, normalization and spatiotemporal alignment on the multi-source heterogeneous data.

[0022] Specifically, the multi-source heterogeneous data includes at least one of the following: task requirement information, service personnel status information, traffic condition data, and weather data.

[0023] Specifically, task requirements information should include at least one of the following: task location, time window, skill requirements, and priority.

[0024] Service personnel status information includes at least one of the following: real-time location, speed and direction, skill tags, and online status.

[0025] Traffic data should include at least one of the following: average speed of road segment, congestion index, and travel time.

[0026] Weather data must include at least one of the following: temperature, wind speed, precipitation, and visibility.

[0027] Step 2: Based on the preprocessed multi-source heterogeneous data, construct a dynamic spatiotemporal graph, and process the dynamic spatiotemporal graph based on a spatiotemporal graph neural network (ST-GNN) to obtain a high-dimensional feature representation of each node.

[0028] like Figure 3 As shown, the dynamic spatiotemporal diagram is as follows: .

[0029] in, This is the set of nodes at time step t, including personnel nodes and task nodes.

[0030] It is a set of edges, including spatially adjacent edges, temporally related edges, and semantically related edges.

[0031] It is an adjacency matrix.

[0032] Specifically, personnel node characteristics include personnel GPS coordinates, speed, direction, unique skill encoding vector, online duration, and historical completion rate; task node characteristics include task GPS coordinates, time window, unique skill requirement encoding vector, and priority.

[0033] Among them, in the personnel node features, the personnel's GPS coordinates represent the spatial location of the service personnel, speed represents the speed of the personnel's movement, which affects the time to reach the task location, direction represents the direction of the personnel's movement, which is used to predict future location, skill one-hot encoded vector represents the various skills possessed by the personnel (such as delivery, maintenance, cleaning, etc.), with the same dimension as the number of skill types, online duration represents the time that the personnel have been continuously online for service, which is used to assess fatigue, and historical completion rate represents the proportion of past tasks completed by the personnel, reflecting their reliability.

[0034] The task node features include the task GPS coordinates representing the location of the task, the time window representing the task's constraints on service time, the unique skill requirement encoding vector representing the type of skills required by the task, and the priority representing the importance of the task, which is used for preemption decisions during conflict resolution.

[0035] Furthermore, the spatial adjacency edges in the edge set can be used to capture spatial proximity relationships. When the Euclidean distance between two nodes (including person-person, task-task, and person-task) is less than the preset distance threshold D_max (default is 5km), a spatial adjacency edge is established between them. This edge allows spatially close nodes to directly exchange information in the spatiotemporal graph neural network (ST-GNN), which helps to discover opportunities for nearby matching and ensures that the spatiotemporal graph neural network (ST-GNN) prioritizes close-range matching, thereby reducing traffic costs.

[0036] Temporal association edges can be used to capture temporal evolution features. For the same entity (the same person or the same task), a temporal association edge is established between the node at the current time step t and the node at the previous time step t-1. This edge enables the spatiotemporal graph neural network (ST-GNN) to perceive the changing trend of the entity's state over time, such as the movement trajectory of the person's location and the dynamic changes in the urgency of the task, thus avoiding making wrong decisions based on outdated information.

[0037] Semantic association edges are used to capture skill matching relationships. When the matching degree between a person's skill tag and the skill requirements of a task exceeds a preset matching degree threshold, a semantic association edge is established between the corresponding person node and task node. This edge makes the "person-task" pair with high skill matching degree have higher information transmission weight in the spatiotemporal graph neural network (ST-GNN), providing prior structural information for subsequent feature extraction and matching calculation, ensuring the accuracy of skill matching, and avoiding invalid matching of "someone but not able to do it".

