A steel member transportation path scheduling method based on big data analysis

Abstract

The application discloses a steel member transportation path scheduling method based on big data analysis, which comprises the following steps: collecting transportation task data and preprocessing, generating a standardized transportation task sequence; identifying paths and nodes, and constructing a time sequence dynamic graph; constructing an improved time sequence dynamic graph attention network; dividing nodes into subgroups according to feature similarity, and generating a node embedding matrix; iteratively updating the node embedding, completing path-dependent self-loop, and generating a path embedding matrix; fusing the path embedding matrix with the previous iteration, and generating a path history embedding matrix; global multi-objective search of the colonial spider optimization algorithm determines the steel member transportation scheduling scheme. Through the construction of the improved time sequence dynamic graph attention network and the combination of the colonial spider optimization algorithm, the data-driven path embedding modeling, multi-round node iteration optimization and global multi-objective scheduling scheme generation of the steel member transportation task are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics and transportation route optimization technology, and in particular to a method for scheduling the transportation routes of steel components based on big data analysis. Background Technology

[0002] In steel structure construction, the transportation and scheduling of steel components is a crucial aspect of ensuring construction progress, safety, and the rational utilization of resources. Traditional methods for steel component transportation and scheduling often rely on manual experience or static planning, such as selecting transport vehicles based on on-site experience and manually arranging route sequences and node transport sequences. While this approach can maintain basic transportation arrangements in simple construction environments, it easily leads to problems such as route duplication, empty vehicle runs, uneven loading, or excessively long waiting times at nodes when facing complex situations involving large construction sites, multiple transport routes, and numerous transport vehicles. This negatively impacts transportation efficiency, increases construction costs, and may result in resource waste. Existing technologies typically lack the ability to systematically analyze and dynamically learn from historical transportation data, real-time road conditions, and vehicle status. The scheduling optimization process remains at the level of static planning or single-objective optimization, making it difficult to achieve global scheduling optimization and real-time adjustments, and also failing to fully utilize existing transportation records for dynamic prediction and scheduling improvement.

[0003] In recent years, with the development of big data and intelligent algorithms, some studies have attempted to abstract transportation nodes and paths into graph structures, using graph neural networks or data-driven methods to optimize transportation scheduling. However, most existing methods are geared towards static graph processing, failing to effectively characterize the dynamic changes of transportation nodes over time and historical transportation evolution information. Node embedding lacks grouping aggregation, iterative updates, and historical enhancement mechanisms, limiting the refinement and accuracy of scheduling schemes. Simultaneously, most optimization strategies rely on single heuristics or static optimization algorithms, lacking group information propagation, dynamic role allocation, and multi-objective coordination capabilities, failing to fully integrate node characteristics, path attributes, and historical data to achieve global multi-objective optimization. Therefore, existing technologies still have significant shortcomings in terms of the dynamic adaptability, global optimality, and multi-objective coordination of transportation scheduling, making it difficult to meet the demands of modern construction sites for efficient, precise, and intelligent steel component transportation scheduling.

[0004] Therefore, how to provide a method for scheduling the transportation routes of steel components based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for scheduling steel component transportation routes based on big data analysis. This invention utilizes transportation task data acquisition, standardized preprocessing, path and node feature extraction, temporal dynamic graph construction, improved temporal dynamic graph attention network node embedding generation, path embedding iterative update, path history embedding enhancement, multi-objective path search using a swarm spider optimization algorithm, and transportation scheduling scheme generation technology. It details the implementation process of transportation task data processing, path node feature extraction, node embedding generation and iteration, path embedding closed-loop update, history embedding enhancement, optimization search process, and final scheduling scheme generation. This achieves dynamic modeling, multi-objective optimization, and accurate planning of steel component transportation routes and vehicle scheduling. Compared to traditional transportation scheduling methods that rely on manual experience or static planning, this invention has the advantages of comprehensive multi-source data feature representation, accurate depiction of the dynamic state of transportation nodes and routes, full utilization of path history information, strong global multi-objective optimization capability, and accurate and real-time adjustable transportation scheduling scheme generation.

[0006] A method for scheduling steel component transportation routes based on big data analysis according to an embodiment of the present invention includes: Collect transportation task data for steel components, preprocess the transportation task data, and generate standardized transportation task sequences; Based on the standardized transportation task sequence, the paths and nodes are identified, and the node status, vehicle status and path attributes of each path are extracted as path node features to construct a time-series dynamic graph. Based on temporal dynamic graphs, an improved temporal dynamic graph attention network is constructed, including a node embedding adaptive grouping layer, a path-dependent self-looping layer, and a path history embedding enhancement layer. In the node embedding adaptive grouping layer, nodes are divided into subgroups based on feature similarity to generate a node embedding matrix; In the path-dependent self-loop layer, the node embedding matrix of each path is iteratively updated, information is passed between nodes within the path, and a path embedding matrix is ​​generated. In the path history embedding enhancement layer, the path embedding matrix of each path is fused with the corresponding path embedding matrix of the previous iteration, aligned according to the path node order, the node embedding state is updated, and a path history embedding matrix is ​​generated. The algorithm for optimizing the social spider population is based on the path history embedding matrix. Spider individuals are initialized, and female and male individuals are divided according to a dynamic role switching strategy. The individual positions are updated using an iterative step size adaptive update rule. Vibration information is propagated in the group to perform a global multi-objective path search, generate scheduling search results, and determine the steel component transportation scheduling plan, including the routes and time arrangements of each transport vehicle.

[0007] Optionally, the transportation task data specifically includes transportation node information, transportation vehicle information, steel component loading information, transportation route attribute information, transportation time identifier data, and historical transportation record identifier data.

[0008] Optionally, the preprocessing of the transportation task data includes: Identify and remove abnormal information in transportation task data, including missing node status, abnormal vehicle status, or abnormal route attribute data. Missing node status, vehicle status, or path attribute data are filled in by interpolating preceding and following nodes to generate complete information; The node status and vehicle status are sorted according to the transportation node number, vehicle number and task order to form a standardized node list arranged in transportation order. The transportation route attributes are standardized, including route length, capacity, traffic restrictions, and estimated transportation time, and each route attribute is mapped to a unified dimension. The transportation time identification data is corrected to ensure that the time sequence from the origin to the destination is consistent with the node sequence, abnormal time records are corrected, and a standardized transportation time sequence is formed.

[0009] Optionally, the construction of the timing dynamic graph includes: Based on the standardized transportation task sequence, extract the information of each path and node, number them according to the transportation sequence, and generate a path node list; Based on the list of path nodes, vehicle status characteristics are assigned to each path node, including vehicle type, load and current location status, to form path node status information; Extract the attribute information of each path, including path length, traffic capacity, traffic restrictions and estimated transportation time, and associate it with the corresponding path nodes to generate path attribute features; The status information of each path node is combined with the path attribute information to generate a path node feature set, while keeping the node order consistent with the path. A time-series dynamic graph is constructed based on the feature set of path nodes. Path nodes are represented as graph nodes, and the connection relationships between nodes and paths are represented as graph edges. Path attribute information is written into the corresponding graph edge attributes to generate an initial time-series dynamic graph. Assign a unique identifier to each node and edge in the initial time-series dynamic graph, and record the path node features and edge attributes on the corresponding nodes and edges to form a complete time-series dynamic graph.

