An automatic order dispatching recommendation method based on deep learning

Abstract

The application discloses an automatic dispatching recommendation method based on deep learning, and relates to the technical field of intelligent scheduling, comprising the following steps: collecting to-be-dispatched task data, execution resource data, regional operation state and corresponding dispatching element data, and performing space-time correlation matching to generate a dispatching data set; constructing a dispatching element node according to the dispatching element data in the dispatching data set, and establishing a node correlation edge according to the space-time matching relationship in the dispatching data set to form a dispatching collaborative graph; inputting the dispatching collaborative graph into a pre-constructed dispatching collaborative model to perform node embedding, edge weight calculation and attention fusion on the dispatching element node to generate a dispatching collaborative representation; and the space-time relationship among the to-be-dispatched task, the execution resource and the regional operation state is described through the dispatching collaborative representation, and the joint measurement of the adaptation, the timeliness, the risk, the quality and the regional balance is realized through the performance image and the recommendation score, so that the dispatching accuracy, the performance stability and the resource balance are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling technology, and in particular to an automatic order dispatch recommendation method based on deep learning. Background Technology

[0002] With the development of on-demand delivery, door-to-door services, on-demand transportation capacity, emergency operation and maintenance, and regional business scheduling, automatic order dispatch technology is gradually evolving from manual rule-based order dispatch to data-driven order dispatch. Related technologies typically collect information such as task location, resource location, service capacity, task time limit, and regional load. They use distance priority, idle priority, capacity tag matching, rule scoring, or machine learning models to calculate the matching relationship between the task to be dispatched and the available execution resources, and output the candidate resource ranking results to improve the automation level and scheduling response efficiency in the task allocation process.

[0003] In related automatic task dispatching technologies, task, resource, and regional status are often used as independent fields in calculations. Spatiotemporal relationships are usually limited to distance filtering or time window matching. In scenarios with multi-regional resource flow, dense task releases, and rapid changes in resource status, there is a lack of unified expression of the relationship between task and resource spatial reachability, time availability, and regional capacity. At the same time, some task dispatching recommendation processes focus more on single-point matching scores between tasks and resources, lacking continuous prediction and joint characterization of factors such as performance adaptation, completion timeliness, anomaly risk, service quality, and regional balance. In scenarios with high-concurrency task dispatching or regional load fluctuations, the recommendation results are difficult to stably reflect the comprehensive performance capability of the execution resources for dispatched tasks. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based automatic order dispatch recommendation method to solve the problem of inaccurate performance matching judgment in automatic order dispatch recommendation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an automatic order dispatch recommendation method based on deep learning, which includes,

[0008] Collect data on tasks to be dispatched, execution resources, regional operation status, and corresponding dispatch elements, and perform spatiotemporal correlation matching to generate a dispatch dataset.

[0009] Based on the dispatch element data in the dispatch dataset, dispatch element nodes are constructed, and node association edges are established according to the spatiotemporal matching relationship in the dispatch dataset to form a dispatch collaboration graph.

[0010] Input the dispatch collaboration graph into the pre-built dispatch collaboration model, and perform node embedding, edge weight calculation and attention fusion on the dispatch element nodes to generate a dispatch collaboration representation.

[0011] The collaborative recall score of the execution resource data is calculated based on the dispatch collaborative characterization, the resource availability is verified in combination with the execution resource data, and the number of resources recalled is limited according to the regional operation status to generate an execution resource set.

[0012] Input the execution resource set and the dispatch collaboration representation into the dispatch collaboration model, calculate the performance fit value and predict the completion time value, abnormal risk value and service quality value, and map the profile tags according to the preset level range to generate the dispatch performance profile;

[0013] After performing attribution verification and dimensional normalization on the dispatch fulfillment profile, the regional balance factor is calculated in combination with the regional operation status. Then, the fulfillment gain, risk attenuation, quality enhancement and regional balance adjustment are calculated to generate dispatch recommendation scores.

[0014] The execution resource set is recommended and ranked according to the dispatch recommendation score, and the target execution resource is automatically determined based on the recommendation ranking result to generate automatic dispatch recommendation results.

[0015] As the deep learning-based automatic task dispatch recommendation method of the present invention, the task data to be dispatched includes task identifier, task location, task release time and expected completion time.

[0016] The execution resource data includes execution resource identifier, real-time location, service capability tag, serviceable area, current task status, and available service time period;

[0017] The regional operating status includes the regional identifier, the amount of available resources in the region, and the regional resource load rate.

[0018] The dispatch data includes task resource distance, task time constraint, resource capacity matching, and regional load.

[0019] The automatic order dispatch recommendation method based on deep learning described in this invention is characterized in that: the generation of the order dispatch dataset specifically includes,

[0020] Based on the task location, real-time location, and area identifier, spatial correlation matching is performed on the task data to be dispatched, the execution resource data, and the area operation status to form a spatial matching result.

[0021] Based on the task release time, expected completion time, current task status, and available service time, time association matching is performed on the task data to be dispatched and the execution resource data to form a time matching result.

[0022] The spatial matching results and temporal matching results are linked and integrated according to the task identifier, execution resource identifier, and region identifier to generate a dispatch dataset.

[0023] The automatic order dispatch recommendation method based on deep learning described in this invention is characterized in that: the formation of the order dispatch collaboration graph specifically includes,

[0024] Extract task identifiers, execution resource identifiers, and region identifiers from the dispatch dataset, and construct task element nodes, resource element nodes, and region element nodes to form a dispatch element node set.

[0025] According to the node type, retrieve the task data to be dispatched, the execution resource data, and the regional operation status from the dispatch dataset, configure the attributes of the dispatch element node set, and form a node attribute table.

[0026] Based on the spatial matching results and the temporal matching results, the spatiotemporal relationships between task element nodes, resource element nodes and regional element nodes are determined, and a node relationship table is formed.

[0027] Configure task resource distance elements, task time constraint elements, resource capacity matching elements, and regional load elements according to the node relationship table to form a node association edge set;

[0028] The task dispatch element node set, node attribute table, and node associated edge set are encapsulated into a graph structure to form a task dispatch collaboration graph.

[0029] As the deep learning-based automatic order dispatch recommendation method of the present invention, its characteristic is that: the generation of order dispatch collaborative representation specifically includes,

[0030] The node attribute table is processed by node type labeling, feature encoding, and scale normalization to form a node encoding feature table;

[0031] Input the node encoding feature table into the dispatch collaboration model, and perform same-dimensional vector mapping on task element nodes, resource element nodes and regional element nodes to form a node embedding matrix.

[0032] According to the edge type encoding in the node association edge set, the adjacent nodes in the node embedding matrix are classified and edge weights are calculated to form an edge weight coefficient table;

[0033] Attention aggregation is performed on the node embedding matrix according to the edge weight coefficient table to form the node collaborative representation matrix;

[0034] The node collaborative representation matrix is ​​processed by aligning node numbers and unifying vector dimensions to generate a dispatch collaborative representation.

[0035] The automatic order dispatch recommendation method based on deep learning described in this invention is characterized in that: the generation of the execution resource set specifically includes,

[0036] The dispatching collaboration representation is split according to the node type of task element nodes, resource element nodes and regional element nodes. The collaboration vector is extracted and collected according to the node identifier to form a recall vector table.

[0037] Availability checks are performed on the service capability tags, serviceable areas, current task status, and available service time periods associated with the recall vector table to obtain the recallable resource table.

[0038] By using the task resource distance element, task time constraint element, resource capacity matching element and regional load element corresponding to the recallable resource table, the execution resources are collaboratively recalled and scored to obtain the resource recall scoring table.

[0039] According to the resource recall scoring table, the execution resource identifiers associated with the same task identifier are sorted in descending order, and execution resource identifiers below the preset recall threshold are removed to obtain the resource candidate ranking table.

[0040] Based on the resource candidate sorting table and the available resources in the region, extract the execution resource identifiers and generate an execution resource set.

[0041] As the deep learning-based automatic order dispatch recommendation method of the present invention, its characteristic is that: the fulfillment adaptation calculation of execution resources specifically includes,

[0042] Based on the same task identifier in the execution resource set, the dispatch coordination representation is retrieved, and the coordination vectors of task element nodes, resource element nodes and regional element nodes are paired to form a performance adaptation vector group.

[0043] Retrieve the task resource distance elements, task time constraint elements, resource capability matching elements, and regional load elements associated with the performance adaptation vector group, and perform normalization and fusion processing to form a performance adaptation feature group.

[0044] Perform capability consistency verification, spatiotemporal reachability verification, and regional load constraint verification on the performance adaptation feature group to form the performance constraint verification result;

[0045] The performance adaptation feature group is input into the dispatch collaboration model to calculate the performance adaptation prediction value, and the performance adaptation prediction value is constrained and filtered according to the performance constraint verification result to form the performance adaptation result.