[0038] The adjacency matrix is ​​generated based on the definitions of nodes and edges mentioned above, generating the adjacency matrix for the current time step t. , Let N×N be a symmetric or asymmetric matrix (N being the total number of nodes), where [i][j] represents the connection weight from node i to node j. For spatial adjacency edges and semantically related edges, the weight can be set as a function value of distance or matching degree. For temporally related edges, the weight can be set as a fixed value or a time decay factor.

[0039] Therefore, the problem of scattered human resource scheduling is modeled as a spatiotemporal graph neural network (ST-GNN), which can naturally and uniformly express multi-dimensional information such as personnel, tasks, spatial relationships, temporal evolution and semantic matching, thereby improving scheduling efficiency and matching rate.

[0040] Specifically, the process of processing the dynamic spatiotemporal graph based on the spatiotemporal graph neural network to obtain a high-dimensional feature representation of each node involves: extracting spatial and temporal features from the dynamic spatiotemporal graph based on the spatiotemporal graph neural network, and fusing the extracted multi-source features through a multi-head attention mechanism to obtain a high-dimensional feature representation of each node.

[0041] The spatial feature extraction employs a graph attention network (GAT), which aggregates neighborhood information by calculating the attention coefficients between nodes.

[0042] Specifically, for dynamic spatiotemporal graphs First, a Graph Attention Network (GAT) is used to extract features in the spatial dimension. The GAT allows nodes to assign different attention weights to different neighbors when aggregating neighborhood information, thereby focusing on more important neighbor nodes. The specific calculation process is as follows: For node i and its neighbor node j (j∈N(i)), the attention coefficient is calculated as follows: The formula for calculating the attention coefficient is: in, and Let be the input feature vectors of nodes i and j, respectively, and W be a learnable weight matrix used to perform a linear transformation on the input features. This is the transpose of the attention vector. This is the activation function, used to introduce nonlinearity.

[0043] Then, the attention coefficients are normalized using the Softmax function to obtain the final attention weights. :

[0044] Finally, the new feature of node i is represented as a weighted sum of the features of all its neighboring nodes:

[0045] in By using activation functions such as ELU and stacking multiple layers of graph attention networks (GAT), nodes can aggregate information from multiple neighborhoods and capture higher-order spatial structure features. This enables the spatiotemporal graph neural network (ST-GNN) model to aggregate information based on the actual degree of correlation between nodes (rather than fixed weights) and adapt to dynamically changing graph structures.

[0046] The temporal feature extraction employs a gated temporal convolutional network (GTCN), which extracts time-series features through convolution operations and a gating mechanism.

[0047] Specifically, the Gated Temporal Convolutional Network (GTCN) combines dilated causal convolution with a gating mechanism, enabling it to efficiently capture temporal dependencies over long periods while avoiding gradient vanishing in recurrent neural networks (RNNs). The specific computation process is as follows: For each node, take its feature sequence from the past k time steps. This is achieved through two parallel causal convolutional layers: Main convolutional layer output: , Gated convolutional layer output: , Then, the fusion is performed through a gating mechanism: , in: The function is a hyperbolic tangent activation function with an output range of (-1, 1). The Sigmoid activation function has an output range of (0,1) and is used as a gating signal.

[0048] This gating mechanism enables the Spatiotemporal Graph Neural Network (ST-GNN) model to adaptively select which historical information to retain or forget, thereby capturing temporal evolution patterns more accurately. The temporal feature extraction of the Gated Temporal Convolutional Network (GTCN) can efficiently capture long-term sequence dependencies, and the gating mechanism avoids gradient vanishing, making it suitable for processing high-frequency real-time data streams.

[0049] The multi-head attention mechanism is used to fuse spatial features, temporal features, traffic features, skill features, and weather features.

[0050] Specifically, after obtaining spatial and temporal features, it is necessary to integrate features from other dimensions (such as traffic conditions and weather) to form a comprehensive node representation. Multi-head attention is used to fuse multi-source features, taking spatial, temporal, traffic, skill, and weather features as different input sources. Multi-head attention calculates the interaction relationships between the source features, with each attention head independently calculating its attention weight. Finally, the outputs of all heads are concatenated or averaged to obtain the final fused feature vector.

