Logistics transportation optimization method and system based on traffic logistics large model

Abstract

The application provides a logistics transportation optimization method and system based on a traffic logistics large model, which extracts elements by receiving basic scheduling instructions of a logistics transportation task, generates core element information through semantic scene analysis, adapts a transportation scene based on the core element information, and generates a set of scene influence factors in combination with deep feature mining results of a historical transportation scene database; then, a pre-trained traffic logistics large model is called to make a collaborative path decision on the set of scene influence factors, iteratively interacts a path generation unit and a cost evaluation unit in the model, and outputs a candidate path scheme population containing a path topology structure and a cost feature distribution; the candidate path scheme population is subjected to multi-round evolutionary screening according to a preset optimization target, an optimal path scheme meeting transportation requirements is determined, and a converted set of transportation scheduling instructions is pushed to a transportation execution terminal. The application can effectively improve adaptability, decision accuracy and reliability in a logistics transportation optimization process.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a logistics transportation optimization method and system based on a large-scale transportation and logistics model. Background Technology

[0002] With the rapid development of the logistics industry, logistics transportation optimization technology, as a core technology for realizing transportation route planning and resource allocation, plays a crucial role in improving transportation efficiency and cost control. Current logistics transportation optimization methods typically extract fixed element information from basic scheduling instructions, generate static scenario conditions based on historical scenario data matched with a preset rule base, calculate candidate routes using traditional route models, and then select the optimal solution through a single round of decision-making, finally converting it into a fixed-format instruction sent to the terminal. However, these methods rely on fixed template parsing in the element extraction stage, making it difficult to capture the implicit complex scenario intent; in scenario adaptation, historical data and the current scenario are mostly statically superimposed, failing to reflect the spatiotemporal changes in the transportation scenario; route planning and cost assessment lack interactive feedback, making it difficult to balance the adaptation relationship between route topology and cost characteristics; the optimization target weight is fixed during route selection, resulting in insufficient adaptability in complex scenarios; and the instruction generation stage lacks real-time interactive verification at the terminal, easily leading to format mismatch problems and affecting the smooth progress of transportation tasks. Summary of the Invention

[0003] In view of this, the present invention provides a logistics transportation optimization method and system based on a large-scale transportation and logistics model.

[0004] The technical solution of this invention is implemented as follows:

[0005] On one hand, embodiments of the present invention provide a logistics transportation optimization method based on a large-scale transportation and logistics model. The method includes: receiving basic scheduling instructions for logistics transportation tasks; extracting elements from the basic scheduling instructions; generating core element information containing spatial associations of the origin and destination of the path and cargo transportation needs through semantic scene parsing; adapting the transportation scenario based on the core element information; generating a set of scenario influence factors required for transportation route planning by combining the deep feature mining results of a historical transportation scenario database; calling a pre-trained large-scale transportation and logistics model to perform collaborative path decision-making on the set of scenario influence factors; outputting a candidate path scheme population containing path topology and cost feature distribution through interactive iteration between the path generation unit and cost evaluation unit within the model; performing multiple rounds of evolutionary screening on the candidate path scheme population according to a preset optimization objective; conducting a comprehensive performance evaluation of each scheme to determine the optimal path scheme that meets transportation needs; converting the optimal path scheme into instructions to generate a transportation scheduling instruction set and pushing it to the transportation execution terminal through a real-time data interaction interface.

[0006] On the other hand, embodiments of the present invention provide a computer system including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the program to implement the steps in the above-described method.

[0007] The logistics transportation optimization method based on a large-scale transportation and logistics model provided by this invention extracts elements from the basic scheduling instructions of logistics transportation tasks, combines semantic scene parsing to cover contextual semantic associations and multi-turn interaction intent recognition, and generates core element information including spatial associations of path origin and destination and cargo transportation needs. This enables more accurate capture of implicit transportation needs and complex scenario intents in scheduling instructions. Based on the core element information, transportation scenarios are adapted, and combined with the deep feature mining results of historical transportation scenario databases, a set of scenario influencing factors is formed through spatial feature evolution and time series coupling. This achieves the evolution and real-time coupling of transportation scenario features, effectively adapting to the characteristics of transportation scenarios changing with time and space. Finally, a pre-trained large-scale transportation and logistics model is invoked to perform collaborative path decision-making on the set of scenario influencing factors. This method, through the interactive iteration of the path generation unit and cost evaluation unit within the model, outputs a population of candidate path solutions containing path topology and cost feature distribution, promoting the collaborative optimization of path generation and cost evaluation, and improving the matching accuracy between path solutions and cost features. Based on a preset optimization objective, the population of candidate path solutions undergoes multiple rounds of evolutionary screening, performing comprehensive performance evaluation and combining real-time feedback and iterative optimization, gradually approaching the optimal path solution that meets transportation needs, improving the optimization accuracy and scenario adaptability of the path solution. The optimal path solution is then converted into instructions to generate a transportation scheduling instruction set, which is pushed to the transportation execution terminal through a real-time data interaction interface. Combined with instruction format adaptation and terminal interaction response verification, the compatibility and reliability of the scheduling instructions with the transportation execution terminal are ensured. Through the synergistic effect of the above technical features, this method can effectively improve the adaptability, decision-making accuracy, and reliability in the logistics transportation optimization process, meeting the needs of complex and ever-changing logistics transportation scenarios. Attached Figure Description