[0010] Optionally, the construction of the improved temporal dynamic graph attention network includes: The improved temporal dynamic graph attention network includes a node embedding adaptive grouping layer, a path-dependent self-looping layer, and a path history embedding enhancement layer. The parameters of each layer are initialized, including weights, biases, and node embedding vector dimension information. Based on the time-series dynamic graph, the path node feature information and graph edge attributes of each node in the time-series dynamic graph are used as node input features to record the path index, sequence position and adjacency relationship of each node. The improved temporal dynamic graph attention network was trained using a combination of node embedding difference error, path embedding consistency error, path history embedding fusion error, node order preservation error, and adjacency information preservation error as joint optimization objectives. The network parameters of the node embedding adaptive grouping layer, path-dependent self-looping layer, and path history embedding enhancement layer were continuously optimized. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved temporal dynamic graph attention network was considered to have completed convergence training.

[0011] Optionally, the generation of the node embedding matrix includes: In the node embedding adaptive grouping layer, feature similarity calculation is performed on the nodes of each path. Based on the feature similarity, the nodes are divided into several subgroups. Within each subgroup, information aggregation is performed according to the node order and feature similarity relationship to generate the subgroup node embedding representation. The initial node embedding matrix of the entire path is formed by combining all subgroup node embeddings in the order of path nodes. Batch normalization and nonlinear transformation are performed on the initial node embedding matrix of each path to generate the final node embedding matrix.

[0012] Optionally, the generated path embedding matrix includes: Input a node embedding matrix, arrange the node embeddings of each path according to the path number and node order, and record the path index, sequence position and adjacent node information of each node; In the path-dependent self-loop layer, information iterative update operation is performed on the node embedding matrix of each path. In each iteration, the current node embedding and the adjacent node embedding are sequentially transferred, merged and weighted along the path, and the node embedding change state is recorded to form a closed-loop representation of the node embedding within the path. The node embedding within the repeated path is iterated in a closed loop for several rounds. In each round of iteration, the node order is kept consistent with the path sequence, and the node embedding state is updated until convergence, generating the stable state of the node embedding for each transportation path and forming the initial path embedding representation. The initial path embedding representation of each path is organized and serialized. The node embeddings within the path are arranged in the transportation order, and the node index, adjacency relationship and path index are recorded to form a complete path embedding matrix.

[0013] Optionally, the generated path history embedding matrix includes: Input the path embedding matrix for each path, arrange the node embeddings according to the path node order, and record the path index, sequence position and node embedding state of the previous iteration for each node; In the path history embedding enhancement layer, the node embedding of each path is iteratively updated. The current round path embedding is aligned with the corresponding path embedding of the previous round according to the node order and then fused. A preliminary path history embedding representation is generated through information aggregation. Multiple iterations are performed on the initial path history embedding representation. In each iteration, the node order is kept consistent with the path sequence. After each iteration, the node embedding state is updated until the node embedding state is stable. The stable path history embedding representation is normalized to eliminate numerical differences, and a nonlinear transformation is applied to generate an updated path history embedding representation, while maintaining the integrity of node order and path sequence information. The updated path history embedding representation for each path is organized and serialized, and the index of each node and the relationship between adjacent nodes are recorded to generate a path history embedding matrix.

[0014] Optionally, determining the steel component transportation and scheduling plan specifically includes: The path history embedding matrix is ​​used as input. The node embedding of each path is arranged according to the path number and node order. The individual spider state in the social spider optimization algorithm is initialized, and the path selection, position parameters and path history embedding information of each individual are recorded. The female and male spiders in the group are divided into groups according to a dynamic role switching strategy, the group structure is established, and each individual is assigned an iteration step size, vibration sensing parameters and search state identifier to form the initial state set of the group. In each iteration, the position update operation is performed on each spider individual based on the path history embedding matrix information. The individual path position is adjusted by the node order and path embedding relationship. At the same time, the individual vibration intensity is calculated, and the vibration information is transmitted through the group propagation mechanism to generate the updated spider individual state. Based on the updated individual spider status, perform multi-objective evaluation, including transportation time, path length, load constraints and road condition adaptability. Update the individual's priority path selection and iteration step size based on the multi-objective evaluation results, and save the current group's optimal path combination. Repeatedly perform individual spider position updates and multi-objective evaluation operations until the optimal path combination in the group does not change in continuous iterations, forming the optimal node embedding order for each path and generating the group's optimal path state set; Organize the path information of each spider individual in the optimal path state set of the group and the corresponding multi-objective evaluation results, arrange the optimal node embedding order of each path according to the path number, and generate a scheduling search result matrix; Read the scheduling search result matrix and obtain the node information and path combination information for each path according to the path number and node order; Based on the node information and route combination information of each route, the transport vehicles are matched with the routes, a driving node sequence for each vehicle is generated, and the vehicle number, node order and route index are recorded to form a preliminary vehicle route plan. Schedule the travel node sequence of each transport vehicle, determine the estimated arrival and departure times of each node, record the time information in relation to the vehicle number and route index, and generate a vehicle route timetable. The preliminary vehicle route plan and vehicle route timetable were compiled, and the node sequence and time arrangement of all transport vehicles were merged to form a complete steel component transportation scheduling plan, including the driving route and corresponding time arrangement of each vehicle.

[0015] The beneficial effects of this invention are: This invention realizes an intelligent scheduling system for steel component transportation by constructing an improved temporal dynamic graph attention network and a social spider optimization algorithm through deep coupling. Compared with traditional transportation scheduling methods that rely on human experience or static planning, this invention can dynamically model transportation task data, integrating transportation nodes, path attributes, vehicle states, and historical transportation information into the network structure. Through node embedding adaptive grouping, path dependency self-looping, and path history embedding enhancement mechanisms, it accurately characterizes the temporal changes and path features during steel component transportation, providing comprehensive data representation for subsequent scheduling optimization. The improved temporal dynamic graph attention network can dynamically update node and path embeddings in multiple iterations, realizing continuous learning of transportation node states and effective utilization of historical information, thus improving the accuracy and continuity of transportation state modeling.

[0016] This invention utilizes a swarm spider optimization algorithm to perform a global multi-objective search on the path history embedding matrix. Through dynamic role switching, adaptive iteration step size updates, and a swarm vibration information propagation mechanism, it achieves global optimization of transportation routes and vehicle scheduling. The optimization process simultaneously considers transportation time, path length, vehicle load constraints, and road condition adaptability, ensuring that the generated scheduling search result matrix comprehensively reflects the dynamic characteristics and multi-objective constraints of each transportation route. The steel component transportation scheduling scheme generated by this invention maintains high consistency and continuity in node sequence, path arrangement, and vehicle time planning, and can dynamically adjust and adapt to the complex conditions of the construction site, achieving optimal allocation of transportation resources. This method significantly overcomes the shortcomings of traditional methods, such as weak static and local optimization capabilities and insufficient utilization of historical data, improving the accuracy, global optimization capability, and dynamic adaptability of transportation scheduling. It provides a comprehensive technical effect of intelligent, efficient, and sustainable optimization for steel component transportation at construction sites. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a steel component transportation route scheduling method based on big data analysis proposed in this invention; Figure 2 This is a schematic diagram of an improved temporal dynamic graph attention network for a steel component transportation path scheduling method based on big data analysis proposed in this invention. Figure 3 This is a schematic diagram of the gregarious spider optimization algorithm for a steel component transportation route scheduling method based on big data analysis proposed in this invention, applied in multi-objective path search. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A method for scheduling the transportation routes of steel components based on big data analysis, comprising: Collect transportation task data for steel components, preprocess the transportation task data, and generate standardized transportation task sequences; Based on the standardized transportation task sequence, the paths and nodes are identified, and the node status, vehicle status and path attributes of each path are extracted as path node features to construct a time-series dynamic graph. Based on temporal dynamic graphs, an improved temporal dynamic graph attention network is constructed, including a node embedding adaptive grouping layer, a path-dependent self-looping layer, and a path history embedding enhancement layer. In the node embedding adaptive grouping layer, nodes are divided into subgroups based on feature similarity to generate a node embedding matrix; In the path-dependent self-loop layer, the node embedding matrix of each path is iteratively updated, information is passed between nodes within the path, and a path embedding matrix is ​​generated. In the path history embedding enhancement layer, the path embedding matrix of each path is fused with the corresponding path embedding matrix of the previous iteration, aligned according to the path node order, the node embedding state is updated, and a path history embedding matrix is ​​generated. The algorithm for optimizing the social spider population is based on the path history embedding matrix. Spider individuals are initialized, and female and male individuals are divided according to a dynamic role switching strategy. The individual positions are updated using an iterative step size adaptive update rule. Vibration information is propagated in the group to perform a global multi-objective path search, generate scheduling search results, and determine the steel component transportation scheduling plan, including the routes and time arrangements of each transport vehicle.