[0051] The formula for calculating multi-head attention is: Each head is: Here, features from different sources are mapped to queries, keys, and values, respectively. The correlation weights between features from different sources are automatically learned through an attention mechanism to achieve adaptive fusion. This eliminates the need to manually set the fusion weights for each feature, thereby improving the generalization ability of the spatiotemporal graph neural network (ST-GNN) model.

[0052] Step 3: Based on the high-dimensional feature representation, calculate the feature similarity between personnel nodes and task nodes, and select Top-K candidate personnel for each task to form a candidate set.

[0053] Step 4: Based on the candidate set, construct a constrained combinatorial optimization model with the objective of minimizing the global scheduling cost, and solve for a preliminary supply and demand matching scheme.

[0054] Specifically, after feature extraction in the spatiotemporal graph neural network (ST-GNN) is completed, each person node and task node obtains a high-dimensional feature vector. In order to reduce the complexity of subsequent optimization, candidate set screening is first performed. For each task, the feature cosine similarity between it and all people is calculated. The top-K people are selected as candidates for the task according to the similarity from high to low (K takes the default value of 20).

[0055] The formula for calculating cosine similarity is: in, and These are the fused feature vectors for task nodes and personnel nodes, respectively.

[0056] By filtering the candidate set, the original full matching space of "all people × all tasks" is compressed into a sparse matching space of "each task × Top-K candidates", which greatly reduces the subsequent optimization problems.

[0057] Furthermore, based on the candidate set, a constrained combinatorial optimization model is constructed to find the matching scheme that minimizes the global scheduling cost.

[0058] Among them, the decision variables are: , indicating whether to assign person i to task j.

[0059] Objective function (minimize global scheduling cost):

[0060] in, The distance cost is calculated based on the shortest path distance from the current location of person i to the location of task j. The time cost is calculated based on the deviation between the estimated arrival time of person i for task j and the time window of task j. The skill deviation cost is calculated based on the difference between the skill vector of person i and the skill requirement vector of task j. The load balancing cost is calculated based on the number of tasks currently assigned to person i, encouraging balanced allocation.

[0061] , , , These are the weighting coefficients for each cost item, which can be dynamically adjusted according to actual business needs.

[0062] Furthermore, the constraints include task coverage constraints, personnel capacity constraints, time window constraints, skill matching constraints, and service radius constraints.

[0063] The task coverage constraint means that each task can be assigned to at most one person, or exactly one person, depending on the business rules.

[0064] Personnel capacity constraint means that each person can only undertake a maximum of [number] tasks. One task, This is the maximum concurrent task capacity for personnel.

[0065] The time window constraint means that the time when person i arrives at task j must be within the time window of task j. Within.

[0066] The skill matching constraint means that the skill label of person i must cover the skill requirements of task j, that is, the skill deviation cost is lower than the allowable threshold.

[0067] The service radius constraint means that the distance from person i to task j must not exceed the maximum service radius. .

[0068] Step 5: Perform conflict detection on the preliminary supply and demand matching scheme. If a conflict exists, resolve the conflict to obtain the final scheduling scheme and output it for execution.

[0069] Specifically, conflict detection mainly includes checking whether a person has been assigned to multiple tasks and exceeds their capacity limit, checking whether a task has been assigned to multiple people, and checking whether a task has not been covered by any person.

[0070] Furthermore, when multiple tasks compete for the same person, the task with the higher priority is given priority. The task that is preempted enters the queue for reallocation. For tasks that are preempted or not covered, the second-best available person is selected from its candidate set to reassign the task. The above process is repeated until all conflicts are resolved or the maximum number of retries is reached.