[0008] Figure 1 This is a schematic diagram illustrating the implementation process of a logistics transportation optimization method based on a large-scale transportation and logistics model, provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the hardware entity of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0010] This invention provides a logistics transportation optimization method based on a large-scale transportation and logistics model, which can be executed by a computer system processor. The computer system can refer to devices with data processing capabilities, such as servers, laptops, tablets, and desktop computers.

[0011] Figure 1 This invention provides a schematic diagram illustrating the implementation process of a logistics transportation optimization method based on a large-scale transportation and logistics model, as shown in the embodiment of the invention. Figure 1 As shown, the method includes:

[0012] Step S100: Receive the basic scheduling instructions for logistics transportation tasks, extract elements from the basic scheduling instructions, and generate core element information including the spatial association of the origin and destination of the path and the cargo transportation requirements through semantic scene parsing. The element extraction process covers the contextual semantic association of the instruction text and the recognition of multi-round interactive intents.

[0013] The basic scheduling instruction for a logistics transportation task refers to the initial instruction issued when initiating a logistics transportation task. This instruction contains basic information about the task, such as the origin, destination, and relevant requirements of the goods. Element extraction is the process of identifying and extracting key information from the basic scheduling instruction, aiming to extract useful information from the instruction. Semantic scenario analysis involves understanding and analyzing the semantics of the instruction text, combining it with scenario information to clarify the actual meaning of the instruction. Path origin-end spatial association represents the spatial relationship between the origin and end points of the transportation path, such as the distance between the two points and their geographical location. Cargo transportation requirements refer to various requirements of the goods during transportation, such as transportation time, the nature of the goods (e.g., fragile, perishable), and transportation mode requirements. Core element information is a set of key information obtained after processing, which is crucial for subsequent transportation path planning and task execution. Contextual semantic association refers to the semantic connection between the preceding and following texts of the instruction text; considering the context allows for a more accurate understanding of the instruction's meaning. Multi-turn interaction intent recognition identifies the user's true intent during multi-turn interactions to ensure that the extracted element information accurately reflects the user's needs.

[0014] Specifically, receiving basic dispatch instructions can be achieved through an information receiving interface, which can receive instruction information from various channels, such as logistics management systems and mobile applications. For element extraction, natural language processing techniques can be employed, such as named entity recognition algorithms, to identify key entities in the instructions, such as location, cargo type, and time. In terms of semantic scene parsing, pre-trained language models, such as the BERT model, can be used to semantically understand the instruction text, combining it with scene information to generate core element information such as the spatial association of the origin and destination of the route and cargo transportation needs. During the element extraction process, to cover the contextual semantic association of the instruction text and the recognition of multi-turn interaction intentions, a dialogue management system can be built to record multi-turn interaction information with the user, and context-aware algorithms can be used to comprehensively analyze the instructions.

[0015] Step S200: Based on the core element information, adapt the transportation scenario and generate the set of scenario influencing factors required for transportation route planning by combining the deep feature mining results of the historical transportation scenario database. The set of scenario influencing factors is formed by coupling spatial feature evolution with time series.

[0016] Transportation scenario adaptation is the process of matching the current transportation task with a suitable transportation scenario based on core element information. A historical transportation scenario database stores information related to past transportation scenarios, including the characteristics and results of various scenarios. Deep feature mining is the process of extracting deep, valuable feature information from the historical transportation scenario database. By analyzing these features, patterns and regularities under different transportation scenarios can be understood. The scenario influencing factor set is a collection of factors that affect transportation route planning. These factors comprehensively consider spatial and temporal factors, enabling a more accurate reflection of the actual situation of the transportation scenario. Spatial feature evolution refers to the process of spatial features changing over time or due to other factors, such as changes in road congestion over time. Time series coupling combines spatial features with time series information to reflect changes in the transportation scenario at different points in time.

[0017] As one implementation method, step S200 can be implemented as the following steps S210~S250:

[0018] Step S210: Perform grid-based modeling on the spatial relationship between the origin and destination of the path in the core element information, and generate a spatial grid index that matches the coordinate system of the historical transportation scenario database. The grid-based modeling process includes coordinate system unification and spatial scale adjustment.

[0019] Grid-based modeling of the spatial association between the start and end points of a path involves dividing the spatial region containing the path's starting and ending points into several grids. This method allows for a more precise representation of spatial location information. The spatial grid index is an index structure used for rapid location and retrieval of spatial data, associating spatial data with grids to improve data query efficiency. The coordinate system of the historical transportation scenario database is the coordinate system used by the database to represent spatial location information within the transportation scenario. Coordinate system unification involves converting coordinate data from different sources into a unified coordinate system to ensure data consistency and comparability. Spatial scale adjustment involves adjusting the spatial scale of the grid according to actual needs, such as adjusting the grid size, to adapt to different application scenarios.