[0046] As the deep learning-based automatic order dispatch recommendation method of the present invention, its characteristic is that: the generation of the order dispatch fulfillment profile specifically includes,

[0047] Match the performance adaptation results with the dispatch dataset according to the task identifier and execution resource identifier, and normalize the matching content to form a performance prediction feature table.

[0048] The order dispatching collaboration model is used to calculate the completion time value, abnormal risk value and service quality value corresponding to the performance prediction feature table to form the performance prediction result.

[0049] Based on the task identifier and execution resource identifier, the performance adaptation results and performance prediction results are merged and profile tags are mapped to generate a dispatch performance profile.

[0050] As the deep learning-based automatic order dispatch recommendation method of the present invention, its characteristic is that: the generation of order dispatch recommendation scores specifically includes,

[0051] The fulfillment matching value, completion timeliness value, abnormal risk value and service quality value in the order fulfillment profile are subjected to attribution verification and unit normalization processing to form a profile scoring table.

[0052] The regional identifiers in the portrait scoring table are matched with the regional operating status and the proportions are normalized to form a regional balance factor table.

[0053] The performance gain, risk reduction, quality enhancement and regional balance adjustment are calculated on the profile scoring table and the regional balance factor table to form the order recommendation scoring table.

[0054] The execution resource identifiers in the dispatch recommendation scoring table are scored and sorted according to the same task identifier to generate a dispatch recommendation ranking table.

[0055] The automatic order dispatch recommendation method based on deep learning described in this invention is characterized in that: the generation of automatic order dispatch recommendation results specifically includes,

[0056] The execution resource identifiers under the same task identifier in the dispatch recommendation scoring table are deduplicated, effective scores are filtered, and profile tags are matched to form a recommendation candidate table.

[0057] The execution resource identifiers in the recommended candidate table are sorted in descending order according to the dispatch recommendation score, and the score difference between adjacent execution resources is calculated to form a recommendation ranking table;

[0058] Confidence checks are performed on the score difference, completion time value, abnormal risk value, service quality value, regional available resources and regional resource load rate in the recommendation ranking table to form a target confirmation table;

[0059] Select the execution resource identifier that meets the confirmation threshold and is ranked first in the target confirmation table to form the target resource table;

[0060] The task identifier, target execution resource identifier, dispatch recommendation score, and profile tag in the target resource table are encapsulated into dispatch instructions to generate automatic dispatch recommendation results.

[0061] The beneficial effects of this invention are as follows: By generating a dispatch dataset, constructing a dispatch collaboration graph, and generating a dispatch collaboration representation, a unified expression is formed for the spatiotemporal, capability, and load relationships between tasks to be dispatched, execution resources, and regional operating status. Resource recall can simultaneously consider spatial reachability, temporal availability, capability adaptability, and regional carrying capacity, improving the accuracy and stability of execution resource selection. Through performance adaptation calculation, completion time prediction, anomaly risk prediction, service quality prediction, dispatch performance profile generation, and dispatch recommendation scoring, the automatic dispatch recommendation results can comprehensively reflect task acceptance capacity, performance reliability, risk controllability, service quality level, and regional balance, improving task execution efficiency, scheduling stability, and regional resource utilization. Attached Figure Description

[0062] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a flowchart of a deep learning-based automatic order dispatch recommendation method.

[0064] Figure 2 A flowchart for generating the dispatch dataset and dispatch collaboration graph.

[0065] Figure 3 A flowchart for generating a dispatch collaboration representation.

[0066] Figure 4 A flowchart for generating automatic order dispatch recommendation results. Detailed Implementation

[0067] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0068] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0069] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0070] Reference Figures 1-4 This is one embodiment of the present invention, which provides an automatic order dispatch recommendation method based on deep learning, including the following steps:

[0071] S1: Collect data on tasks to be dispatched, execution resources, regional operation status, and corresponding dispatch elements, and perform spatiotemporal correlation matching to generate a dispatch dataset.

[0072] It should be noted that the data for tasks to be dispatched is collected from the task submission records and the dispatch queue generated by the task issuing end; the execution resource data is collected from the location reporting records, status reporting records, and service capability registration records generated by the execution resource end; the regional operation status is collected from the regional scheduling records; and the dispatch element data is collected from the dispatch element records generated by the task issuing end, the execution resource end, and the regional scheduling end. The dispatch cycle starts with the generation of a task submission record by the task issuing end and entry into the dispatch queue, and ends with the generation of the automatic dispatch recommendation result or the expiration of the preset scheduling refresh time. The generation of task submission records and the refresh of regional scheduling records are all part of the process. New or reported changes in the status of execution resources serve as trigger conditions for the dispatch cycle. Based on the association range of task identifier, execution resource identifier, and region identifier, corresponding content is synchronously collected from task submission records generated by the task issuing end, location reporting records, status reporting records, and service capability registration records generated by the execution resource end, as well as regional scheduling records. The collected content is then marked with the collection time and source, enabling the same task to be dispatched, the corresponding execution resource, and the operating status of the region to establish a continuous correspondence within the dispatch cycle with a clear start point, end point, and trigger conditions, forming a collection basis that can be used for subsequent spatiotemporal correlation matching.

[0073] It should also be noted that the scheduling refresh duration is set based on the historical update interval of task submission records, execution resource status reporting records, and regional scheduling records, as well as the allowed dispatch response delay. It is used to end the current dispatch cycle and re-collect the data of tasks to be dispatched, execution resource data, regional operation status, and dispatch element data when no new dispatch cycle triggering conditions are generated, so as to avoid the execution resource status or regional operation status not being updated for a long time; the value is 30 seconds.

[0074] S1.1: Based on the task location, real-time location, and area identifier, perform spatial correlation matching on the task data to be dispatched, the execution resource data, and the area operation status to form a spatial matching result.

[0075] It should be noted that spatial correlation matching is carried out according to task location, real-time location, and region identifier. The coordinate reference system, coordinate axis order, and unit of measurement used by the task location, real-time location, and region scheduling record are compared. When the coordinate format is consistent, the coordinate fields are directly mapped. When the coordinate formats are inconsistent, coordinate transformation is used to uniformly convert the task location and real-time location to the coordinate field format used by the region scheduling record. When the task location is latitude and longitude coordinates, the task location coordinates are directly extracted. When the real-time location is the positioning coordinates reported by the execution resource end, the real-time location coordinates are directly extracted. According to the correspondence between the region identifier and the coordinate range in the region scheduling record, the task location coordinates and real-time location coordinates are matched to the corresponding region identifiers respectively. The task location is mapped to the region identifier in the region operation status, and the real-time location is mapped to the region identifier in the region operation status. The spatial correspondence is retained for the task data to be dispatched and the execution resource data with consistent region identifiers. The spatial correspondence is canceled for the task data to be dispatched and the execution resource data with inconsistent region identifiers. Then, the retained spatial correspondence is associated and organized according to the task identifier, execution resource identifier, and region identifier to form the spatial matching result.

[0076] S1.2: Based on the task release time, expected completion time, current task status, and available service time period, perform time correlation matching on the task data to be dispatched and the execution resource data to form a time matching result.

[0077] It should be noted that time-related matching is performed based on task release time, expected completion time, current task status, and available service time period. The task release time and expected completion time in the data of tasks to be dispatched are organized into task time ranges, and the current task status and available service time period in the execution resource data are organized into resource time ranges. The time overlap between the task time range and the resource time range is compared. When the current task status is idle or in execution and the available service time period covers the task time range, the corresponding execution resource data is retained. When the current task status is unavailable, offline, or the available service time period does not cover the task time range, the corresponding execution resource data is filtered out. The retained time correspondences are then linked and organized according to the task identifier and the execution resource identifier to form the time matching result.

[0078] It should also be noted that the expected completion time is determined by the task publisher based on the task publication time and the allowed completion time set in the task submission record, and is collected along with the data of the tasks to be dispatched.

[0079] S1.3: The spatial matching results and temporal matching results are associated and integrated according to the task identifier, execution resource identifier and region identifier to generate a dispatch dataset.

[0080] It should be noted that the content with the same task identifier and execution resource identifier in the spatial matching results is compared with the content with the same task identifier and execution resource identifier in the temporal matching results. The content that satisfies both spatial and temporal correlation matching and can be matched with the region identifier is retained. The content with inconsistent task identifiers, inconsistent execution resource identifiers, missing region identifiers, or only satisfying spatial correlation matching or only satisfying temporal correlation matching is removed. The retained content is then merged and organized with the task data to be dispatched, execution resource data, regional operation status, and corresponding dispatch element data, so that the execution resource identifier under the same task identifier is associated with the region identifier, spatial matching result, and temporal matching result, thus generating a dispatch dataset.