[0020] In this embodiment, the transportation task data specifically includes transportation node information, transportation vehicle information, steel component loading information, transportation route attribute information, transportation time identifier data, and historical transportation record identifier data.

[0021] In this embodiment, the preprocessing of the transportation task data includes: Identify and remove abnormal information in transportation task data, including missing node status, abnormal vehicle status, or abnormal route attribute data. Missing node status, vehicle status, or path attribute data are filled in by interpolating preceding and following nodes to generate complete information; The node status and vehicle status are sorted according to the transportation node number, vehicle number and task order to form a standardized node list arranged in transportation order. The transportation route attributes are standardized, including route length, capacity, traffic restrictions, and estimated transportation time, and each route attribute is mapped to a unified dimension. The transportation time identification data is corrected to ensure that the time sequence from the origin to the destination is consistent with the node sequence, abnormal time records are corrected, and a standardized transportation time sequence is formed.

[0022] In this embodiment, the construction of the timing dynamic graph specifically involves: Based on the standardized transportation task sequence, extract the information of each path and node, number them according to the transportation sequence, and generate a path node list; Based on the list of path nodes, vehicle status characteristics are assigned to each path node, including vehicle type, load and current location status, to form path node status information; Extract the attribute information of each path, including path length, traffic capacity, traffic restrictions and estimated transportation time, and associate it with the corresponding path nodes to generate path attribute features; The status information of each path node is combined with the path attribute information to generate a path node feature set, while keeping the node order consistent with the path. A time-series dynamic graph is constructed based on the feature set of path nodes. Path nodes are represented as graph nodes, and the connection relationships between nodes and paths are represented as graph edges. Path attribute information is written into the corresponding graph edge attributes to generate an initial time-series dynamic graph. Assign a unique identifier to each node and edge in the initial time-series dynamic graph, and record the path node features and edge attributes on the corresponding nodes and edges to form a complete time-series dynamic graph.

[0023] In this embodiment, the construction of the improved temporal dynamic graph attention network specifically involves: An improved temporal dynamic graph attention network is constructed, comprising a node embedding adaptive grouping layer, a path-dependent self-looping layer, and a path history embedding enhancement layer. The parameters of each layer are initialized, including weights, biases, and node embedding vector dimension information, wherein: An improved temporal dynamic graph attention network is constructed, specifically as follows: Based on the traditional temporal dynamic graph attention network, the basic node embedding calculation module, path information transmission module, and temporal feature capture module are retained. The node embedding calculation module is expanded into a node embedding adaptive grouping layer, which performs similarity analysis and grouping on the path node feature vectors. The path information transmission module is expanded into a path-dependent self-looping layer, which performs sequential iterative updates on the path node embeddings. A new path history embedding enhancement layer is added and connected to the output of the node embedding adaptive grouping layer and the path-dependent self-looping layer to align and fuse the current round path embedding with the previous round history embedding. The three layers are connected sequentially, and the node embedding vector output is passed to the path-dependent self-looping layer. The path embedding output is passed to the path history embedding enhancement layer, forming an improved temporal dynamic graph attention network. Initialize the parameters of each layer as follows: The node embedding vector dimension is set to 64 for the adaptive grouping layer. The feature vector weight matrix of each node is initialized to a normal distribution with a mean of 0 and a standard deviation of 0.02. The bias vector is initialized to zero. In the subgrouping strategy, nodes are divided into groups of 8 to 12 nodes based on feature similarity. The information aggregation weight within each subgroup is initialized to a uniform distribution of 0.1 to 0.5. The number of iterations for the path-dependent self-looping layer is set to 10. The node information transfer weight matrix is ​​initialized to a normal distribution with a mean of 0 and a standard deviation of 0.02. The non-linear activation function is ReLU. Each node embedding vector is input into the ReLU activation function after weighted summation and bias operation. The output is max(0,x), with negative values ​​set to zero and positive values ​​kept. The output dimension of the node embedding is consistent with the input dimension. The node information weighting coefficients are initialized to a uniform distribution in the range of 0.3 to 0.7. The initial value of the fusion coefficient vector for the path history embedding enhancement layer is set to 0.5, and the node embedding normalization parameter is initialized to a unit vector. Specifically, for the embedding vector of each node, the values ​​of each dimension of the embedding vector are adjusted according to the proportion of the overall length so that the overall length of the vector in the initial state is 1, ensuring that the initial scale of all node embedding vectors is consistent. Based on the time-series dynamic graph, the path node feature information and graph edge attributes of each node in the time-series dynamic graph are used as node input features to record the path index, sequence position and adjacency relationship of each node. An improved temporal dynamic graph attention network was trained using a combination of node embedding difference error, path embedding consistency error, path history embedding fusion error, node order preservation error, and adjacency information preservation error as joint optimization objectives. The network parameters of the node embedding adaptive grouping layer, path-dependent self-looping layer, and path history embedding enhancement layer were continuously optimized. The improved temporal dynamic graph attention network was considered to have completed convergence training when the change in the joint loss value over five consecutive training rounds was less than 0.001. The improved temporal dynamic graph attention network is trained as follows: The standardized node feature sequence, node embedding matrix, and path history embedding matrix of each path are input into the improved temporal dynamic graph attention network. The output is the node embedding prediction value, path embedding prediction value, and historical embedding prediction value for each path. The node embedding difference error is obtained by averaging the squared differences between the node embedding prediction value and the true node embedding matrix element by element. The path embedding consistency error is obtained by averaging the squared differences between the path embedding prediction value and the true path embedding matrix element by element. The path history embedding fusion error is obtained by averaging the squared differences between the path history embedding prediction value and the true history embedding matrix element by element. The node order preservation error is obtained by averaging the squared node order deviation. The adjacency information preservation error is obtained by averaging the squared adjacency information difference. The errors are summed according to their weights to form a joint loss value: node embedding difference error 0.30, path embedding consistency error 0.25, path history embedding fusion error 0.20, node order preservation error 0.15, and adjacency information preservation error 0.10. The joint loss value is used to calculate the gradient during backpropagation, and the parameters of the node embedding adaptive grouping layer, path dependency self-looping layer, and path history embedding enhancement layer are updated. The gradient is multiplied by the learning rate of 0.001 and then subtracted from the current parameters to generate the updated network parameters, which are then written back to the corresponding layer. The training process uses a batch size of 64 paths, with 5–15 nodes per path and a node embedding vector dimension of 64. Iterates for 50–100 rounds, updating parameters using the complete training set in each round. When the change in the joint loss value is less than 0.001 for 5 consecutive rounds, it is determined that the improved temporal dynamic graph attention network has converged and the network parameters are stable.