[0020] Specifically, Geographic Information System (GIS) technology can be used to perform gridded modeling of the spatial relationship between the origin and destination of a route. First, the coordinate data of the origin and destination of the route are obtained from the core element information. Then, based on the coordinate system of the historical transportation scenario database, these coordinate data are standardized to a unified coordinate system. Coordinate transformation algorithms, such as the seven-parameter transformation method, can be used to convert coordinate data from different coordinate systems into a unified coordinate system. When adjusting the spatial scale, the characteristics and requirements of the transportation scenario need to be considered. If the transportation area is large, the grid size can be appropriately increased; if more precise positioning is required, the grid size can be decreased. Spatial grid indexes can be generated using spatial index structures such as quadtree indexes and R-tree indexes.

[0021] Step S220: Extract demand features from the core element information to generate a demand vector that includes the urgency of transportation demand and the adaptability of goods. The dimension of the demand vector is the same as the feature dimension of the scenario influencing factors.

[0022] Demand feature extraction is the process of extracting feature information related to transportation demand from core element information. Transportation demand urgency refers to the urgency of the cargo transportation task, such as whether there are strict delivery time requirements. Cargo adaptability features refer to the adaptability of the cargo's nature and characteristics to transportation methods and conditions, such as the cargo's weight, volume, and fragility. A demand vector is a vector that quantifies the extracted demand features; each dimension of this vector corresponds to a demand feature. The feature dimensions of scenario influencing factors refer to the number and types of features contained in the scenario influencing factors.

[0023] In practical applications, feature engineering techniques can be used to extract demand features from core element information. For the urgency of transportation needs, it can be quantified based on the delivery time requirements of the goods; for example, tasks with delivery times closer to the present time have higher urgency settings. For the adaptability features of goods, they can be classified and quantified based on the nature of the goods and transportation requirements; for example, the adaptability feature of fragile goods can be set to a value indicating higher requirements for transportation methods and conditions. When generating the demand vector, the extracted transportation demand urgency and goods adaptability features are arranged in a specific order to form a vector. To ensure that the dimension of the demand vector is the same as the feature dimension of the scene influencing factors, features need to be screened and adjusted during the feature extraction process to ensure consistency in feature types and quantities.

[0024] Step S230: Call the feature mining unit of the historical transportation scenario database, perform similar scenario retrieval based on the spatial grid index, and extract the implicit association patterns between scenario features and transportation results through the association rule mining algorithm. The association patterns are represented by the feature interaction weight matrix.

[0025] The feature mining unit of the historical transportation scenario database is a module specifically designed to extract valuable information from this database. The spatial grid index is used to quickly locate and retrieve historical transportation scenarios related to the current transportation task. Similar scenario retrieval searches the historical transportation scenario database for scenarios similar to the current task in terms of spatial location, cargo demand, etc. Association rule mining algorithms are data mining algorithms used to discover the relationships between different variables in a dataset. Implicit association patterns between scenario features and transportation outcomes refer to relationships hidden in the data and not easily discovered directly, such as the association between certain road conditions, weather conditions, and transportation time, transportation costs, etc. The feature interaction weight matrix is ​​a matrix used to represent the interactions and influences between scenario features; the elements in the matrix represent the weight relationships between different features.

[0026] As one implementation method, step S230 can be implemented as the following steps S231~S235:

[0027] Step S231: Convert the spatial grid index into an entity identifier sequence that can be recognized by the feature mining unit, construct a query statement containing the origin and destination entities of the path and the cargo type entities, and adjust the entity relationship of the query statement through temporal association weights.

[0028] Spatial grid indexes are index structures used for spatial data retrieval. Feature mining units typically require formatted data for processing. Converting the spatial grid index into a sequence of entity identifiers recognizable by the feature mining unit involves transforming spatial grid information into entity identifiers that the feature mining unit can understand and process. Path origin and destination entities represent the starting and ending points of a transportation path at the entity level, such as specific warehouses or distribution centers. Cargo type entities refer to the specific types of goods, such as electronic products or food. A query statement is a statement used to retrieve relevant data from a historical transportation scenario database. By constructing a query statement that includes path origin and destination entities and cargo type entities, historical scenarios related to the current transportation task can be located more accurately. Temporal association weights are used to adjust the weights of entity relationships in the query statement. Considering the temporal sequence during transportation and the temporal association between entities, adjusting the weights can more accurately reflect the relationships between entities.

[0029] Specifically, converting a spatial grid index into a sequence of entity identifiers can be achieved using a mapping table that records the correspondence between spatial grids and entity identifiers. When constructing queries, Structured Query Language (SQL) or other query languages ​​are used to combine path origin / destination entities and cargo type entities into query conditions. When adjusting the temporal correlation weights of entity relationships, the degree of temporal correlation between different entities can be determined based on the time requirements of the transportation task and time information in historical data.