[0081] S2: Construct dispatch element nodes based on dispatch element data in the dispatch dataset, and establish node association edges based on the spatiotemporal matching relationship in the dispatch dataset to form a dispatch collaboration graph.

[0082] S2.1: Extract task identifier, execution resource identifier, and region identifier from the dispatch dataset, construct task element nodes, resource element nodes, and region element nodes to form a dispatch element node set.

[0083] It should be noted that task identifiers, execution resource identifiers, and region identifiers are extracted from the dispatch dataset. The task identifier is used as the node identifier for the task element node, the execution resource identifier as the node identifier for the resource element node, and the region identifier as the node identifier for the region element node. Duplicate task identifiers, execution resource identifiers, and region identifiers are merged and retained. Corresponding node types are added to the task element node, resource element node, and region element node, so that the task element node can correspond to the task data to be dispatched, the resource element node can correspond to the execution resource data, and the region element node can correspond to the region operation status, forming a dispatch element node set.

[0084] S2.2: Retrieve task data to be dispatched, execution resource data, and regional operation status from the dispatch dataset according to node type, configure the attributes of the dispatch element node set, and form a node attribute table.

[0085] It should be noted that the node type in the dispatch element node set is used as the basis for data retrieval. The task element node retrieves the pending task data that matches the task identifier in the dispatch data set. The resource element node retrieves the execution resource data that matches the execution resource identifier in the dispatch data set. The region element node retrieves the region operation status that matches the region identifier in the dispatch data set. The retrieved data content is then matched with the attributes of the task element node, resource element node, and region element node, so that the task element node carries the attribute corresponding to the pending task data, the resource element node carries the attribute corresponding to the execution resource data, and the region element node carries the attribute corresponding to the region operation status, forming a node attribute table.

[0086] S2.3: Based on the spatial matching results and temporal matching results, determine the spatiotemporal relationships between task element nodes, resource element nodes, and regional element nodes, and form a node relationship table.

[0087] It should be noted that the matching content with task identifier, execution resource identifier, and region identifier in the spatial matching results is compared with the matching content with the same task identifier and execution resource identifier in the temporal matching results. For the matching content with the same task identifier and execution resource identifier that can correspond to the region identifier, the spatiotemporal relationship between task element nodes and resource element nodes is determined, the regional affiliation relationship between task element nodes and region element nodes is determined, and the regional affiliation relationship between resource element nodes and region element nodes is determined. Matching content with inconsistent task identifiers, inconsistent execution resource identifiers, or that cannot correspond to the region identifier is eliminated. The retained spatiotemporal relationship and regional affiliation relationship are then matched with the task element nodes, resource element nodes, and region element nodes to form a node relationship table.

[0088] S2.4: Configure task resource distance elements, task time constraint elements, resource capacity matching elements, and regional load elements according to the node relationship table to form a node association edge set.

[0089] It should be explained that the relationships between task element nodes, resource element nodes, and regional element nodes are read from the node relationship table. Bidirectional directed edges are established according to the node relationships, and edge type codes are configured for the relationships between task element nodes and resource element nodes, between task element nodes and regional element nodes, and between resource element nodes and regional element nodes, respectively. Task-resource distance elements, task-time constraint elements, and resource-capability matching elements are configured between task element nodes and resource element nodes, and regional load elements are configured between task element nodes and regional element nodes, and between resource element nodes and regional element nodes. Duplicate relationships are merged, and relationships lacking the necessary elements for the corresponding edge type are removed to form a set of node-related edges.

[0090] S2.5: Encapsulate the dispatch element node set, node attribute table, and node association edge set into a graph structure to form a dispatch collaboration graph.

[0091] It should be noted that the dispatch element node set is used as the node content in the graph structure. The attribute content in the node attribute table is attached to the task element node, resource element node, and region element node with the same node identifier. The edge content in the node association edge set is connected to the corresponding task element node, resource element node, and region element node. The node content, attribute content, and edge content are checked for node identifier consistency. Node content with missing node identifiers, attribute content that cannot match node identifiers, and edge content that cannot connect to the corresponding node are deleted. Task element nodes with complete node identifiers and attribute content but not yet associated with resource element nodes are retained, and spatial association matching and temporal association matching are re-executed in the next dispatch cycle. The node content, attribute content, and edge content that have completed the verification are organized into a unified graph structure to form the dispatch collaboration graph.

[0092] S3: Input the dispatch collaboration graph into the pre-built dispatch collaboration model, perform node embedding, edge weight calculation and attention fusion on the dispatch element nodes, and generate the dispatch collaboration representation.

[0093] S3.1: Perform node type labeling, feature encoding, and scale normalization on the node attribute table to form a node encoding feature table.

[0094] It should be noted that the node identifiers in the node attribute table are compared with the task element nodes, resource element nodes, and regional element nodes in the dispatch element node set to confirm the node type corresponding to the node identifier. The task element node type, resource element node type, or regional element node type is then marked in the attribute record corresponding to the node identifier in the node attribute table. Using the global capability dictionary, status dictionary, and regional dictionary formed and solidified during the dispatch collaboration model training phase, service capability tags are split into multi-bit binary codes according to capability categories. The code bit matching the corresponding capability category is set to 1, and the code bit not matching the corresponding capability category is set to 0. The current task status is converted into status code values ​​according to idle, running, unavailable, and offline. The serviceable area is converted into a region coverage code value according to the region identifier. The code bit covering the corresponding region identifier is set to 1, and the code bit not covering the corresponding region identifier is set to 0. The value is 0; the task location and real-time location are represented by the coordinate field in the regional scheduling record as location coordinate values; the task release time, expected completion time and available service period are represented by the time offset within the collection date as time feature values; the available resources in the region are represented by the resource quantity value; and the regional resource load rate is represented by the load ratio value. Based on the maximum and minimum values ​​of each continuous field statistically fixed during the order dispatch collaboration model training phase, the location coordinate value, time feature value, resource quantity value and load ratio value are normalized. Continuous feature values ​​that exceed the fixed value range are processed according to the corresponding boundary value. When the maximum value and the minimum value are equal, the corresponding normalization result is set to 0. The node identifier, node type, multi-bit binary code, status code value, regional coverage code value and normalized continuous feature value are organized according to the same attribute record to form a node code feature table.

[0095] S3.2: Input the node encoding feature table into the dispatch collaboration model, and perform same-dimensional vector mapping on task element nodes, resource element nodes and regional element nodes to form a node embedding matrix.

[0096] It should be noted that the training set of the dispatch collaboration model is divided into training input samples and supervision labels. The training input samples consist of the node encoding feature table and node association edge set corresponding to historical dispatch records. The supervision labels consist of historical dispatch matching results, historical fulfillment adaptation results, historical completion time results, historical abnormal risk results, and historical service quality results. The supervision labels are only used for training error calculation and are not used as input features of the dispatch collaboration model. Successful dispatch markers are used as matching labels, and fulfillment adaptation labels are determined based on resource order acceptance status, task completion status, and fulfillment constraint satisfaction. The expected completion time is subtracted from the task release time to obtain the allowed completion time. The actual completion time is not greater than the allowed completion time. The completion time label is set to 1. When the actual completion time is longer than the allowed completion time, the ratio of the allowed completion time to the actual completion time is used as the completion time label. Historical order records with zero or missing actual completion time are not included in the completion time training. Abnormal interruption records, delay records, and failure records are treated as risk event items. The risk weighting value is calculated according to the severity coefficient and the number of occurrences of the risk event item, and the risk weighting value is normalized to 0 to 1 to form a single abnormal risk label. When there is no risk event item, the abnormal risk label is set to 0. Service evaluation records and quality inspection results are forward normalized, and complaint records are reverse normalized and weighted and summarized to form a service quality label.

[0097] The dispatching collaboration model includes an input layer, a node embedding layer, an edge weight calculation layer, a two-layer graph attention fusion layer, a matching prediction output head, a fulfillment adaptation output head, a completion time output head, an anomaly risk output head, and a service quality output head. The node embedding layer maps the node type, encoded value, and normalized continuous feature values ​​from the node encoding feature table into a d-dimensional node embedding vector. The edge weight calculation layer encodes edges according to their type and calculates edge weight coefficients by calling the corresponding task resource distance elements, task time constraint elements, resource capacity matching elements, or regional load elements. The first-layer graph attention fusion layer performs a linear transformation on the node embedding vector, calculates attention weights by combining the edge weight coefficients, and uses the LeakyReLU activation function to process attention scoring. The adjacency aggregation result is processed using the ELU activation function to form the first-layer node collaborative vector. The second-layer graph attention fusion layer takes the first-layer node collaborative vector as input and forms the second-layer node collaborative vector using the same processing method. The two graph attention fusion layers are configured with linear transformation weights, attention weights, and bias values ​​respectively, and do not share training parameters. The matching prediction output head, fulfillment adaptation output head, completion timeout output head, abnormal risk output head, and service quality output head are respectively processed using fully connected layers and the Sigmoid activation function, and the output values ​​are restricted to 0 to 1 to obtain historical order matching prediction results, historical fulfillment adaptation prediction results, historical completion timeout prediction results, historical abnormal risk prediction results, and historical service quality prediction results.