[0071] In one embodiment, the step of constructing a constrained combinatorial optimization model based on the candidate set, with the objective of minimizing the global scheduling cost, and solving for a preliminary supply-demand matching scheme specifically involves: For small-scale problems, Mixed Integer Linear Programming (MILP) is used for exact solutions, and branch-and-bound algorithms or commercial solvers (such as Gurobi and CPLEX) are used to obtain the global optimal solution.

[0072] In one embodiment, the step of constructing a constrained combinatorial optimization model based on the candidate set, with the objective of minimizing the global scheduling cost, and solving for a preliminary supply-demand matching scheme specifically involves: For large-scale problems, a hybrid heuristic algorithm combining an improved Hungarian algorithm with ant colony optimization and simulated annealing is used to solve them.

[0073] Specifically, the hybrid heuristic algorithm is as follows: First, an improved Hungarian algorithm is used to quickly generate a feasible initial solution on the candidate set. Then, the Ant Colony Optimization (ACO) algorithm is used to perform a global search in the solution space. Finally, the Simulated Annealing (SA) algorithm is used to perform a local fine search based on the results of the Ant Colony Optimization, so as to avoid getting trapped in local optima.

[0074] In one embodiment, actual execution result data is collected, and the parameters of the spatiotemporal graph neural network and the combined optimization model are updated online to achieve continuous automatic optimization.

[0075] Specifically, the optimization method is as follows: for features that change slowly (such as personnel skill tags, task skill requirements, road network static attributes, etc.), Redis caching is used for storage. When the ST-GNN needs these features, it will first read them from the cache and only load them from the database when the cache is not hit. This can effectively reduce the overhead of repeated database access and feature recalculation, and reduce the system load.

[0076] Secondly, in continuous time steps, changes in the dynamic spatiotemporal graph are usually local (such as updates to a small number of node positions or additions of a small number of tasks) rather than global reconstruction. This embodiment adopts an incremental update strategy, which only updates the nodes that have changed and their neighborhoods locally, while maintaining the features of the unchanged parts of the graph. This avoids the huge overhead of global recalculation and is suitable for high-frequency real-time scheduling scenarios.

[0077] For all scheduling requests arriving within a short time window, such as 1 second, a batch processing approach is adopted to merge the matching requests of multiple tasks into a single batch matching calculation, rather than solving each task independently. Batch processing can fully utilize the vectorization advantage of matrix operations, increasing the system throughput by about 3 times. Batch processing parallelizes serial tasks, significantly improving the system throughput.

[0078] Furthermore, since spatiotemporal graph neural network (ST-GNN) models typically have a large number of parameters and high inference latency, knowledge distillation techniques can be used to transfer the knowledge of the trained large model to a lightweight small model. The number of parameters in the lightweight small model is about 1 / 10 of that in the large model, which significantly improves the inference speed. At the same time, the matching accuracy loss is controlled within an acceptable range (≤2%), which is beneficial for deployment at resource-constrained edge devices.

[0079] Furthermore, since spatiotemporal feature extraction and graph neural network inference computation have high parallelism, the feature extraction and graph attention computation of spatiotemporal graph neural networks (ST-GNN) can be deployed on GPUs to perform these computations. By leveraging the parallel computing capabilities of CUDA, matrix multiplication and attention computation can be accelerated, meeting the millisecond-level response requirements.

[0080] In summary, this embodiment achieves end-to-end intelligent scheduling through the above steps, including real-time perception, dynamic mapping, deep feature extraction, efficient candidate selection, global optimization, conflict resolution, scheme execution, and continuous self-learning. Compared with existing technologies, this invention significantly improves response speed, global cost, matching coverage, skill accuracy, and concurrency scale. It also features dynamic environment adaptation and continuous self-optimization, effectively solving the problems of scattered tasks and widely distributed service providers, the inability of traditional manual scheduling methods to achieve efficient matching, resulting in a large amount of idle resources and task delays, and the inability to adapt to complex scene changes by simply relying on manually set rules. This invention further improves the efficiency, adaptability, and scalability of scattered human resource scheduling while also reducing costs.