[0024] In this embodiment, generating the node embedding matrix specifically involves: In the node embedding adaptive grouping layer, feature similarity calculation is performed on the nodes of each path. Based on feature similarity, the nodes are divided into several subgroups. Within each subgroup, information aggregation is performed according to the node order and feature similarity relationship to generate subgroup node embedding representations, where: The node embedding adaptive grouping layer includes: Feature Similarity Calculator: Calculates the feature similarity between each node and other nodes along the same path, generating similarity information; Subgroup partitioning table: Divides nodes into several subgroups based on similarity information and records the order of nodes within each group; Information aggregation queue: Receives node embedding vectors within subgroups in node order, performs cumulative and weighted processing, and generates subgroup node embedding representations; Normalization unit: Normalizes the embedding vectors of subgroup nodes to maintain consistent vector scale; Output cache table: Stores the node embedding representation of each subgroup in the original path node order; In the node embedding adaptive grouping layer, the node similarity information generated by the feature similarity calculation unit is input into the subgroup partitioning table. The subgroup partitioning table records the subgroup to which the node belongs and the order within the group. The subgroup information aggregation queue receives the node order and the corresponding node embedding vector in the subgroup partitioning table and generates the subgroup node embedding representation. The node embedding normalization unit receives the subgroup node embedding representation, normalizes the vector, and outputs the normalized subgroup node embedding representation to the output buffer table, arranged according to the original path node order. For each path's nodes, perform feature similarity calculations, specifically: The embedding vector of each node in the path is compared with the embedding vectors of other nodes in the same path one by one. The similarity of each pair of nodes is calculated by multiplying the corresponding values ​​of each dimension of the embedding vectors of the two nodes, summing them, and dividing by the product of the square root of the sum of the squares of the values ​​of the two node embedding vectors. Perform feature similarity calculation on all nodes within the path to form a node similarity matrix, where the rows and columns of the node similarity matrix correspond to the path node numbers, and the elements record the similarity values ​​between node pairs; Nodes are divided into several subgroups based on feature similarity, specifically: Read the node similarity matrix and sort the nodes in each path from high to low according to the sum of their similarity scores. The sorting is based on the sum of the similarity scores between each node and other nodes in the path. Nodes are assigned to subgroups in order of sorting results. Each subgroup contains 8 to 12 nodes. When assigning nodes, if the subgroup capacity reaches 12, a new subgroup is created. When there are fewer than 8 nodes remaining, the remaining nodes are evenly distributed to the existing subgroups in order. Each subgroup can add a maximum of two nodes. During the allocation process, the path index, original path sequence position, and sequence position within the subgroup of each node are recorded, and the embedding vector index and node similarity vector information of each node are appended to the subgroup partitioning table. The entire subgroup partitioning table is sorted in path number order. Each record includes the subgroup number, the number of nodes contained in the subgroup, and the embedding vector index of each node in the subgroup. Performing information aggregation operations, specifically: Each subgroup receives node embedding vectors according to the node order recorded in the subgroup partitioning table. For each node embedding vector in the subgroup, a weighted sum is performed on the corresponding dimension of the embedding vectors of other nodes in the same group, with the weight values ​​being uniformly distributed from 0.1 to 0.5. The initial aggregated vector is obtained by weighted averaging the cumulative result of the node embedding vectors within the subgroup and the node's own embedding vector. The initial aggregated vector is then input into the normalization unit in node order. The values ​​of each dimension of the vector are divided by the square root of the sum of the squares of the overall length to form a standardized embedding vector with a length of 1. After normalizing all nodes within a subgroup, the normalized embedding vectors are arranged in subgroup order and combined to form a subgroup node embedding representation, recording the path index, original sequence position, and subgroup sequence position of each node. The initial node embedding matrix of the entire path is formed by combining all subgroup node embeddings in the order of path nodes. Batch normalization and nonlinear transformation are performed on the initial node embedding matrix of each path to generate the final node embedding matrix.

[0025] In this embodiment, the generation of the path embedding matrix specifically refers to: Input a node embedding matrix, arrange the node embeddings of each path according to the path number and node order, and record the path index, sequence position and adjacent node information of each node; In the path-dependent self-looping layer, an iterative information update operation is performed on the node embedding matrix of each path. In each iteration, the current node embedding and the embeddings of adjacent nodes are sequentially combined, information is passed, fused, and weighted along the path. The changing state of the node embedding is recorded to form a closed-loop representation of the node embedding within the path, where: The path-dependent self-loop layer includes: Node embedding register array: Stores the embedding vector and node index information of each node in the order of path nodes; Adjacency information cache table: records the indexes of each node's preceding and following neighbor nodes and their connection weights; Weighted fusion unit: performs element-wise combination of the current node's embedding vector and the embedding vectors of its neighboring nodes according to a weighted combination rule; Iteration state register: stores the changes in the node embedding vector and the cumulative update value in each iteration; Normalization and Nonlinear Mapping Unit: Receives the weighted and fused node embedding vector, normalizes it, and maps it through the ReLU activation function to form the node embedding output of the current iteration; Output cache table: Stores the node embedding vectors after each iteration in the order of path nodes; In the path-dependent self-looping layer, the node embedding vector stored in the node embedding register array, together with the neighbor node index and weight recorded in the adjacency information cache table, is input into the weighted fusion unit to generate a weighted combined node embedding vector. The weighted combined node embedding vector is then passed to the normalization and nonlinear mapping unit for vector normalization and ReLU mapping. The output node embedding vector is written to the output cache table, and the iteration state register is updated to record the node embedding change information. In the next iteration, the node embedding vector in the output cache table is again used as input, and combined with the adjacency information to repeat the weighted fusion and normalization mapping process until all iterations are completed, forming a closed-loop representation of the node embedding within the path. The current node embedding and the neighboring node embeddings are used to exchange information, specifically as follows: Read the node embedding vector and node index of each path sequentially from the node embedding register array, and look up the indexes of the forward and backward neighbors of the current node according to the adjacency information cache table; In the node embedding register array, locate and read the corresponding neighbor node embedding vector by index, and arrange each neighbor node embedding and the current node embedding in sequence order to form a node embedding sequence; The node embedding sequence is written back to the node embedding register array in sequence. The node embedding register array stores the embedding vector combination of each node and its preceding and following neighbors in the order of the path nodes, keeping the path order and node index consistent. After collecting the embedding vector information of all nodes in the path, a set of node information transfer for each path is formed. The fusion and weighted combination method is as follows: Read the nodes and forward and backward neighbor embedding vectors of each path in the order of the path. Arrange the 64 dimensions of the current node embedding vector and the 64 dimensions of each neighbor embedding vector element by element to form a node embedding combination matrix. Each row corresponds to the current node and a neighbor. The total number of rows in the matrix is ​​equal to the number of neighbors plus 1. In the weighted fusion unit, an element-level weighted summation operation is performed on each column of the node embedding combination matrix. The current node embedding weight is fixed at 0.6, the forward neighbor embedding weight is 0.2, and the backward neighbor embedding weight is 0.2. The weighted summation result is written into the node temporary fusion vector. The temporary fusion vector of each node is arranged in the order of the path nodes to generate a path fusion matrix. The number of rows in the path fusion matrix is ​​equal to the number of nodes, and the number of columns is 64. Each row records the fusion embedding information of the node and its neighbors. The path fusion matrix is ​​iterated 10 times. In each iteration, the weighted summation and path fusion matrix update operations are repeated, and the change magnitude of the embedding vector of each node is recorded in the iteration state register. Each path fusion matrix generated through iteration is mapped to a node index and arranged in the order of the paths to form complete path embedding fusion data; The path node embedding closed-loop representation is formed as follows: After the path iteration is completed, the node embedding vectors updated in each iteration are arranged in the order of the path nodes. The connection relationship between the first and last node embedding vectors is mapped to the adjacency relationship between the last node and the first node in the path sequence to obtain the first and last mapping information. Integrate all node embedding vectors and start-end mapping information to generate a closed node embedding sequence, ensuring that the embedding connections at the start and end of the path remain consistent; Write the closed node embedding sequence of each path back into the node embedding register array, arrange them in order of path number, and form the node embedding closed-loop representation of each path. The node embedding within the repeated path is iterated in a closed loop for several rounds. In each round, the node order is kept consistent with the path sequence, and the node embedding state is updated until convergence, generating a stable node embedding state for each transportation path, forming the initial path embedding representation, where: Update the node embedding state until convergence, specifically as follows: In each iteration, for each node in the path, the element-wise difference between the current node embedding vector and the corresponding node embedding vector in the previous iteration is calculated, the sum of squares is taken and the square root is obtained to obtain the change in the node embedding vector. The average change value of all nodes in the path is calculated to form the average change value of the path. When the average change value of the path is less than 0.01 in three consecutive iterations, the node embedding is considered to be converged. The initial path embedding representation of each path is organized and serialized. The node embeddings within the path are arranged in the transportation order, and the node index, adjacency relationship and path index are recorded to form a complete path embedding matrix.