[0098] The model performs a multiplication operation on the natural logarithm of the label value and the predicted value. It then subtracts 1 from both the label value and the predicted value, multiplies the label difference from the predicted difference by the natural logarithm of the subtraction results, adds the results of these two multiplications, and takes the opposite of the result to obtain the single prediction loss. Finally, it adds all single prediction losses and divides them by the number of historical dispatch records to obtain the matching prediction error and fulfillment adaptation error. The model then subtracts the historical completion time prediction results, historical abnormal risk prediction results, and historical service quality prediction results from their corresponding labels, multiplies each subtraction result by itself, adds all multiplication results, and divides them by the number of historical dispatch records to obtain the completion time error, abnormal risk error, and service quality error. Finally, it adjusts the input layer, node embedding layer, edge weight calculation layer, two-layer graph attention fusion layer, and the connection weights and biases of each output head according to the model training error until the model training error meets the error threshold or the maximum number of training rounds is reached, resulting in the trained dispatch collaboration model.

[0099] It should also be noted that the error threshold is set based on the model training error formed by the weighted sum of five errors in the validation set and the change range within consecutive training rounds. Training ends when the model training error is not greater than the error threshold, which is set to 0.01.

[0100] The maximum number of training rounds is set based on the number of training sample entries, the number of node encoding feature fields, and the model training error reduction rate. Training ends when the model training error has not reached the error threshold but the number of training rounds has reached the maximum number of training rounds. The maximum number of training rounds is 100 rounds.

[0101] The node encoding feature table is input into the node embedding layer of the trained dispatch collaboration model. The node embedding layer extracts the node type, encoding value, and normalized continuous feature value corresponding to task element nodes, resource element nodes, and regional element nodes according to the node identifier. The connection weights and bias values ​​corresponding to the node embedding layer are obtained by back-updating the model training error during the dispatch collaboration model training process, and are solidified as training completion parameters when the model training error meets the set error threshold or the number of training rounds reaches the maximum. The node embedding layer calls the solidified connection weights and bias values ​​to perform same-dimensional vector mapping on the feature content of different node types, so that task element nodes, resource element nodes, and regional element nodes are converted into node embedding vectors of the same dimension. The node embedding vectors are arranged according to the node identifier and node type to form a node embedding matrix.

[0102] S3.3: According to the edge type encoding in the node association edge set, classify the adjacent nodes in the node embedding matrix and calculate the edge weights to form an edge weight coefficient table.

[0103] It should be noted that bidirectional directed edges between task element nodes and resource element nodes, between task element nodes and regional element nodes, and between resource element nodes and regional element nodes are distinguished according to edge type encoding. The task-resource distance element, task-time constraint element, resource-capability matching element, and regional load element are normalized to 0 to 1 respectively. The normalized result of the task-resource distance element is subtracted from the value of 1 to obtain the inverse distance value. The normalized results of the inverse distance value, task-time constraint element, and resource-capability matching element are multiplied by their corresponding edge weights, and the results of these three multiplications are added to obtain the edge weight coefficient between task element nodes and resource element nodes. The normalized result of the value of 1 is subtracted from the value of 1 to obtain the edge weight coefficient between task element nodes and regional element nodes, and between resource element nodes and regional element nodes. The edge weight coefficients are then associated with their corresponding edge type encodings to form an edge weight coefficient table.

[0104] It should also be noted that the edge weights include the edge weights of task resource distance, task time constraint, and resource capability matching, which are set based on the degree of influence of spatial accessibility, time accessibility, and capability adaptability on successful dispatch and fulfillment results in historical dispatch records. The task resource distance and task time constraint factors affect the cost of reaching the execution resource and the degree of fulfillment of the task deadline, respectively, with a weight of 0.30. The resource capability matching factor directly affects whether the execution resource has the conditions to undertake the task, with a weight of 0.40. The sum of the three edge weights is 1.

[0105] S3.4: Perform attention aggregation on the node embedding matrix according to the edge weight coefficient table to form the node collaborative representation matrix.

[0106] It should be noted that the edge weight coefficients in the edge weight coefficient table are matched with the corresponding node embedding vectors in the node embedding matrix. Each edge weight coefficient associated with the same target node is subjected to exponential operation, and the results of each exponential operation are added to obtain the sum of exponents. The results of each exponential operation are then divided by the sum of exponents to obtain the attention weights of the corresponding adjacent nodes, ensuring that the sum of the attention weights corresponding to the same target node is 1. The attention weights are then multiplied item by item with the node embedding vectors of the corresponding adjacent nodes, and the results of each multiplication under the same target node are added to obtain the adjacency aggregation vector. The first graph attention fusion layer uses the node embedding matrix as the aggregation object, and the second graph attention fusion layer uses the node collaboration vector formed in the first layer as the aggregation object. Each graph attention fusion layer sets a set of adjacency aggregation fusion weights and self-node fusion weights for task element nodes, resource element nodes, and regional element nodes. Both weights are node type-level scalar parameters and are not applied to individual nodes. Each node or edge is set individually; the adjacency aggregation fusion weight and the self-node fusion weight are exponentially calculated separately, and the results of the two exponential calculations are added to obtain the sum of the fusion weights. The results of the two exponential calculations are then divided by the sum of the fusion weights to make the sum of the adjacency aggregation fusion weight and the self-node fusion weight equal to 1. The adjacency aggregation vector and the vector of the corresponding node are multiplied by the adjacency aggregation fusion weight and the self-node fusion weight, respectively, and the results of the two multiplications are added to obtain the node collaboration vector. The adjacency aggregation fusion weight and the self-node fusion weight are updated by the two-layer graph attention fusion layer according to the model training error during the order dispatch collaboration model training process, and are solidified into the attention fusion layer parameters after the order dispatch collaboration model training is completed, so that task element nodes, resource element nodes, and regional element nodes can obtain the collaboration information of the associated nodes respectively. The node collaboration vectors formed by the second-layer graph attention fusion layer are arranged according to the node identifier and node type to form a node collaboration representation matrix.

[0107] S3.5: Perform node index alignment and vector dimension unification processing on the node collaborative representation matrix to generate the dispatch collaborative representation.

[0108] It should be noted that the consistency verification between the vector content in the node collaborative representation matrix and the node identifiers in the dispatch element node set is performed. Vector content that corresponds to task element nodes, resource element nodes, and regional element nodes is retained. The nodes are grouped according to the node type order of task element nodes, resource element nodes, and regional element nodes, and arranged in ascending order of node identifier within the same node type. Continuous node numbers are assigned to the arranged dispatch element nodes, and the arrangement of vector content is adjusted according to the continuous node numbers to complete the node number alignment. The vector content in the node collaborative representation matrix is ​​obtained by aggregating d-dimensional node embedding vectors from two layers of graph attention fusion layers. The vector dimension is consistent with the node embedding vector dimension. The d-dimensional consistency verification is performed on the vector content after node number alignment to confirm that the vector content corresponding to task element nodes, resource element nodes, and regional element nodes all maintain a d-dimensional vector structure. Dispatch collaborative representation record rows are established according to node identifiers, with the node number as the record row index, the node type as the record row type field, and the d-dimensional vector content as the record row feature field. The same node identifier corresponds to only one record row, generating the dispatch collaborative representation.

[0109] It should also be noted that d is the node embedding vector dimension determined and fixed during the training phase of the dispatch collaboration model. By comparing the training error of the validation set model and the stability of matching prediction, performance adaptation prediction, completion time prediction, abnormal risk prediction and service quality prediction corresponding to different candidate vector dimensions, the candidate vector dimension with lower model training error and stable prediction results is selected, with a value of 64.

[0110] S4: Calculate the collaborative recall score of execution resource data based on the dispatch collaborative characterization, perform resource availability verification in combination with execution resource data, limit the number of resource recalls based on the regional operation status, and generate an execution resource set.

[0111] S4.1: Decompose the dispatch collaboration representation according to the node type of task element nodes, resource element nodes and regional element nodes, extract the collaboration vector and collect them according to the node identifier to form a recall vector table.