[0081] To facilitate understanding of the embodiments of this solution by those skilled in the art, the working principle of this solution will now be briefly explained in conjunction with specific application scenarios: Taking the real-time delivery scheduling scenario of food delivery as an example: In the urban delivery area of ​​a certain food delivery platform, there are currently multiple online riders and multiple orders to be delivered (including those that have been ordered but not picked up and those that have been picked up but not delivered). The system needs to assign a suitable rider to each order in real time.

[0082] The rider's app reports their location (GPS coordinates), speed, direction, and number of orders at regular intervals. The merchant and user apps obtain order information (pickup address, delivery address, estimated food preparation time, and expected delivery time). Real-time traffic data and estimated travel time are obtained through third-party map services, and the current weather conditions in the area are obtained through weather services.

[0083] Construct a dynamic spatiotemporal graph containing multiple rider nodes and multiple order nodes. Establish spatial adjacency edges based on the real-time distance between riders and orders, establish temporal association edges based on the rider's position sequence at each time step, and establish semantic association edges based on the rider's delivery skills (such as whether they have an insulated box or are familiar with the area) and order requirements.

[0084] The spatial features of riders and orders are extracted using a graph attention network (GAT) (such as the distance between riders and pickup points, and the relative position distribution among riders). The evolutionary features of rider position changes and order waiting time are extracted using a gated temporal convolutional network (GTCN). Spatial features, temporal features, road condition features, and weather features are fused using multi-head attention to output a 128-dimensional fused feature vector for each rider and order.

[0085] For each order, calculate the cosine similarity of features with all riders, and select the Top-15 riders as the candidate set.

[0086] An optimization model is constructed with the goal of minimizing the weighted average delivery time and the number of overdue orders. Constraints include: the maximum number of orders each rider can carry at the same time, the order delivery time must not be later than the expected delivery time by a certain number of minutes, and riders must not exceed the delivery area. An improved Hungarian algorithm combined with simulated annealing is used to solve for the optimal matching scheme.

[0087] Check if any riders have been assigned an excessive number of orders or if any orders have not been covered. Prioritize riders with excessive orders, reallocate low-priority orders, and select replacement riders from the candidate set for orders that have not been covered.

[0088] Output the optimal rider allocation plan, estimated pickup time and estimated delivery time for each order, plan the optimal pickup and delivery route for each rider that includes all orders (considering pickup order and delivery order), issue dispatch instructions via push notification, and display the navigation route on the rider's APP.

[0089] Collect execution data such as actual delivery time, actual arrival time, and rider feedback. Fine-tune the parameters of the Spatiotemporal Graph Neural Network (ST-GNN) model and optimize the model weight coefficients online.

[0090] Example 2 like Figure 2As shown, this implementation is a fragmented human resource intelligent scheduling system proposed based on implementation one. It adopts a five-layer architecture design, mainly including the client layer, gateway layer, business service layer, algorithm engine layer and data layer.

[0091] The client layer serves as the entry point for interaction between the system and end users, and includes: Demand-side APP: Used by task publishers to create tasks, view task status, and receive scheduling result notifications.

[0092] Service provider app: Used by service personnel, such as delivery riders or freelancers, to report their location, receive task assignments, and view navigation routes.

[0093] Management Web: Used by operations and management personnel to view global scheduling status, configure scheduling parameters, and handle exceptions.

[0094] Web-based operations and maintenance (O&M) platform: Used by system O&M personnel to monitor system operation status, view logs, and perform O&M operations.

[0095] Open API: Allows third-party systems to integrate and call, enabling the external access to scheduling capabilities.

[0096] The gateway layer uses an API gateway to uniformly manage all external requests, and its main functions include: Load balancing: Distributes requests to multiple business service instances to improve system throughput.