[0030] Step S232: The feature mining unit performs a depth-first scene traversal based on the query statement, retrieves historical transportation scene nodes associated with the origin and destination entities, records the scene feature attributes and transportation result feedback data of each historical transportation scene node, and adapts the traversal depth according to the scene complexity.

[0031] Depth-first scene traversal is a graph traversal algorithm. The feature mining unit starts from the initial node and visits nodes as deep as possible along a path until it cannot continue. Then, it backtracks to the previous node and continues exploring other paths. In this way, it retrieves historical transportation scene nodes associated with the origin and destination entities. Historical transportation scene nodes are data nodes in a historical transportation scene database. Each node represents a specific transportation scene and contains various feature attributes and transportation result feedback data for that scene. Scene feature attributes describe various characteristics of the transportation scene, such as road conditions, weather conditions, and transportation vehicles. Transportation result feedback data refers to the result information obtained after the transportation task is completed, such as transportation time, transportation cost, and cargo damage status. The traversal depth is adapted to the scene complexity, meaning that the traversal depth is determined based on the complexity of the transportation scene to ensure that sufficient relevant information is retrieved while avoiding unnecessary computation.

[0032] Specifically, after receiving the query, the feature mining unit performs a depth-first traversal of the scene, starting from the initial node related to the origin and destination entities. A stack data structure can be used to implement this depth-first traversal, pushing nodes to be visited onto the stack and popping one node from the stack for each visit. During the traversal, the scene feature attributes and transportation result feedback data of each historical transportation scene node are recorded. This data can be stored in temporary data storage using a data recording module. The traversal depth can be adjusted based on factors such as the number of nodes in the transportation scene and the connections between them. For example, if the transportation scene is simple and the number of nodes is small, the traversal depth can be set smaller; if the transportation scene is complex and the relationships between nodes are intricate, the traversal depth needs to be increased appropriately.

[0033] Step S233: Perform feature standardization processing on the retrieved historical scene nodes, converting scene feature attributes of different dimensions into standardized feature values ​​of a unified range.

[0034] Feature standardization of historical scene nodes aims to eliminate the influence of different units of measurement among scene feature attributes, making different features comparable. Scene feature attributes with different units of measurement refer to features with different units of measurement; for example, road length might be measured in kilometers, while transportation costs might be measured in yuan. Converting these feature attributes with different units of measurement into standardized feature values ​​within a unified range allows feature mining algorithms to process data more effectively, improving algorithm performance and accuracy. Specifically, feature standardization methods include min-max standardization and Z-score standardization, without further limitation.

[0035] Step S234: Perform association analysis on standardized feature values ​​and transportation result feedback data using association rule mining algorithm, filter key association rules by support-confidence threshold, and generate a rule set containing antecedent feature items, consequent result items and association strength.

[0036] Association rule mining algorithms are used to discover the relationships between different variables in a dataset. Standardized feature values ​​are the scene feature attribute values ​​after standardization, and transportation result feedback data is the result information after the transportation task is completed. Association analysis identifies association rules by analyzing the relationship between standardized feature values ​​and transportation result feedback data. Support refers to the frequency of association rules appearing in the dataset, reflecting the generality of the rule. Confidence refers to the probability that the consequent result term is satisfied given that the antecedent feature term is satisfied, reflecting the reliability of the rule. By setting a support-confidence threshold, key association rules with a certain degree of generality and reliability can be screened out. The antecedent feature term is the condition part of the association rule, and the consequent result term is the conclusion part. Association strength is an indicator that measures the strength of an association rule, usually calculated by combining factors such as support and confidence. Specifically, feasible association rule mining algorithms include the Apriori algorithm and the FP-growth algorithm.

[0037] As one implementation method, step S234 can be implemented as the following steps S2341~S2345:

[0038] Step S2341: Discretize the standardized feature values ​​to generate a feature term set. Convert continuous feature values ​​into discrete feature labels by dividing the feature intervals. The interval division is adjusted according to the feature distribution density.

[0039] Standardized feature values ​​are the scene feature attribute values ​​after standardization, usually continuous numerical values. Discretization is the process of converting continuous feature values ​​into discrete feature labels, facilitating association rule mining. The feature term set is the set containing discrete feature labels obtained after discretization. Feature interval partitioning divides the range of continuous feature values ​​into several intervals, each interval corresponding to a discrete feature label. Feature distribution density refers to the distribution of feature values ​​across different value ranges; adjusting the interval partitioning based on the feature distribution density can make the discretized features better reflect the actual situation of the data.

[0040] Specifically, feasible discretization methods include equal-width discretization, equal-frequency discretization, and cluster-based discretization. Equal-width discretization divides the range of feature values ​​into intervals of equal width; equal-frequency discretization divides feature values ​​into several intervals according to their frequency of occurrence, with approximately the same number of samples in each interval; cluster-based discretization uses a clustering algorithm to cluster the feature values, with each cluster corresponding to a discrete feature label.