[0112] It should be explained that the vector content carrying task element node types in the dispatch collaboration representation is divided into collaboration vectors corresponding to task element nodes; the vector content carrying resource element node types in the dispatch collaboration representation is divided into collaboration vectors corresponding to resource element nodes; and the vector content carrying region element node types in the dispatch collaboration representation is divided into collaboration vectors corresponding to region element nodes. The partitioned collaboration vectors are then validated for node identifiers. Collaboration vectors that correspond to task identifiers, execution resource identifiers, and region identifiers are retained, while collaboration vectors with missing node types or whose node identifiers cannot be matched are removed. The retained collaboration vectors are then aggregated and organized with their corresponding node types, node identifiers, task identifiers, execution resource identifiers, and region identifiers to form a recall vector table.

[0113] S4.2: Perform availability verification on the service capability tags, serviceable areas, current task status, and available service time periods associated with the recall vector table to obtain the recallable resource table.

[0114] It should be noted that the execution resource identifier corresponding to the resource element node in the recall vector table is matched with the service capability tag, serviceable area, current task status, and available service time period in the execution resource data. It verifies whether the service capability tag can match the resource capability matching element, whether the serviceable area covers the area corresponding to the task location in the task data to be dispatched, whether the current task status allows participation in order dispatch, and whether the available service time period covers the time range from the task release time to the expected completion time. Execution resource identifiers that meet the verification requirements for service capability tag, serviceable area, current task status, and available service time period are included in the recall scope and are then organized with the corresponding collaboration vector, task identifier, and area identifier to obtain the recallable resource table.

[0115] S4.3: By using the task resource distance element, task time constraint element, resource capacity matching element and regional load element corresponding to the recallable resource table, the execution resources are collaboratively recalled and scored to obtain the resource recall score table.

[0116] It should be explained that, according to the task identifier, execution resource identifier, and region identifier, the corresponding collaborative vectors of task element nodes, resource element nodes, and region element nodes are extracted from the recall vector table, and concatenated in a fixed order to form recall collaborative features. The collaborative matching value is calculated through the matching prediction output head of the dispatch collaborative model. The task resource distance element, task time constraint element, resource capability matching element, and region load element are normalized to a unified value scale. 1 is subtracted from the normalized results of the task resource distance element and the region load element to obtain the distance inversion value and the load inversion value. The collaborative matching value, distance inversion value, normalized result of task time constraint element, normalized result of resource capability matching element, and load inversion value are multiplied with the corresponding recall weights, and the results of the five multiplication operations are added to obtain the collaborative recall score. The collaborative recall score is then organized with the task identifier, execution resource identifier, and region identifier to obtain the resource recall score table.

[0117] It should also be noted that the recall weights include the collaborative matching value recall weight, the task resource distance factor recall weight, the task time constraint factor recall weight, the resource capacity matching factor recall weight, and the regional load factor recall weight; the collaborative matching value reflects the overall correlation between task element nodes, resource element nodes, and regional element nodes, and has a value of 0.30; the resource capacity matching factor directly determines whether the execution resource has the ability to undertake the task, and has a value of 0.25; the task time constraint factor affects whether the task can be completed within the expected completion time, and has a value of 0.20; the task resource distance factor affects the cost of reaching the execution resource and the response speed, and has a value of 0.15; the regional load factor is used to suppress the concentrated recall of resources in high-load areas, and its impact on the undertaking of a single task is relatively indirect, and has a value of 0.10; the sum of the five weights is 1.

[0118] The fixed order is determined by the arrangement of the input fields in the fulfillment adaptation output header during the training phase of the dispatch collaboration model, in order to maintain the consistency of feature positions between the training phase and the actual calculation phase. The fixed order is as follows: task element node corresponding to collaboration vector, resource element node corresponding to collaboration vector, regional element node corresponding to collaboration vector, task resource distance element, task time constraint element, resource capability matching element, and regional load element.

[0119] S4.4: Sort the execution resource identifiers associated with the same task identifier in descending order according to the resource recall scoring table, and remove the execution resource identifiers that are below the preset recall threshold to obtain the resource candidate ranking table.

[0120] It should be noted that, based on the task identifier in the resource recall scoring table, execution resource identifiers associated with the same task identifier and their corresponding collaborative recall scores are grouped into the same group. Execution resource identifiers within the same group are then arranged from highest to lowest according to their collaborative recall scores. The arranged collaborative recall scores are compared item by item with the preset recall threshold. Execution resource identifiers whose collaborative recall scores reach the preset recall threshold are retained, while those whose collaborative recall scores are lower than the preset recall threshold are removed. The retained execution resource identifiers, their corresponding collaborative recall scores, task identifiers, and region identifiers are then sorted in order to obtain the resource candidate ranking table.

[0121] It should also be noted that the recall threshold is determined based on the collaborative recall score distribution corresponding to the successfully fulfilled execution resources in the historical order. When the collaborative recall score is expressed as a normalized numerical value, it is 0.60.

[0122] S4.5: Extract execution resource identifiers based on the resource candidate sorting table and the available resource quantity in the region, and generate an execution resource set.

[0123] It should be noted that the execution resource identifiers associated with the same task identifier in the resource candidate sorting table are matched with the region identifiers according to the order of the collaborative recall score from high to low. The region identifiers are then matched with the amount of available resources in the region's operating status. The number of execution resources currently in the dispatchable state in the region's operating status is taken as the region's retainable quantity, which is an integer value corresponding to the region's available resources. When the number of execution resource identifiers under the same task identifier exceeds the region's retainable quantity, execution resource identifiers are truncated according to the order of the collaborative recall score from high to low. Execution resource identifiers whose ranking is not greater than the region's retainable quantity are included in the execution resource set, while those whose ranking is greater than the region's retainable quantity are not included in the execution resource set. When the number of execution resource identifiers under the same task identifier does not exceed the region's retainable quantity, the execution resource identifiers in the resource candidate sorting table are retained, and the retained execution resource identifiers are matched with the task identifier, region identifier, and collaborative recall score to generate the execution resource set.

[0124] S5: Input the execution resource set and dispatch collaboration representation into the dispatch collaboration model, calculate the performance fit value and predict the completion time value, abnormal risk value and service quality value, and map the profile tags according to the preset level range to generate the dispatch performance profile.

[0125] S5.1: Retrieve the dispatch collaboration representation according to the same task identifier in the execution resource set, and pair the collaboration vectors of task element nodes, resource element nodes and regional element nodes to form a performance adaptation vector group.

[0126] It should be noted that the execution resource identifier and region identifier corresponding to the same task identifier in the execution resource set are used as the positioning basis. In the dispatching collaboration representation, the collaboration vectors corresponding to the task element nodes, resource element nodes, and region element nodes are located respectively. Content where the task identifier, execution resource identifier, and region identifier are all consistent, and the collaboration vectors corresponding to the task element nodes, resource element nodes, and region element nodes are all complete, is determined as valid pairing content. The collaboration vectors corresponding to the task element nodes, resource element nodes, and region element nodes in the valid pairing content are paired and organized according to the same task identifier and the same execution resource identifier to form a performance adaptation vector group.

[0127] S5.2: Retrieve the task resource distance elements, task time constraint elements, resource capability matching elements, and regional load elements associated with the performance adaptation vector group, and perform normalization and fusion processing to form a performance adaptation feature group.

[0128] It should be noted that, based on the task identifier, execution resource identifier, and region identifier in the performance adaptation vector group, the task resource distance element, task time constraint element, resource capability matching element, and region load element corresponding to the same task identifier, the same execution resource identifier, and the same region identifier are matched from the dispatch dataset. Missing data is checked and the units are transformed for the task resource distance element, task time constraint element, resource capability matching element, and region load element. The distance size, time satisfaction degree, capability matching degree, and region load degree are converted to establish performance adaptation feature records according to the task identifier, execution resource identifier, and region identifier. The collaborative vectors corresponding to the task element node, the resource element node, and the region element node are concatenated in a fixed order to form a collaborative concatenation vector. The normalized task resource distance element, task time constraint element, resource capability matching element, and region load element are appended to the end of the collaborative concatenation vector, so that the same performance adaptation feature record simultaneously contains collaborative vector features and dispatch element features, forming a performance adaptation feature group.

[0129] S5.3: Perform capability consistency verification, spatiotemporal reachability verification, and regional load constraint verification on the performance adaptation feature group to form the performance constraint verification result.

[0130] It should be explained that the resource capability matching elements are matched with the service capability tags in the execution resource data to confirm whether the execution resource identifier has the service capability to undertake the corresponding task to be dispatched; when performing spatiotemporal reachability verification on the performance adaptation feature group, the arrival relationship between the task location and the real-time location is determined according to the task resource distance element, and the available service period is confirmed to cover the task time range corresponding to the task release time to the expected completion time according to the task time constraint element; the regional load element is matched with the available regional resources and regional resource load rate in the regional operation status. When the available regional resources are greater than 0 and the regional resource load rate is not higher than the preset regional load threshold, the regional load constraint verification is determined to be passed; when the available regional resources are equal to 0 or the regional resource load rate is higher than the preset regional load threshold, the regional load constraint verification is determined to be failed; the passed and failed contents obtained from the capability consistency verification, spatiotemporal reachability verification, and regional load constraint verification are organized according to the task identifier, execution resource identifier, and regional identifier to form the performance constraint verification result.