[0097] Rate limiting and circuit breaking: Provides rate limiting and circuit breaking protection during peak traffic periods or service anomalies to prevent system cascading failure.

[0098] Authentication and authorization: Verify the identity and permissions of requests to ensure system security.

[0099] Protocol conversion: Supports conversion of multiple protocols such as HTTP, WebSocket, and gRPC to adapt to different clients.

[0100] The business service layer adopts a microservice architecture, with each service deployed and scaled independently, including: Task Service: Responsible for the lifecycle management of tasks (creation, update, cancellation, completion).

[0101] Personnel Services: Responsible for the registration, status management, and skill tag maintenance of service personnel.

[0102] Matching service: responsible for calling the algorithm engine to perform matching calculations and managing the matching results.

[0103] Scheduling service: Responsible for converting matching results into scheduling instructions and coordinating task allocation.

[0104] Notification service: Responsible for sending notifications to users via push notifications, SMS, email, etc.

[0105] Reporting services: Responsible for generating various operational reports and data statistics.

[0106] Alarm service: Responsible for monitoring abnormal system indicators and triggering alarms.

[0107] Configuration service: Responsible for centrally managing the dynamic configuration parameters of the system.

[0108] The algorithm engine layer is the core decision-making layer of the system, and includes: Spatiotemporal Feature Extraction Engine: Based on the Spatiotemporal Graph Neural Network (ST-GNN) model, it extracts high-dimensional features of people and tasks from dynamic spatiotemporal graphs.

[0109] Supply and demand matching optimization engine: Based on MILP or hybrid heuristic algorithms, it solves the optimal matching scheme.

[0110] Route planning engine: Plans the optimal navigation route for matched users, taking into account real-time traffic conditions.

[0111] Prediction engine: Used to predict task demand and personnel distribution in future periods to assist in pre-scheduling.

[0112] The data layer is responsible for persistent storage and real-time data transfer, including: MySQL: Stores structured business data (task information, personnel information, matching records, etc.).

[0113] PostGIS: Stores spatial geographic data and supports spatial indexing and spatial querying.

[0114] Redis: As a cache database, it stores frequently accessed data (such as real-time location of people and feature vectors).

[0115] Kafka: As a message queue, it enables asynchronous data flow and peak smoothing.

[0116] ElasticSearch: As a search engine, it supports log retrieval and data analysis.

[0117] Specifically, the client layer accesses the business service layer through the gateway layer, the business service layer calls the algorithm engine layer to complete intelligent decision-making, and the data layer provides data persistence and caching support for the upper layer. The five-layer architecture realizes the separation of interface display, business logic, algorithm decision-making and data storage, which facilitates the independent evolution and expansion of each layer.

[0118] Example 3 This embodiment proposes an intelligent scheduling device for scattered human resources based on Embodiment 1, including: Memory is used to store computer programs.

[0119] The processor is used to execute the computer program stored in the memory to implement the steps of the method for intelligent scheduling of scattered human resources.

[0120] Specifically, the memory is used to store computer programs and related data. It can be non-volatile memory (such as NAND Flash, SSD) or volatile memory (such as DDR4 DRAM), with a storage capacity of not less than 64GB to meet the storage requirements of large-scale graph data and model parameters.

[0121] Processor: Used to execute computer programs stored in the memory, and may employ a heterogeneous computing architecture of multi-core CPU (such as Intel Xeon or AMD EPYC series) and GPU (such as NVIDIA A100 or V100), where the CPU is responsible for business logic and data preprocessing, and the GPU is responsible for parallel inference computing of spatiotemporal graph neural network (ST-GNN).