[0026] In this embodiment, the generation of the path history embedding matrix includes: Input the path embedding matrix for each path, arrange the node embeddings according to the path node order, and record the path index, sequence position and node embedding state of the previous iteration for each node; In the path history embedding enhancement layer, the node embeddings of each path are iteratively updated. The current round path embeddings are aligned with the corresponding path embeddings of the previous round in node order and then fused. A preliminary path history embedding representation is generated through information aggregation, where: The path history embedding enhancement layer includes: Path embedding register array: Stores the embedding vector and node index of the current round node for each path in the order of path nodes; Historical embedding register table: records the embedding vector and index information of the path nodes corresponding to the previous iteration; Node alignment table: Align the current round node embedding vector with the corresponding node embedding vector of the previous round according to the node order, and generate a list of node pair indices; Information aggregation queue: Receives aligned node embedding vector pairs and arranges the node embedding vectors according to the node order to form a temporary aggregation sequence; Fusion coefficient register array: stores the coefficient values ​​of each node's embedding vector during the fusion process; Preliminary path history embedding cache table: Stores the node embedding vectors after each round of fusion in the order of path nodes to form a preliminary path history embedding representation; In the path history embedding enhancement layer, the current round node embedding vector in the path embedding register array is paired with the previous round node embedding vector in the history embedding register table according to the index in the node alignment table to generate a node embedding vector pair sequence and input it to the information aggregation queue. The information aggregation queue performs weighted combination on each node pair according to the coefficient value in the fusion coefficient register array to form a temporary fusion vector. The fused node embedding vector is written to the preliminary path history embedding cache table in the order of path nodes. The current round path embedding is aligned with the corresponding path embedding of the previous round iteration according to the node order and then merged. Specifically: The node embedding vectors of the current round path are read sequentially from the path embedding register array. Each node embedding vector has a 64-dimensional dimension and records the node index and sequence position. Read the embedding vector and index information of the corresponding node in the previous iteration from the historical embedding register table. The embedding vector of each historical node has a dimension of 64, and the sequence position corresponds to the index of the node in the current round. In the node alignment table, the current round node index is matched with the historical node index to form a node pair index list. The length of the node pair index list generated for each path is equal to the number of nodes in the path. Each row contains a combined vector of the current node embedding vector and the historical node embedding vector, with a total length of 128 dimensions. Arrange all node pairs in the order of path nodes to form a temporary node pair sequence for each path. Record the path index, sequence position and offset of each node in the temporary node pair sequence. A preliminary path history embedding representation is generated through information aggregation, specifically as follows: The temporary node pair sequence input information of each path is aggregated into a queue, and each node pair is processed one by one according to the order of the temporary node pair sequence. For each node pair, the current round node embedding vector and the previous round node embedding vector are weighted and combined according to the fusion coefficient vector. The current node embedding weight is set to 0.5, the historical node embedding weight is set to 0.5, and each node embedding vector has a dimension of 64. After combination, a fusion vector with a length of 64 dimensions is formed. Arrange all fusion vectors in the order of path nodes to form a temporary path fusion matrix. The number of rows is equal to the number of path nodes, and the number of columns is 64. Each row records the node and the corresponding historical embedded weighted vector information. Write the temporary path fusion matrix into the preliminary path history embedding cache table, keeping the node order and path index consistent, to form the preliminary path history embedding representation; Multiple iterations are performed on the initial path history embedding representation. In each iteration, the node order is kept consistent with the path sequence. After each iteration, the node embedding state is updated until the node embedding state is stable. The stable path history embedding representation is normalized to eliminate numerical differences, and a nonlinear transformation is applied to generate an updated path history embedding representation, preserving the integrity of node order and path sequence information. The updated path history embedding representation is generated as follows: The stable path history embedding representation is input into the normalization unit in the order of path nodes. For the 64-dimensional embedding vector of each node, the square root of the sum of squares of each dimension of the vector is calculated. The value of each dimension is divided by the norm of the vector so that the overall length of the vector is 1. The normalized node embedding vectors are sequentially input into the nonlinear mapping unit and processed using the ReLU activation function. Max(0,x) is calculated for each dimension, negative values ​​are set to zero, and positive values ​​are kept in their original values ​​to form the node embedding vectors after nonlinear mapping. The node embedding vectors after nonlinear mapping are arranged in the order of path nodes to form a path normalization fusion matrix with the number of rows equal to the number of path nodes and the number of columns being 64. Each row records the updated corresponding node embedding vector. Write the path normalization fusion matrix into the path history embedding cache table to maintain the integrity of node order and path sequence information, and form the updated path history embedding representation. The updated path history embedding representation for each path is organized and serialized, and the index of each node and the relationship between adjacent nodes are recorded to generate a path history embedding matrix.