[0131] It should also be noted that the regional load threshold is set based on the corresponding distribution between regional resource load rate and task delays, abnormal interruptions, and order failures in historical dispatch records, and is set to a value of 0.80.

[0132] S5.4: Input the performance adaptation feature group into the dispatch collaboration model to calculate the performance adaptation prediction value, and perform constraint filtering on the performance adaptation prediction value based on the performance constraint verification result to form the performance adaptation result.

[0133] It should be noted that the performance adaptation feature records corresponding to the same task identifier, the same execution resource identifier, and the same region identifier in the performance adaptation feature group should be aligned with the corresponding capability consistency verification content, spatiotemporal reachability verification content, and regional load constraint verification content in the performance constraint verification results. The task element node collaboration vector, resource element node collaboration vector, regional element node collaboration vector, task resource distance element, task time constraint element, resource capability matching element, and regional load element in the performance adaptation feature records should be input into the performance adaptation output header of the dispatch collaboration model in a fixed order, ensuring that the input dimension in the actual calculation stage remains 3D+4D, and aligning with the input fields in the dispatch collaboration model training stage. The arrangement is consistent; the performance adaptation output head takes the feature values ​​of each dimension in the performance adaptation feature record as input, multiplies each feature value with the corresponding connection weights after training, accumulates the product results and adds them with the bias value corresponding to the performance adaptation output head to obtain the performance adaptation prediction value under a unified value scale; when the capability consistency verification content, spatiotemporal reachability verification content, and regional load constraint verification content are all passed, the corresponding performance adaptation prediction value is retained; when any verification content fails, the corresponding performance adaptation prediction value is set to 0; the task identifier, execution resource identifier, and regional identifier of the performance adaptation prediction value after constraint screening are sorted together to form the performance adaptation result.

[0134] S5.5: Match the performance adaptation results with the dispatch dataset according to the task identifier and execution resource identifier, and normalize the matched content to form a performance prediction feature table.

[0135] It should be noted that, using task identifier and execution resource identifier as corresponding fields, the performance adaptation value and region identifier in the performance adaptation result are matched with the pending task data, execution resource data, region operation status, and dispatch element data corresponding to the same task identifier and execution resource identifier in the dispatch dataset. Matching content where both task identifier and execution resource identifier are consistent and the performance adaptation value, region identifier, and dispatch element data are complete is retained, while content with inconsistent task identifiers, inconsistent execution resource identifiers, or missing matching content is removed. The performance adaptation value, task resource distance element, task time constraint element, resource capacity matching element, region load element, region available resource quantity, and region resource load rate in the retained matching content are converted to a unified value scale, while maintaining the correspondence between task identifier, execution resource identifier, and region identifier. The normalized matching content is then aggregated according to task identifier and execution resource identifier to form a performance prediction feature table.

[0136] S5.6: Calculate the completion time value, abnormal risk value and service quality value corresponding to the performance prediction feature table through the dispatch collaboration model to form the performance prediction result.

[0137] It should be explained that after inputting the performance prediction feature table into the trained dispatch collaboration model, the completion time output head calls the normalized results corresponding to the performance adaptation value and task time constraint elements, and performs weighted summation according to the trained connection weights and bias values ​​to obtain the completion time value; the anomaly risk output head calls the normalized results corresponding to the regional load element, regional resource load rate, regional available resources, and performance adaptation value, and performs a reverse transformation on the normalized results corresponding to the regional available resources and performance adaptation value, and performs weighted summation according to the trained connection weights and bias values ​​to obtain the anomaly risk value; the service quality output head calls the normalized results corresponding to the resource capability matching element, task resource distance element, and performance adaptation value, and performs a reverse transformation on the normalized results corresponding to the task resource distance element, and performs weighted summation according to the trained connection weights and bias values ​​to obtain the service quality value; the completion time value, anomaly risk value, and service quality value are then matched with the task identifier, execution resource identifier, and regional identifier to form the performance prediction result.

[0138] S5.7: Merge the performance adaptation results and performance prediction results based on the task identifier and execution resource identifier, and perform profile tag mapping to generate a dispatch performance profile.

[0139] It should be explained that, based on the task identifier and execution resource identifier, the performance adaptation value in the performance adaptation result and the completion timeliness value, abnormal risk value, and service quality value in the performance prediction result are merged. Content where both the task identifier and execution resource identifier are consistent is retained, while content with inconsistent task identifiers, inconsistent execution resource identifiers, or missing performance adaptation value, completion timeliness value, abnormal risk value, and service quality value is removed. The merged performance adaptation value, completion timeliness value, abnormal risk value, and service quality value are compared with preset level ranges to determine the level range into which each value falls. Performance adaptation profile labels, completion timeliness profile labels, abnormal risk profile labels, and service quality profile labels are then assigned according to the level range into which they fall. Finally, these labels are mapped and organized with the task identifier, execution resource identifier, region identifier, performance adaptation value, completion timeliness value, abnormal risk value, and service quality value to generate a dispatch performance profile.

[0140] It should also be noted that the grade range is set based on the distribution of performance adaptation value, completion timeliness value, abnormal risk value, and service quality value in the completed order dispatch records. It is used to distinguish the performance adaptation degree, completion timeliness level, abnormal risk level, and service quality level corresponding to different execution resource identifiers. When the performance adaptation value, completion timeliness value, abnormal risk value, and service quality value are expressed using normalized numerical values, the grade range includes a low grade range, a medium grade range, and a high grade range. The low grade range is greater than or equal to 0 and less than 0.4, the medium grade range is greater than or equal to 0.4 and less than 0.7, and the high grade range is greater than or equal to 0.7 and less than or equal to 1.

[0141] S6: After performing attribution verification and dimension normalization on the dispatch fulfillment profile, calculate the regional balance factor in conjunction with the regional operation status, and perform fulfillment gain, risk attenuation, quality enhancement and regional balance adjustment calculations to generate dispatch recommendation scores.

[0142] S6.1: Perform attribution verification and dimension normalization on the fulfillment adaptation value, completion timeliness value, abnormal risk value and service quality value in the order fulfillment profile to form a profile scoring table.

[0143] It should be noted that, using task identifiers and execution resource identifiers as verification fields, the performance adaptation value, completion timeliness value, abnormal risk value, and service quality value in the dispatch performance profile are checked for consistency with the corresponding task identifier, execution resource identifier, and region identifier. Profile content with consistent task identifiers and execution resource identifiers and complete performance adaptation value, completion timeliness value, abnormal risk value, and service quality value is retained, while profile content with inconsistent task identifiers, inconsistent execution resource identifiers, missing region identifiers, or missing values ​​is removed. Dimensional normalization is performed using minimum and maximum value normalization methods. The retained performance adaptation value, completion timeliness value, abnormal risk value, and service quality value are converted to a unified scoring scale. The normalized performance adaptation value, completion timeliness value, abnormal risk value, and service quality value are then correlated with the task identifier, execution resource identifier, region identifier, and profile label to form a profile scoring table.

[0144] S6.2: Perform field matching and ratio normalization on the region identifier and region operation status in the portrait scoring table to form a region balance factor table.

[0145] It should be noted that the regional identifiers in the profile scoring table are matched item by item with the regional identifiers in the regional operation status. Profile scoring content with consistent regional identifiers and that can be associated with the available regional resources and regional resource load rate is retained, while profile scoring content with missing regional identifiers or regional operation status that cannot be matched is removed. After the field matching is completed, the available regional resources are divided by the maximum available regional resources in the same dispatch cycle to obtain the regional available resource ratio. 1 is subtracted from the regional resource load rate to obtain the regional load inverse ratio. The regional available resource ratio and the regional load inverse ratio are then weighted and merged to obtain the regional balance factor with a value ranging from 0 to 1. The regional balance factor is then correlated with the task identifier, execution resource identifier, regional identifier, and the normalized performance adaptation value, completion time value, abnormal risk value, and service quality value in the profile scoring table to form the regional balance factor table.

[0146] S6.3: Calculate performance gain, risk reduction, quality enhancement and regional balance adjustment for the profile scoring table and regional balance factor table to form a dispatch recommendation scoring table.