[0122] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for intelligent scheduling of scattered human resources, characterized in that, Includes the following steps: Real-time acquisition of multi-source heterogeneous data, and processing of the multi-source heterogeneous data; Based on the preprocessed multi-source heterogeneous data, a dynamic spatiotemporal graph is constructed, and the dynamic spatiotemporal graph is processed based on the spatiotemporal graph neural network to obtain a high-dimensional feature representation of each node; Based on the high-dimensional feature representation, the feature similarity between personnel nodes and task nodes is calculated, and the Top-K candidate personnel are selected for each task to form a candidate set. Based on the candidate set, a constrained combinatorial optimization model is constructed to minimize the global scheduling cost, and a preliminary supply and demand matching scheme is obtained by solving the model. The initial supply and demand matching scheme is subjected to conflict detection. If a conflict exists, it is resolved to obtain the final scheduling scheme and output it for execution.

2. The method for intelligent scheduling of scattered human resources according to claim 1, characterized in that, The real-time acquisition of multi-source heterogeneous data and the processing of the multi-source heterogeneous data specifically involve: performing data cleaning, normalization, and spatiotemporal alignment processing on the multi-source heterogeneous data. The multi-source heterogeneous data includes at least one of the following: task requirement information, service personnel status information, traffic condition data, and weather data.

3. The method for intelligent scheduling of scattered human resources according to claim 1, characterized in that, The dynamic spatiotemporal diagram is as follows: ; in, This is the set of nodes at time step t, containing personnel nodes and task nodes; It is a set of edges, including spatially adjacent edges, temporally related edges, and semantically related edges; It is an adjacency matrix.

4. The method for intelligent scheduling of scattered human resources according to claim 3, characterized in that, The spatial adjacent edges are determined based on whether the Euclidean distance between two spatial nodes is less than a preset distance threshold. The time-related edges are used to connect nodes of the same entity at adjacent time steps to capture temporal evolution features; The semantic association edge is determined based on whether the matching degree between personnel skills and task requirements exceeds a preset matching degree threshold.

5. The method for intelligent scheduling of scattered human resources according to claim 1, characterized in that, The process of processing the dynamic spatiotemporal graph based on the spatiotemporal graph neural network to obtain the high-dimensional feature representation of each node is specifically as follows: Spatial and temporal features are extracted from the dynamic spatiotemporal graph based on a spatiotemporal graph neural network, and the extracted multi-source features are fused through a multi-head attention mechanism to obtain a high-dimensional feature representation for each node.

6. The method for intelligent scheduling of scattered human resources according to claim 5, characterized in that, The spatial feature extraction employs a graph attention network, which aggregates neighborhood information by calculating the attention coefficients between nodes; The time feature extraction employs a gated temporal convolutional network, which extracts time series features through convolution operations and a gating mechanism. The multi-head attention mechanism is used to fuse spatial features, temporal features, traffic features, skill features, and weather features.

7. The method for intelligent scheduling of scattered human resources according to claim 1, characterized in that, The specific steps for constructing a constrained combinatorial optimization model based on the candidate set, aiming to minimize the global scheduling cost, and solving for the preliminary supply-demand matching scheme are as follows: For small-scale problems, mixed-integer linear programming (MILP) is used for exact solutions. For large-scale problems, a hybrid heuristic algorithm combining an improved Hungarian algorithm with ant colony optimization and simulated annealing is used to solve them.

8. The method for intelligent scheduling of scattered human resources according to claim 1, characterized in that, The process of performing conflict detection on the preliminary supply and demand matching scheme, resolving conflicts if they exist, and obtaining the final scheduling scheme and outputting the execution details are as follows: conflict resolution is performed through priority preemption to obtain the final scheduling scheme. And / or, conflict resolution can be achieved through alternative personnel recommendation strategies to obtain the final scheduling scheme.

9. The method for intelligent scheduling of scattered human resources according to claim 1, characterized in that, Also includes: Collect actual execution result data and update the parameters of the spatiotemporal graph neural network and the combined optimization model online to achieve continuous automatic optimization.

10. A device for intelligent scheduling of scattered human resources, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program stored in the memory for implementing the steps of the method as described in any one of claims 1-9.

Images