[0027] In this embodiment, generating a determined steel component transportation and scheduling plan specifically involves: Using the path history embedding matrix as input, the node embeddings of each path are arranged according to path number and node order. This initializes the individual spider states in the gregarious spider optimization algorithm, recording the path selection, position parameters, and path history embedding information of each individual. Initialize the individual spider states in the social spider optimization algorithm as follows: The number and order of nodes in each path are determined based on the path history embedding matrix, and a path selection vector is generated for each spider individual, with each element being the index of the corresponding path node. The initial chaotic sequence is generated iteratively using the Logistic chaotic mapping function. The iterative relationship is: the next value is equal to the difference between the mapping coefficient, the current value, and 1. The mapping coefficient is 4, and the initial value is randomly selected from the range of [0.1, 0.9]. The initial chaotic sequence is linearly mapped proportionally to the path node index range. That is, each chaotic number is multiplied by the difference between the maximum and minimum index of the path node, plus the minimum index of the path node, to form the initial position vector of each spider individual. The length is the same as the path selection vector, and each element corresponds to the numerical representation of the path node index. Based on the reverse learning method, a reverse position vector is calculated for each initial position vector. The reverse position vector is equal to the sum of the minimum and maximum values ​​of the path index minus the current path index value. The corresponding elements of the forward position vector and the reverse position vector are averaged to obtain the final position vector. Each element is scaled according to the path node index range and corresponds to the path history embedding matrix. Generate a velocity vector with the same length as the position vector. Each element is equal to the absolute value of the difference between the current position vector and the corresponding reverse position vector, multiplied by a random coefficient in the range [0.05, 0.12]. Set the vibration threshold to be uniformly distributed in the range of [0.02, 0.04]. The historical best position vector is equal to the final position vector. The path history embedding index corresponds one-to-one with the path history embedding matrix. All spider individual states are arranged according to path number and node order to form a complete set of spider individual states. The spiders in the group were divided into female and male individuals according to a dynamic role switching strategy. The group structure was established, and each individual was assigned an iteration step size, vibration sensing parameters, and search state identifier, forming the initial state set of the group, where: The dynamic role switching strategy is as follows: Read the current position vector and the historical best position vector of each spider individual. The length of each vector is equal to the number of paths. For the i-th individual, subtract each element of the current position vector from the corresponding element of the historical best position vector to obtain the element difference sequence. Square each element in the element difference sequence, sum them, and then take the square root to obtain the individual's difference value. The variance values ​​of all individuals in the population are sorted from largest to smallest. The top 60% of individuals are marked as female and the remaining 40% as male. Before each iteration, the individual variance values ​​are recalculated, and the population variance mean and standard deviation are calculated. The individual variance values ​​are compared with the population variance mean plus or minus the standard deviation. Individuals with variance values ​​greater than the mean plus the standard deviation are marked as female, and those with variance values ​​less than the mean minus the standard deviation are marked as male. After the role label is updated, the iteration step size is linearly scaled to the range of [0.05, 0.12] according to the ratio of individual difference value to population mean, and the vibration sensing parameters are linearly mapped to the range of [0.02, 0.04] according to the ratio of individual difference value to population standard deviation. Each individual records the path selection index, position vector and node embedding index, which correspond to the path history embedding matrix. Each individual is assigned an iteration step size, vibration sensing parameters, and search state identifier, specifically: Based on the relationship between individual differences and the mean and standard deviation of group differences, the iteration step size is linearly mapped to the range [0.05, 0.12], and the calculation formula is as follows: Iteration step size = 0.05 + (individual difference value - minimum group difference) × (0.12 - 0.05) / (maximum group difference - minimum group difference); Individual differences and population standard deviations are linearly mapped proportionally to the vibration sensing parameter range [0.02, 0.04], and the mapped value for each individual is recorded as the vibration sensing parameter. The search status identifier is generated by comparing the individual difference value with the group mean. If the difference value is greater than the mean plus the standard deviation, it is marked as an active state; if it is less than the mean minus the standard deviation, it is marked as a pending activation state; and the rest are marked as normal states. In each iteration, a position update operation is performed on each spider individual along a path based on the path history embedding matrix information. The individual path position is adjusted according to the node order and path embedding relationship. At the same time, the individual vibration intensity is calculated, and the vibration information is transmitted through the group propagation mechanism to generate the updated spider individual state, where: Perform a position update operation on each spider individual along the path, specifically: Read the node embedding vector of each path in the order of the path history embedding matrix, and match it with the corresponding vector of the individual current path selection; Each element of the current position vector is combined with the corresponding element of the historical best position vector in a weighted linear combination. The update formula is that the new element value is equal to the current element value multiplied by 0.7, plus the corresponding element value of the historical best position multiplied by 0.3. Each element of the updated position vector is adjusted by adding the value of the corresponding element of the velocity vector multiplied by the iteration step size, where the iteration step size is a random value of [0.05, 0.12]. Arrange the updated position vectors in the order of path nodes to generate new path position vectors for each individual, with each element corresponding to a path node index. Boundary constraints are applied to the path position vector, adjusting elements that exceed the minimum and maximum node indices to the corresponding range to ensure that the path position vector and path node indices are perfectly matched. The individual vibration intensity is calculated as follows: Read the node position elements of each path in the order of the path position vector; Calculate the absolute value of the difference between each element of the current path position vector and the historical best path position vector, and sum all the differences to obtain the overall path position deviation value. Divide the overall positional deviation of the path by the number of path nodes to obtain the average positional deviation of the path. The vibration intensity is equal to the average positional deviation of the path multiplied by a random coefficient, which takes a uniform value between [0.02, 0.04]. The vibration intensity of each path is summarized to form a vibration intensity vector of the individual spider, with each element corresponding to one path, and the path numbering order is kept consistent. Vibration information is transmitted through a group propagation mechanism, specifically as follows: Arrange the vibration intensity vector of each spider in the group in order of path number; For each individual, calculate the absolute value of the difference between the path position vector and that of other individuals in the group, multiply the absolute value of the difference by the vibration intensity of the corresponding individual, and then sum them up to obtain the propagation coefficient of each path position; The vibration information update of the current individual is obtained by averaging the propagation coefficient at each path location; The vibration information update is added to the vibration intensity vector of the current individual to form the updated vibration intensity vector, while keeping the path numbering order unchanged; Based on the updated individual spider states, a multi-objective evaluation is performed, including transportation time, path length, load constraints, and road condition adaptability. The individual's path selection vector and iteration step size are updated based on the multi-objective evaluation results, and the optimal path combination information for the current group is saved. Perform a multi-objective assessment, specifically: For each spider individual's path selection vector, the corresponding path's transportation time, path length, load, and road condition adaptability index are input into a multi-objective evaluation function. Each index value is standardized to 0–1, where the transportation time is normalized to the interval [0,1], the path length is normalized to the interval [0,1], the load constraint exceeding the vehicle's maximum load is normalized to 1, and the rest are mapped to 0–1 proportionally. The road condition adaptability score is linearly mapped to 0–1 in the interval [0.7,1.0]. For each individual spider, a multi-objective comprehensive score is calculated according to the following weights: transportation time multiplied by 0.35, path length multiplied by 0.25, load constraint multiplied by 0.25, and road condition adaptability multiplied by 0.15. The four weighted values ​​are summed to obtain the comprehensive score vector of the path corresponding to the individual spider. Each element corresponds to one path in the path selection vector. Sort the comprehensive score vectors from high to low to generate a path index sorting list. Rearrange the path selection vectors according to the path index sorting list to form the priority path selection vector for the corresponding spider individual in the current iteration. The path selection vector and iteration step size of each individual are updated based on the multi-objective evaluation results, specifically as follows: For each spider individual, the priority path selection vector of the current iteration is mapped to the comprehensive score vector, and the comprehensive score of each path determines the ranking position of the path in the priority path selection vector. Based on the sorting results, the path selection vector is regenerated to ensure that the path order in the path selection vector is consistent with the overall score, and each element corresponds to the path number and path node index information. Based on the overall path score and the magnitude of the ranking position change, update the iteration step vector. Each element corresponds to the path index in the path selection vector. The iteration step range is set to 0.01–0.06. The iteration step of each path element = 0.01 + (0.06–0.01) × (1 – corresponding path overall score / highest path score). Keep the length of the iteration step vector consistent with the length of the path selection vector. Repeatedly perform individual spider position updates and multi-objective evaluation operations until the optimal path combination in the group does not change in continuous iterations, forming the optimal node embedding order for each path and generating the group's optimal path state set; Organize the path information of each spider individual in the optimal path state set of the group and the corresponding multi-objective evaluation results, arrange the optimal node embedding order of each path according to the path number, and generate a scheduling search result matrix; Read the scheduling search result matrix and obtain the node information and path combination information for each path according to the path number and node order; Based on the node information and route combination information of each route, the transport vehicles are matched with the routes, a driving node sequence for each vehicle is generated, and the vehicle number, node order and route index are recorded to form a preliminary vehicle route plan. Schedule the travel node sequence of each transport vehicle, determine the estimated arrival and departure times of each node, record the time information in relation to the vehicle number and route index, and generate a vehicle route timetable. The preliminary vehicle route plan and vehicle route timetable were compiled, and the node sequence and time arrangement of all transport vehicles were merged to form a complete steel component transportation scheduling plan, including the driving route and corresponding time arrangement of each vehicle.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to the transportation and scheduling of steel components at a large steel structure construction site. The system received transportation task data, including 50 transport vehicles, 120 transport routes, and 450 transport nodes. Each route contained 3 to 7 nodes, with an average distance of approximately 2.3 kilometers between nodes. Some routes had capacity restrictions (approximately 12%), and the load capacity of the transport vehicles ranged from 5 to 30 tons, with a total transported steel component weight of 680 tons. The original data had approximately 5% missing node status and approximately 8% inconsistent route attributes. After standardization processing, including anomaly removal, missing information filling, node and vehicle status sorting, route attribute normalization, and transportation time identifier correction, data integrity reached 100%, and the node status order error rate was reduced to 0.