[0147] It should be explained that the task identifier, execution resource identifier, and region identifier in the profile scoring table are matched with the task identifier, execution resource identifier, and region identifier in the region balance factor table, retaining profile scoring content and region balance factors with consistent fields; the normalized performance adaptation value and completion timeliness value are added together, and the result of the addition is divided by the number of scoring items involved in the addition to obtain the performance gain value; the upper limit of the normalized scoring scale is subtracted from the normalized abnormal risk value to obtain the risk decay value; the normalized service quality value is used as the quality enhancement value; the region balance factor is used as the region equilibrium adjustment value; the performance gain value, risk decay value, quality enhancement value, and region equilibrium adjustment value are multiplied by the preset scoring weights respectively, and the results of each multiplication are added together to obtain the dispatch recommendation score; the dispatch recommendation score is then matched with the task identifier, execution resource identifier, region identifier, performance adaptation value, completion timeliness value, abnormal risk value, service quality value, and region balance factor to form the dispatch recommendation scoring table.

[0148] It should also be noted that the scoring weights include performance gain weight, risk attenuation weight, quality enhancement weight, and regional balance adjustment weight. These weights are set based on the degree of influence of performance completion, anomaly occurrence, service feedback, and regional resource utilization on the dispatch results in the completed dispatch business records. Performance completion directly reflects whether the execution resources can complete the dispatch task as required, and has the largest impact on the dispatch recommendation score, with a value of 0.35. Anomaly occurrence reflects the risk of interruption, delay, and failure of execution resources during the performance process, and has a deductive impact on the dispatch recommendation score, with a value of 0.25. Service feedback reflects the service quality level after the execution resources complete the task, and has an enhancing impact on the dispatch recommendation score, with a value of 0.25. Regional resource utilization is used to adjust the balance of execution resource allocation within the same region, and plays a supplementary adjustment role in the dispatch recommendation score, with a value of 0.15. The sum of the performance gain weight, risk attenuation weight, quality enhancement weight, and regional balance adjustment weight after normalization is 1.

[0149] The upper limit of the normalized scoring scale is set based on the normalized performance fit value, completion timeliness value, abnormal risk value and service quality value in the profile scoring table, as well as the range of proportional normalized values ​​corresponding to the regional balance factors in the regional balance factor table; when using normalized numerical representation, the example value of the upper limit of the normalized scoring scale is 1.

[0150] S6.4: Sort the execution resource identifiers in the dispatch recommendation scoring table according to the same task identifier, and generate a dispatch recommendation ranking table.

[0151] It should be noted that execution resource identifiers with the same task identifier in the dispatch recommendation score table should be grouped into the same sorting group, and the correspondence between execution resource identifiers, dispatch recommendation scores, region identifiers, and profile tags should be verified, retaining content with complete dispatch recommendation scores and unique execution resource identifiers; execution resource identifiers in the same sorting group should be arranged from high to low according to dispatch recommendation scores. If dispatch recommendation scores are the same, the abnormal risk value, completion time value, and service quality value should be compared in sequence, and the order of execution resource identifiers should be determined according to the order of smaller abnormal risk value, larger completion time value, and larger service quality value; the sorted execution resource identifiers should be matched with their corresponding task identifiers, region identifiers, profile tags, and dispatch recommendation scores to generate a dispatch recommendation sorting table.

[0152] S7: Sort the execution resource set according to the order dispatch recommendation score, and automatically determine the target execution resource based on the recommendation ranking result to generate automatic order dispatch recommendation results.

[0153] S7.1: Perform deduplication verification, effective score filtering, and profile tag matching on the execution resource identifiers under the same task identifier in the dispatch recommendation scoring table to form a recommendation candidate table.

[0154] It should be noted that execution resource identifiers with the same task identifier in the dispatch recommendation scoring table are grouped into the same candidate range, and duplicate execution resource identifiers are deduplicated. When duplicate content corresponds to the same dispatch recommendation score, it is merged into one record. When duplicate content corresponds to different dispatch recommendation scores, the record corresponding to the higher dispatch recommendation score is taken as the valid record. The deduplicated execution resource identifiers are then filtered for valid scores. Records with complete dispatch recommendation scores, task identifiers, and execution resource identifiers are included in the valid scoring range. Then, the performance matching profile label, completion time profile label, abnormal risk profile label, and service quality profile label are matched with the same task identifier and the same execution resource identifier. Records with complete profile labels and consistent correspondences are determined as recommended candidate records. The execution resource identifiers, dispatch recommendation scores, profile labels, task identifiers, and region identifiers in the recommended candidate records are then organized to form a recommended candidate table.

[0155] S7.2: Sort the execution resource identifiers in the recommended candidate table in descending order according to the dispatch recommendation score, and calculate the score difference between adjacent execution resources to form a recommendation ranking table.

[0156] It should be noted that the execution resource identifiers and dispatch recommendation scores corresponding to the same task identifier in the recommended candidate table are grouped together. The execution resource identifiers under the same task identifier are sorted in descending order of dispatch recommendation score from high to low. After sorting, the dispatch recommendation score corresponding to the current ranking position is subtracted from the dispatch recommendation score corresponding to the next ranking position to obtain the score difference between adjacent execution resources. For the execution resource identifier with the lowest ranking position, the score difference is recorded as a null value or marked as the lowest difference value. Using the task identifier and execution resource identifier as the main corresponding fields, the ranking position is written into the record where the corresponding execution resource identifier is located. The dispatch recommendation score, score difference, region identifier, profile tag, completion time value, abnormal risk value, service quality value, region available resources, and region resource load rate are retained in the same record. This ensures that the execution resource identifiers under the same task identifier simultaneously correspond to the ranking position, dispatch recommendation score, score difference, region identifier, profile tag, and fields required for confidence verification, forming a recommendation ranking table.

[0157] S7.3: Perform confidence checks on the score difference, completion time value, abnormal risk value, service quality value, regional available resources and regional resource load rate in the recommendation ranking table to form a target confirmation table.

[0158] It should be explained that the score difference, completion timeliness, service quality, and available regional resources in the recommendation ranking table are used as positive confidence content, while the abnormal risk value and regional resource load rate are used as negative confidence content. The score difference, completion timeliness, service quality, and available regional resources are proportionally normalized. The abnormal risk value and regional resource load rate are subtracted from the upper limit of the normalized score scale to obtain the negatively normalized content. The normalized score difference, completion timeliness, service quality, available regional resources, negatively normalized abnormal risk value, and negatively normalized regional resource load rate are added together, and then divided according to the number of confidence content involved in the addition to obtain the target confirmation value. Using the task identifier and execution resource identifier as the primary corresponding fields, the target confirmation value is written to the record corresponding to the ranking position, and the dispatch recommendation score, score difference, completion timeliness, abnormal risk value, service quality value, available regional resources, and regional resource load rate in the same record are retained to form the target confirmation table.

[0159] S7.4: Select the execution resource identifier that meets the confirmation threshold and is ranked first according to the target confirmation table to form the target resource table.

[0160] It should be explained that the execution resource identifiers corresponding to the same task identifier in the target confirmation table are arranged from first to last according to their sorting position. The target confirmation value corresponding to the execution resource identifier is compared with the confirmation threshold item by item. Execution resource identifiers whose target confirmation value reaches the confirmation threshold are retained, and execution resource identifiers whose target confirmation value does not reach the confirmation threshold are removed. Among the retained execution resource identifiers, the execution resource identifier with the highest sorting position is selected as the target execution resource identifier. The target execution resource identifier, task identifier, region identifier, sorting position, dispatch recommendation score, target confirmation value, and profile tag are written into the same record to form the target resource table.

[0161] It should also be noted that the confirmation threshold is set based on the distribution of target confirmation values ​​of successfully executed automatic dispatch recommendation results in the completed dispatch business records, and is used to filter out execution resource identifiers with insufficient confidence; when the target confirmation value is expressed as a normalized numerical value, the value is 0.70.

[0162] S7.5: Encapsulate the task identifier, target execution resource identifier, dispatch recommendation score, and profile tag in the target resource table into dispatch instructions to generate automatic dispatch recommendation results.

[0163] It should be explained that the task identifier, target execution resource identifier, dispatch recommendation score, and profile tag are extracted from the target resource table. The task identifier and target execution resource identifier are checked for completeness, and content with missing task identifiers, target execution resource identifiers, dispatch recommendation scores, or profile tags is removed. After the check passes, the task identifier is used as the target content for the task to be dispatched, the target execution resource identifier is used as the target content for the execution resource, the dispatch recommendation score is used as the basis for dispatch priority, and the profile tag is used as supplementary description of performance adaptation, completion time, abnormal risks, and service quality. The dispatch instruction is encapsulated according to the field order of task identifier, target execution resource identifier, dispatch recommendation score, and profile tag, so that the same task identifier can correspond to a unique target execution resource identifier and a corresponding dispatch recommendation score and profile tag, generating automatic dispatch recommendation results.