[0041] Step S2342: Classify the transportation result feedback data into result categories and generate result level labels. The result level labels are dynamically related to the degree of satisfaction of transportation needs.

[0042] Transportation outcome feedback data refers to the information obtained after a transportation task is completed, such as transportation time, transportation cost, and cargo damage status. Outcome categorization is the process of classifying transportation outcome feedback data according to certain standards. Outcome level labels are tags used to mark the categorized outcome types, representing different outcome levels. Transportation demand satisfaction refers to the degree to which the transportation outcome meets the needs of cargo transportation. Outcome level labels are dynamically related to transportation demand satisfaction, meaning that outcome level labels will change according to different transportation demands.

[0043] Step S2343: Initialize the set of individual scene feature units as an empty set, scan the historical scene database through a sliding window to calculate the occurrence frequency of individual scene feature units, and filter significant individual scene feature units through a frequency threshold. The frequency threshold is adaptively adjusted according to the number of scene features.

[0044] A single scene feature unit refers to a single scene feature in the historical scene database, such as road congestion or weather conditions. A set of single scene feature units is the collection containing these single scene feature units. Initializing the single scene feature unit set to an empty set involves clearing the set before computation to allow for recalculation and filtering. Sliding window scanning is a data processing method that processes data within a fixed-size window that slides across the historical scene database. The frequency of occurrence of a single scene feature unit refers to the ratio of the number of times that feature unit appears in the historical scene database to the total number of samples. A frequency threshold is a standard used to filter significant single scene feature units; only feature units with a frequency higher than this threshold are considered significant. The number of scene features refers to the number of different scene features contained in the historical scene database. Adaptively adjusting the frequency threshold based on the number of scene features ensures that the selected significant feature units are representative.

[0045] Specifically, when using a sliding window to scan a historical scene database, the window size can be determined based on the characteristics of the data and the needs of the analysis. For the data within each window, the frequency of occurrence of each individual scene feature unit is calculated. The frequency threshold can be adaptively adjusted based on the number of scene features. If the number of scene features is large, the frequency threshold can be appropriately increased to filter out more representative feature units; if the number of scene features is small, the frequency threshold can be appropriately decreased.

[0046] Step S2344: Generate combined scene feature units containing multiple features from significant single scene feature units through a layer-by-layer iterative approach. Remove combined units containing non-significant features through pruning operations. Calculate the occurrence frequency of combined scene feature units and filter significant combined scene feature units through a threshold.

[0047] A salient single scene feature unit is a representative single scene feature unit obtained after frequency thresholding. A combined scene feature unit is a feature unit composed of multiple salient single scene feature units. The iterative approach involves starting with a single scene feature unit and gradually generating combined scene feature units containing more features. Pruning removes combined scene feature units containing non-salient features during the generation process to reduce unnecessary computation and analysis. The frequency of occurrence of a combined scene feature unit refers to the ratio of the number of times the combined feature unit appears in the historical scene database to the total number of samples. Thresholding salient combined scene feature units ensures that the selected combined feature units have a certain degree of universality and representativeness.

[0048] Specifically, the iterative generation of combined scene feature units can utilize the layer-by-layer search approach of the Apriori algorithm. Starting with a salient single scene feature unit, combinations containing more features are generated at each stage. For example, combinations containing two salient single scene feature units are generated first, then combinations containing three features, and so on. Pruning is performed during the combination generation process. If a combined unit contains non-salient features, it is removed. For each generated combined scene feature unit, its frequency of occurrence is calculated, and salient combined scene feature units are selected using the same or different frequency thresholds as used for selecting individual scene feature units.

[0049] Step S2345: For each significant combination scene feature unit, generate all possible non-empty true subsets as rule antecedents and the remaining parts as rule consequents. Filter key association rules through confidence thresholds to generate a rule set containing antecedent feature items, consequent result items, and association strength. The association strength is calculated by weighted fusion of occurrence frequency and confidence.

[0050] A salient combination of scene feature units is a representative combination of feature units containing multiple features, obtained after screening. A non-empty proper subset is a subset of a set that is not empty and is not equal to the original set. The antecedent of a rule is the condition part of the association rule, and the consequent is the conclusion part. The confidence threshold is a standard used to screen key association rules; only association rules with a confidence level higher than this threshold are considered reliable. Association strength is an indicator that measures the strength of an association rule. It is calculated by weighting and fusing frequency and confidence levels to more comprehensively reflect the reliability and universality of the association rule.

[0051] Step S235: Perform redundant rule pruning on the rule set, eliminate semantically repetitive association rules by calculating rule coverage, and retain implicit association patterns with unique predictive value. The pruning process includes ranking rules by importance and filtering rules with low contribution.