[0029] Standardized data is mapped to a temporal dynamic graph, where nodes represent transportation nodes and edges represent transportation paths and their attributes. Each node includes vehicle type, load, and current location status, and each path is associated with path length, capacity, traffic restrictions, and estimated transportation time. An improved temporal dynamic graph attention network performs adaptive grouping of node embeddings on 450 nodes, dividing them into 45 subgroups, each containing approximately 10 nodes. Within each subgroup, information is aggregated based on node order and feature similarity, with each node embedding vector having a dimension of 64, generating an initial node embedding matrix. Subsequently, a path-dependent self-looping layer performs 10 iterations of updates. In each iteration, the current node embedding is fused with the embeddings of neighboring nodes along the path order, and a weighted combination is formed to create the path embedding matrix. After a single iteration, the average change in node embeddings is approximately 0.045, and after 10 iterations, the change in node embeddings converges to 0.014. The path history embedding enhancement layer further aligns and merges the path embeddings from the previous iteration with the node embeddings from the current iteration. After 5 iterations, a stable path history embedding matrix is ​​generated, and the average change in node embedding status decreases from the initial 0.27 to 0.013, forming stable historical embedding data that can be used for population optimization.

[0030] The social spider optimization algorithm initializes 100 spider individuals (60 females and 40 males), with an initial iteration step size ranging from 0.05 to 0.12 and a vibration sensing threshold of 0.03. In each iteration, individual spiders adjust their path order based on the path history embedding matrix. Each path embedding vector has a length of 64, with the corresponding node number and order strictly maintained. After calculating individual vibration intensity, the state of other individuals is updated through group propagation. Multi-objective evaluation includes transportation time, path length, load constraints, and road condition adaptability, with evaluation scores ranging from 0 to 100. After 20 consecutive iterations, the optimal path combination of the group stably generates the scheduling search result matrix. In the final scheduling scheme, the average transportation time for 50 vehicles decreased from 78 minutes to 51 minutes, the number of path conflicts decreased from 16 to less than 1, the average waiting time decreased from 23 minutes to 7 minutes, the overall transportation efficiency improved by 34%, and the on-time arrival rate of steel components increased from 72% to 96%. This embodiment fully demonstrates that the present invention can effectively reduce transportation delays, optimize route combinations, and improve on-time performance in the transportation scheduling of complex multi-objective steel components, thus realizing the practical engineering feasibility of dynamic multi-objective scheduling.

[0031] Table 1. Performance Comparison of the Invention Method and Traditional Methods Regarding Steel Component Transportation Route Scheduling

[0032] As can be seen from the statistical results in Table 1, under the same number of transportation tasks and vehicles, the method of this invention significantly improves several key performance indicators compared to the traditional scheduling method. The average transportation time decreased from 78 minutes using the traditional method to 51 minutes, indicating that the vehicle travel routes have been optimized and the overall transportation efficiency has significantly improved. The number of route conflicts decreased from 16 to less than 1, showing that the method of this invention can effectively avoid multiple vehicles occupying the same route simultaneously, reducing transportation delays. The average vehicle waiting time decreased from 23 minutes to 7 minutes, reflecting more rational route scheduling, a significant reduction in vehicle dwell time during transportation, and a significant improvement in resource utilization.

[0033] Regarding scheduling accuracy, the average on-time arrival rate of the method in this invention increased from 72% to 96%, indicating that steel components can arrive at designated nodes more accurately according to the planned time, ensuring the continuity of construction nodes and the efficiency of on-site operations. Simultaneously, the multi-objective scheduling score improved from 62 points in the traditional method to 91 points, fully demonstrating the optimization capability of the method in this invention under multiple objective conditions, including comprehensive consideration of transportation time, path length, load constraints, and road condition adaptability. These data show that by constructing a temporal dynamic graph attention network and combining it with the social spider optimization algorithm, this invention can achieve dynamic modeling of transportation node states, multi-objective constraint optimization, and utilization of historical embedded information, thereby generating efficient, stable, and executable scheduling schemes in complex transportation environments.

[0034] Overall, this comparative analysis fully demonstrates the advantages of the method of the present invention in terms of transportation efficiency, scheduling accuracy, and multi-objective optimization capabilities, and verifies the feasibility of the simulated data in the embodiments and its practical engineering application value.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for scheduling the transportation routes of steel components based on big data analysis, characterized in that, include: Collect transportation task data for steel components, preprocess the transportation task data, and generate standardized transportation task sequences; Based on the standardized transportation task sequence, the paths and nodes are identified, and the node status, vehicle status and path attributes of each path are extracted as path node features to construct a time-series dynamic graph. Based on temporal dynamic graphs, an improved temporal dynamic graph attention network is constructed, including a node embedding adaptive grouping layer, a path-dependent self-looping layer, and a path history embedding enhancement layer. In the node embedding adaptive grouping layer, nodes are divided into subgroups based on feature similarity to generate a node embedding matrix; In the path-dependent self-loop layer, the node embedding matrix of each path is iteratively updated, information is passed between nodes within the path, and a path embedding matrix is ​​generated. In the path history embedding enhancement layer, the path embedding matrix of each path is fused with the corresponding path embedding matrix of the previous iteration, aligned according to the path node order, the node embedding state is updated, and a path history embedding matrix is ​​generated. The algorithm for optimizing the social spider population is based on the path history embedding matrix. Spider individuals are initialized, and female and male individuals are divided according to a dynamic role switching strategy. The individual positions are updated using an iterative step size adaptive update rule. Vibration information is propagated in the group to perform a global multi-objective path search, generate scheduling search results, and determine the steel component transportation scheduling plan, including the routes and time arrangements of each transport vehicle.

2. The method for scheduling steel component transportation routes based on big data analysis according to claim 1, characterized in that, The transportation task data specifically includes transportation node information, transportation vehicle information, steel component loading information, transportation route attribute information, transportation time identifier data, and historical transportation record identifier data.