[0164] In summary, this invention, through the generation of dispatch datasets, the construction of dispatch collaboration graphs, and the generation of dispatch collaboration representations, enables a unified expression of the spatiotemporal, capability, and load relationships among tasks to be dispatched, execution resources, and regional operational status. Resource retrieval can simultaneously consider spatial reachability, temporal availability, capability adaptability, and regional carrying capacity, improving the accuracy and stability of execution resource selection. Through performance adaptation calculation, completion time prediction, anomaly risk prediction, service quality prediction, dispatch performance profile generation, and dispatch recommendation scoring, the automatic dispatch recommendation results can comprehensively reflect task acceptance capacity, performance reliability, risk controllability, service quality level, and regional balance, improving task execution efficiency, scheduling stability, and regional resource utilization.

[0165] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based automatic order dispatch recommendation method, characterized in that: include, Collect data on tasks to be dispatched, execution resources, regional operation status, and corresponding dispatch elements, and perform spatiotemporal correlation matching to generate a dispatch dataset. Based on the dispatch element data in the dispatch dataset, dispatch element nodes are constructed, and node association edges are established according to the spatiotemporal matching relationship in the dispatch dataset to form a dispatch collaboration graph. Input the dispatch collaboration graph into the pre-built dispatch collaboration model, and perform node embedding, edge weight calculation and attention fusion on the dispatch element nodes to generate a dispatch collaboration representation. The collaborative recall score of the execution resource data is calculated based on the dispatch collaborative characterization, the resource availability is verified in combination with the execution resource data, and the number of resources recalled is limited according to the regional operation status to generate an execution resource set. Input the execution resource set and the dispatch collaboration representation into the dispatch collaboration model, calculate the performance fit value and predict the completion time value, abnormal risk value and service quality value, and map the profile tags according to the preset level range to generate the dispatch performance profile; After performing attribution verification and dimensional normalization on the dispatch fulfillment profile, the regional balance factor is calculated in combination with the regional operation status. Then, the fulfillment gain, risk attenuation, quality enhancement and regional balance adjustment are calculated to generate dispatch recommendation scores. The execution resource set is recommended and ranked according to the dispatch recommendation score, and the target execution resource is automatically determined based on the recommendation ranking result to generate automatic dispatch recommendation results.

2. The automatic order dispatch recommendation method based on deep learning as described in claim 1, characterized in that: The data of tasks to be dispatched includes task identifier, task location, task release time, and expected completion time. The execution resource data includes execution resource identifier, real-time location, service capability tag, serviceable area, current task status, and available service time period; The regional operating status includes the regional identifier, the amount of available resources in the region, and the regional resource load rate. The dispatch data includes task resource distance, task time constraint, resource capacity matching, and regional load.

3. The automatic order dispatch recommendation method based on deep learning as described in claim 2, characterized in that: The generated dispatch dataset specifically includes, Based on the task location, real-time location, and area identifier, spatial correlation matching is performed on the task data to be dispatched, the execution resource data, and the area operation status to form a spatial matching result. Based on the task release time, expected completion time, current task status, and available service time, time association matching is performed on the task data to be dispatched and the execution resource data to form a time matching result. The spatial matching results and temporal matching results are linked and integrated according to the task identifier, execution resource identifier, and region identifier to generate a dispatch dataset.

4. The deep learning-based automatic order dispatch recommendation method as described in claim 3, characterized in that: The formation of the dispatch coordination diagram specifically includes, Extract task identifiers, execution resource identifiers, and region identifiers from the dispatch dataset, and construct task element nodes, resource element nodes, and region element nodes to form a dispatch element node set. According to the node type, retrieve the task data to be dispatched, the execution resource data, and the regional operation status from the dispatch dataset, configure the attributes of the dispatch element node set, and form a node attribute table. Based on the spatial matching results and the temporal matching results, the spatiotemporal relationships between task element nodes, resource element nodes and regional element nodes are determined, and a node relationship table is formed. Configure task resource distance elements, task time constraint elements, resource capacity matching elements, and regional load elements according to the node relationship table to form a node association edge set; The task dispatch element node set, node attribute table, and node associated edge set are encapsulated into a graph structure to form a task dispatch collaboration graph.

5. The deep learning-based automatic order dispatch recommendation method as described in claim 4, characterized in that: The generated dispatch collaborative representation specifically includes, The node attribute table is processed by node type labeling, feature encoding, and scale normalization to form a node encoding feature table; Input the node encoding feature table into the dispatch collaboration model, and perform same-dimensional vector mapping on task element nodes, resource element nodes and regional element nodes to form a node embedding matrix. According to the edge type encoding in the node association edge set, the adjacent nodes in the node embedding matrix are classified and edge weights are calculated to form an edge weight coefficient table; Attention aggregation is performed on the node embedding matrix according to the edge weight coefficient table to form the node collaborative representation matrix; The node collaborative representation matrix is ​​processed by aligning node numbers and unifying vector dimensions to generate a dispatch collaborative representation.

6. The automatic order dispatch recommendation method based on deep learning as described in claim 5, characterized in that: The generated execution resource set specifically includes, The dispatching collaboration representation is split according to the node type of task element nodes, resource element nodes and regional element nodes. The collaboration vector is extracted and collected according to the node identifier to form a recall vector table. Availability checks are performed on the service capability tags, serviceable areas, current task status, and available service time periods associated with the recall vector table to obtain the recallable resource table. By using the task resource distance element, task time constraint element, resource capacity matching element and regional load element corresponding to the recallable resource table, the execution resources are collaboratively recalled and scored to obtain the resource recall scoring table. According to the resource recall scoring table, the execution resource identifiers associated with the same task identifier are sorted in descending order, and execution resource identifiers below the preset recall threshold are removed to obtain the resource candidate ranking table. Based on the resource candidate sorting table and the available resources in the region, extract the execution resource identifiers and generate an execution resource set.

7. The deep learning-based automatic order dispatch recommendation method as described in claim 6, characterized in that: The aforementioned performance adaptation calculation for execution resources specifically includes, Based on the same task identifier in the execution resource set, the dispatch coordination representation is retrieved, and the coordination vectors of task element nodes, resource element nodes and regional element nodes are paired to form a performance adaptation vector group. Retrieve the task resource distance elements, task time constraint elements, resource capability matching elements, and regional load elements associated with the performance adaptation vector group, and perform normalization and fusion processing to form a performance adaptation feature group. Perform capability consistency verification, spatiotemporal reachability verification, and regional load constraint verification on the performance adaptation feature group to form the performance constraint verification result; The performance adaptation feature group is input into the dispatch collaboration model to calculate the performance adaptation prediction value, and the performance adaptation prediction value is constrained and filtered according to the performance constraint verification result to form the performance adaptation result.

8. The automatic order dispatch recommendation method based on deep learning as described in claim 7, characterized in that: The generation of the order fulfillment profile specifically includes: Match the performance adaptation results with the dispatch dataset according to the task identifier and execution resource identifier, and normalize the matching content to form a performance prediction feature table. The order dispatching collaboration model is used to calculate the completion time value, abnormal risk value and service quality value corresponding to the performance prediction feature table to form the performance prediction result. Based on the task identifier and execution resource identifier, the performance adaptation results and performance prediction results are merged and profile tags are mapped to generate a dispatch performance profile.

9. The automatic order dispatch recommendation method based on deep learning as described in claim 8, characterized in that: The generated order dispatch recommendation score specifically includes, The fulfillment matching value, completion timeliness value, abnormal risk value and service quality value in the order fulfillment profile are subjected to attribution verification and unit normalization processing to form a profile scoring table. The regional identifiers in the portrait scoring table are matched with the regional operating status and the proportions are normalized to form a regional balance factor table. The performance gain, risk reduction, quality enhancement and regional balance adjustment are calculated on the profile scoring table and the regional balance factor table to form the order recommendation scoring table. The execution resource identifiers in the dispatch recommendation scoring table are scored and sorted according to the same task identifier to generate a dispatch recommendation ranking table.

10. The deep learning-based automatic order dispatch recommendation method as described in claim 9, characterized in that: The generation of automatic order dispatch recommendation results specifically includes, The execution resource identifiers under the same task identifier in the dispatch recommendation scoring table are deduplicated, effective scores are filtered, and profile tags are matched to form a recommendation candidate table. The execution resource identifiers in the recommended candidate table are sorted in descending order according to the dispatch recommendation score, and the score difference between adjacent execution resources is calculated to form a recommendation ranking table; Confidence checks are performed on the score difference, completion time value, abnormal risk value, service quality value, regional available resources and regional resource load rate in the recommendation ranking table to form a target confirmation table; Select the execution resource identifier that meets the confirmation threshold and is ranked first in the target confirmation table to form the target resource table; The task identifier, target execution resource identifier, dispatch recommendation score, and profile tag in the target resource table are encapsulated into dispatch instructions to generate automatic dispatch recommendation results.

Images