[0052] A rule set is a collection of multiple association rules obtained after association rule mining and filtering. Redundant rule pruning is the process of removing semantically repetitive or low-contribution rules from the rule set to improve its quality and efficiency. Rule coverage refers to the ratio of the number of samples covered by a rule in the dataset to the total number of samples; calculating rule coverage can determine the degree of repetition between rules. Semantically repetitive association rules are rules with the same or similar semantics, which may lead to information redundancy. Implicit association patterns are relationships hidden in the data that are not easily discovered directly; retaining implicit association patterns with unique predictive value can provide more valuable information for transportation route planning. Rule importance ranking sorts the rules in the rule set according to their importance to determine which rules should be retained and which should be removed. Low-contribution rule filtering removes rules that contribute little to the prediction results.

[0053] Specifically, rule coverage can be calculated by counting the number of samples covered by each rule in the historical scenario database. For semantically repetitive rules, this can be determined by comparing the antecedent features and consequents of the rules. For example, if the antecedents and consequents of two rules are essentially the same, they are considered semantically repetitive. Rule importance ranking can be comprehensively evaluated based on metrics such as association strength, support, and confidence. Low-contribution rule filtering can be achieved by setting a contribution threshold; rules with a contribution below this threshold are removed from the rule set.

[0054] Step S240: Perform scenario state evolution analysis on the temporal convolutional network in the implicit association pattern coupled with the demand vector input scenario situation, and generate an influence weight distribution containing time decay characteristics. The time series length of the weight distribution matches the transportation task cycle.

[0055] Implicit association patterns are hidden relationships within data, discovered through association rule mining. These relationships are crucial for understanding the characteristics of transportation scenarios. Demand vectors, containing the urgency of transportation needs and the adaptability characteristics of goods, reflect the specific requirements of cargo transportation. Scenario-situation coupling models are used to comprehensively analyze various factors within a transportation scenario, coupling different information. Temporal convolutional networks (TCNs) are convolutional neural networks that process sequential data, capturing temporal dependencies within the data. Scenario state evolution analysis analyzes the state changes of a transportation scenario over time, understanding its development trend. Time decay characteristics refer to the gradual weakening of the influence of certain factors over time, reflected in the influence weight distribution as the weights decay over time. Influence weight distribution is a distribution that quantifies the degree of influence of different factors during transportation. The time series length of the weight distribution matches the transportation task cycle, ensuring that the weight distribution accurately reflects the situation of the transportation task throughout the entire cycle.

[0056] Specifically, implicit correlation patterns and demand vectors are input into the temporal convolutional network (CCNN) of the scenario situation coupling model. The input layer of the CCNN receives this data and performs feature extraction and processing through convolutional and pooling layers. During convolution, convolutional kernels capture the temporal dependencies in the sequence data, learning the associations between factors at different time points. Through scenario state evolution analysis, the model can predict the state changes of the transportation scenario at different time points. When generating the influence weight distribution, time decay characteristics are considered, such as using an exponential decay function, where the weights of corresponding factors gradually decrease over time. The time series length of the weight distribution is determined based on the transportation task cycle; for example, if the transportation task cycle is three days, the time series length of the weight distribution is also three days, with each time point corresponding to a weight value.

[0057] Step S250: The implicit association pattern, influence weight distribution and spatial grid features are coupled through a feature interaction algorithm to generate a set of scene influence factors with spatiotemporal characteristics. The feature dimensions of each scene influence factor in the set of scene influence factors are consistent.

[0058] Implicit association patterns are relationships hidden within data, reflecting the inherent connections between different factors in a transportation scenario. Influence weight distribution is the distribution of the degree of influence of different factors during transportation, considering time decay characteristics. Spatial grid features are spatial information obtained through gridded modeling of the spatial relationships between the origin and destination of a route, such as spatial grid indexes. Feature interaction algorithms are algorithms used to interact and fuse different types of features, enabling the discovery of potential relationships between features. Spatiotemporal characteristics refer to characteristics that simultaneously consider time and space factors; a set of scenario influence factors with spatiotemporal characteristics can more comprehensively reflect the actual situation of a transportation scenario. The scenario influence factor set is a collection containing multiple scenario influence factors, each of which is a factor that influences transportation route planning after comprehensively considering multiple factors. The consistent feature dimensions of all scenario influence factors ensure that each influence factor can be effectively compared and processed in subsequent calculations and analyses.

[0059] Specifically, the feature interaction algorithm can adopt the attention mechanism in deep learning, calculate the interaction weights between different features through the attention mechanism, and fuse implicit association patterns, influence weight distribution and spatial grid features.

[0060] Step S300: Call the pre-trained large-scale transportation and logistics model to make collaborative path decisions on the set of scene influencing factors. Through the interactive iteration of the path generation unit and the cost evaluation unit inside the model, output a population of candidate path solutions containing path topology and cost feature distribution.