3. The method for scheduling steel component transportation routes based on big data analysis according to claim 1, characterized in that, The preprocessing of transportation task data includes: Identify and remove abnormal information in transportation task data, including missing node status, abnormal vehicle status, or abnormal route attribute data. Missing node status, vehicle status, or path attribute data are filled in by interpolating preceding and following nodes to generate complete information; The node status and vehicle status are sorted according to the transportation node number, vehicle number and task order to form a standardized node list arranged in transportation order. The transportation route attributes are standardized, including route length, capacity, traffic restrictions, and estimated transportation time, and each route attribute is mapped to a unified dimension. The transportation time identification data is corrected to ensure that the time sequence from the origin to the destination is consistent with the node sequence, abnormal time records are corrected, and a standardized transportation time sequence is formed.

4. The method for scheduling steel component transportation routes based on big data analysis according to claim 1, characterized in that, The construction of the timing dynamic graph is specifically as follows: Based on the standardized transportation task sequence, extract the information of each path and node, number them according to the transportation sequence, and generate a path node list; Based on the list of path nodes, vehicle status characteristics are assigned to each path node, including vehicle type, load and current location status, to form path node status information; Extract the attribute information of each path, including path length, traffic capacity, traffic restrictions and estimated transportation time, and associate it with the corresponding path nodes to generate path attribute features; The status information of each path node is combined with the path attribute information to generate a path node feature set, while keeping the node order consistent with the path. A time-series dynamic graph is constructed based on the feature set of path nodes. Path nodes are represented as graph nodes, and the connection relationships between nodes and paths are represented as graph edges. Path attribute information is written into the corresponding graph edge attributes to generate an initial time-series dynamic graph. Assign a unique identifier to each node and edge in the initial time-series dynamic graph, and record the path node features and edge attributes on the corresponding nodes and edges to form a complete time-series dynamic graph.

5. The method for scheduling steel component transportation routes based on big data analysis according to claim 1, characterized in that, The construction of the improved temporal dynamic graph attention network specifically involves: The improved temporal dynamic graph attention network includes a node embedding adaptive grouping layer, a path-dependent self-looping layer, and a path history embedding enhancement layer. The parameters of each layer are initialized, including weights, biases, and node embedding vector dimension information. Based on the time-series dynamic graph, the path node feature information and graph edge attributes of each node in the time-series dynamic graph are used as node input features to record the path index, sequence position and adjacency relationship of each node. The improved temporal dynamic graph attention network was trained using a combination of node embedding difference error, path embedding consistency error, path history embedding fusion error, node order preservation error, and adjacency information preservation error as joint optimization objectives. The network parameters of the node embedding adaptive grouping layer, path-dependent self-looping layer, and path history embedding enhancement layer were continuously optimized. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved temporal dynamic graph attention network was considered to have completed convergence training.

6. The method for scheduling steel component transportation routes based on big data analysis according to claim 1, characterized in that, The generated node embedding matrix is ​​specifically as follows: In the node embedding adaptive grouping layer, feature similarity calculation is performed on the nodes of each path. Based on the feature similarity, the nodes are divided into several subgroups. Within each subgroup, information aggregation is performed according to the node order and feature similarity relationship to generate the subgroup node embedding representation. The initial node embedding matrix of the entire path is formed by combining all subgroup node embeddings in the order of path nodes. Batch normalization and nonlinear transformation are performed on the initial node embedding matrix of each path to generate the final node embedding matrix.

7. The method for scheduling steel component transportation routes based on big data analysis according to claim 1, characterized in that, The generated path embedding matrix is ​​specifically as follows: Input a node embedding matrix, arrange the node embeddings of each path according to the path number and node order, and record the path index, sequence position and adjacent node information of each node; In the path-dependent self-loop layer, information iterative update operation is performed on the node embedding matrix of each path. In each iteration, the current node embedding and the adjacent node embedding are sequentially transferred, merged and weighted along the path, and the node embedding change state is recorded to form a closed-loop representation of the node embedding within the path. The node embedding within the repeated path is iterated in a closed loop for several rounds. In each round of iteration, the node order is kept consistent with the path sequence, and the node embedding state is updated until convergence, generating the stable state of the node embedding for each transportation path and forming the initial path embedding representation. The initial path embedding representation of each path is organized and serialized. The node embeddings within the path are arranged in the transportation order, and the node index, adjacency relationship and path index are recorded to form a complete path embedding matrix.

8. The method for scheduling steel component transportation routes based on big data analysis according to claim 1, characterized in that, The generated path history embedding matrix is ​​specifically as follows: Input the path embedding matrix for each path, arrange the node embeddings according to the path node order, and record the path index, sequence position and node embedding state of the previous iteration for each node; In the path history embedding enhancement layer, the node embedding of each path is iteratively updated. The current round path embedding is aligned with the corresponding path embedding of the previous round according to the node order and then fused. A preliminary path history embedding representation is generated through information aggregation. Multiple iterations are performed on the initial path history embedding representation. In each iteration, the node order is kept consistent with the path sequence. After each iteration, the node embedding state is updated until the node embedding state is stable. The stable path history embedding representation is normalized to eliminate numerical differences, and a nonlinear transformation is applied to generate an updated path history embedding representation, while maintaining the integrity of node order and path sequence information. The updated path history embedding representation for each path is organized and serialized, and the index of each node and the relationship between adjacent nodes are recorded to generate a path history embedding matrix.

9. The method for scheduling steel component transportation routes based on big data analysis according to claim 1, characterized in that, The specific steps for determining the transportation and scheduling plan for steel components are as follows: The path history embedding matrix is ​​used as input. The node embedding of each path is arranged according to the path number and node order. The individual spider state in the social spider optimization algorithm is initialized, and the path selection, position parameters and path history embedding information of each individual are recorded. The female and male spiders in the group are divided into groups according to a dynamic role switching strategy, the group structure is established, and each individual is assigned an iteration step size, vibration sensing parameters and search state identifier to form the initial state set of the group. In each iteration, the position update operation is performed on each spider individual based on the path history embedding matrix information. The individual path position is adjusted by the node order and path embedding relationship. At the same time, the individual vibration intensity is calculated, and the vibration information is transmitted through the group propagation mechanism to generate the updated spider individual state. Based on the updated individual spider status, perform multi-objective evaluation, including transportation time, path length, load constraints and road condition adaptability. Update the individual's priority path selection and iteration step size based on the multi-objective evaluation results, and save the current group's optimal path combination. Repeatedly perform individual spider position updates and multi-objective evaluation operations until the optimal path combination in the group does not change in continuous iterations, forming the optimal node embedding order for each path and generating the group's optimal path state set; Organize the path information of each spider individual in the optimal path state set of the group and the corresponding multi-objective evaluation results, arrange the optimal node embedding order of each path according to the path number, and generate a scheduling search result matrix.

10. Read the scheduling search result matrix, and obtain the node information and path combination information for each path according to the path number and node order; Based on the node information and route combination information of each route, the transport vehicles are matched with the routes, a driving node sequence for each vehicle is generated, and the vehicle number, node order and route index are recorded to form a preliminary vehicle route plan. Schedule the travel node sequence of each transport vehicle, determine the estimated arrival and departure times of each node, record the time information in relation to the vehicle number and route index, and generate a vehicle route timetable. The preliminary vehicle route plan and vehicle route timetable were compiled, and the node sequence and time arrangement of all transport vehicles were merged to form a complete steel component transportation scheduling plan, including the driving route and corresponding time arrangement of each vehicle.

Images