[0061] The pre-trained large-scale transportation and logistics model is a model trained on a large amount of data for use in the transportation and logistics field. This model has powerful learning and predictive capabilities. The scenario influencing factor set is a collection of factors that influence transportation route planning after comprehensively considering multiple factors. Collaborative route decision-making is the decision-making process for selecting the optimal transportation route under the combined influence of multiple factors. The route generation unit is the module in the large-scale transportation and logistics model used to generate transportation routes, generating possible routes based on input information. The cost evaluation unit is the module used to evaluate the cost of transportation routes, considering multiple cost factors such as time cost and resource cost. Interactive iteration is the process of continuous information exchange and adjustment between the route generation unit and the cost evaluation unit, optimizing the route plan through multiple iterations. The route topology refers to the structural representation of the nodes and connections of the transportation route, reflecting the basic form of the route. The cost feature distribution refers to the distribution of transportation routes across different cost factors, such as the distribution of time cost and resource cost. The candidate route plan population is a collection containing multiple candidate route plans, generated through interactive iteration, providing multiple choices for subsequent screening.

[0062] As one implementation method, step S300 can be implemented as the following steps S310~S350:

[0063] Step S310: Input the set of scene influencing factors into the feature encoding layer of the transportation and logistics big model, calculate the correlation weight between scene features through a multi-head self-attention interaction mechanism, and generate scene feature context vectors. The dimension of the context vectors is adapted to the dimension of the model's hidden layer.

[0064] The scene influencing factor set is a collection of factors that influence transportation route planning, obtained by comprehensively considering multiple factors. The feature encoding layer of the large-scale transportation and logistics model is the layer used to encode input features, converting them into a format that the model can process. The multi-head self-attention interaction mechanism is an extension of the attention mechanism; through parallel computation by multiple attention heads, it can capture various relationships between features. The correlation weights between scene features refer to the degree of mutual influence between different scene features; calculating these correlation weights reveals the importance and relevance between features. The scene feature context vector is a vector obtained by comprehensively considering the relationships between scene features, containing contextual information about the scene features and providing a more comprehensive representation. The model hidden layer dimension is the number of neurons in the model's hidden layer; the context vector dimension matches the model hidden layer dimension to ensure that the context vector can be successfully input into the model's hidden layer for subsequent processing.

[0065] Specifically, the set of scene influencing factors is input into the feature encoding layer of the transportation and logistics big data model. The feature encoding layer encodes the input features, converting them into vector representations. Then, a multi-head self-attention interaction mechanism is used to process the encoded features. Each attention head calculates the attention score between features, and converts the attention score into attention weights using a softmax function. The attention weights calculated by different attention heads are concatenated to obtain the final association weights. Based on the association weights, the scene features are weighted and summed to generate a scene feature context vector. During the generation of the context vector, the vector dimension is adjusted to match the dimension of the model's hidden layers.

[0066] Step S320: Initial path sampling is performed through the path generation unit of the large transportation and logistics model. An initial path population containing different topological structures is generated based on the scene feature context vector. Invalid paths are removed through path validity verification. The population size is adjusted according to the scene complexity.

[0067] The path generation unit in the large-scale transportation and logistics model is a module specifically designed to generate transportation routes. Initial path sampling is the process of randomly selecting some paths from the possible path space during the initial stage of path generation. The scene feature context vector is a vector obtained by comprehensively considering the correlations between scene features; it contains contextual information about the scene and provides an important basis for path generation. The initial path population is a set containing multiple initial paths with different topological structures, representing different transportation route choices. Path validity verification checks whether the path meets the basic requirements of the transportation task, such as whether the path contains loops or complies with resource constraints, and removes invalid paths. The population size refers to the number of paths included in the initial path population. Adjusting the population size according to the scene complexity is to cover more possible paths while ensuring search efficiency.

[0068] As one implementation method, step S320 can be implemented as the following steps S321~S325:

[0069] Step S321: Analyze the spatial association features in the scene feature context vector, determine the location codes of the path start and end points in the road network topology, and adjust the location codes through the road network node association weights.

[0070] Scene feature context vectors are vectors obtained by comprehensively considering the relationships between scene features. They include spatial relationship features, reflecting the spatial relationship between the path's origin and destination and its association with other nodes. Road network topology refers to the structural representation of the nodes and connections in a road network, describing the connectivity between roads. Location encoding encodes the position of the path's origin and destination within the road network topology to facilitate location and search within the network. Road network node association weights refer to the degree of association between nodes in the road network. By adjusting the location encoding and considering the association weights between nodes, the actual positional relationship of the path's origin and destination within the road network can be more accurately reflected.

[0071] Specifically, spatial association features in the scene feature context vector can be analyzed using feature extraction algorithms to extract spatial location-related feature information. Based on this feature information, the positions of the path's origin and destination within the road network topology are determined. Location encoding can be represented using node numbers, coordinates, etc. Road network node association weights can be determined by calculating factors such as distance and connectivity between nodes. For example, Dijkstra's algorithm can be used to calculate the shortest path distance between nodes, and the association weight between nodes can be determined based on the distance. When adjusting location encoding, the road network node association weights are used as adjustment factors to weight the location encoding.

[0072] Step S322: Initialize the evolutionary algorithm parameters, set the population size, crossover probability, mutation probability, and maximum number of generations. The population size is adapted to the scene complexity level, and the complexity level is determined by the number of dimensions of the spatially related features.

[0073] Evolutionary algorithms are optimization algorithms that simulate the biological evolution process, used to find optimal solutions in a search space. Evolutionary algorithm parameters affect their performance and search results, including population size, crossover probability, mutation probability, and maximum number of generations. Population size refers to the number of individuals in the population, which represent the path options. Crossover probability is the probability of a crossover operation occurring during evolution; crossover involves exchanging some genes between two individuals to generate a new individual. Mutation probability is the probability of a mutation operation occurring during evolution; mutation randomly alters the genes of an individual. Maximum number of generations is the maximum number of iterations the evolutionary algorithm can perform; after reaching this maximum, the algorithm stops iterating. Scene complexity level is an assessment of the complexity of a transportation scene, determined by the number of dimensions of spatial association features. A higher number of dimensions in the spatial association features indicates a more complex scene, requiring adjustments to parameters such as population size accordingly.

[0074] Specifically, the scene complexity level is first determined based on the number of dimensions of the spatial association features. A higher number of dimensions indicates higher scene complexity, while a lower number of dimensions indicates lower scene complexity. The population size is then adjusted according to the scene complexity level: a higher complexity level results in a larger population size to cover more of the search space, while a lower complexity level results in a smaller population size to improve search efficiency. Crossover probability, mutation probability, and maximum number of generations can be set based on experience or experimentation.

[0075] Step S323: Create an initial population through the path generation strategy, and generate valid paths connecting the origin and destination through road network node selection, ensuring that each path has no loops and the number of nodes is within the threshold range. The threshold is adaptively adjusted according to the road network density.

[0076] Path generation strategy refers to the methods and rules used to generate transportation paths. The initial population is the population at the start of the evolutionary algorithm, containing multiple initial path schemes. Network node selection is the process of choosing suitable nodes from the network to construct paths. An effective path is a path that meets the basic requirements of the transportation task, such as being loop-free and meeting resource constraints. The node number threshold controls the path length, preventing paths from being too long or too short; this threshold is adaptively adjusted based on network density. Network density refers to the density of nodes and edges in the network; when network density is high, the node number threshold can be appropriately reduced; when network density is low, the node number threshold can be appropriately increased.

[0077] Specifically, feasible path generation strategies include random path generation strategies and greedy algorithm strategies. Random path generation strategies randomly select road network nodes to generate possible paths; greedy algorithm strategies select the optimal node at each step, progressively constructing the path. When generating paths, ensuring the path is free of loops can be achieved using loop detection algorithms, such as depth-first search, to check for loops within the path. The number of nodes is adaptively adjusted based on the road network density.

[0078] Step S324: Perform legality verification on each path in the initial population, evaluate the constraint adaptability based on the resource constraint characteristics in the scene influence factor set, eliminate invalid paths that violate hard constraints, and generate new paths through a supplementary mechanism until the preset population size is reached.

[0079] The initial population is the population at the start of the evolutionary algorithm, containing multiple initial path schemes. Validity verification checks whether a path meets the basic requirements of the transportation task. Resource constraint features in the scene influencing factor set refer to resource limitations encountered during transportation, such as vehicle load capacity and transportation time constraints. Constraint fit evaluation assesses whether a path conforms to resource constraints. Hard constraints are mandatory; paths that violate hard constraints are considered invalid. The replenishment mechanism generates new paths to replenish the population after removing invalid paths, ensuring the population size reaches a preset value. The preset population size is a pre-set population size based on factors such as scene complexity.

[0080] Specifically, for each path in the initial population, a constraint suitability assessment is performed based on the resource constraint characteristics in the scene influence factor set. Paths that violate hard constraints are removed from the population. Then, a replenishment mechanism is used to generate new paths. The replenishment mechanism can employ the same method as the path generation strategy, such as a random path generation strategy or a greedy algorithm strategy, to generate new paths. New paths are continuously generated until the population size reaches a preset value.

[0081] Step S325: Convert the verified initial path into a standardized topological structure representation, record the connection relationship between path nodes through an adjacency matrix, and arrange the node order according to the driving direction to form the chromosome encoding set of the initial path population.

[0082] The validated initial path is the path that meets the basic requirements of the transportation task after legality verification. Standardized topology representation transforms the path's topology into a unified and easily processed representation. An adjacency matrix is ​​a matrix used to represent the connection relationships between nodes in a graph; the elements in the matrix indicate whether a connection exists between nodes. The node order is arranged according to the direction of travel to accurately reflect the travel order of the path. The chromosome encoding set is the set obtained after converting the path into chromosome codes, which can be used for genetic operations in evolutionary algorithms.

[0083] Specifically, for the initial paths that pass the verification, they are converted into a standardized topological structure representation. First, the nodes in the path are identified and sorted according to the direction of travel. Then, an adjacency matrix is ​​used to record the connection relationships between the nodes. The rows and columns of the adjacency matrix correspond to the nodes in the path, respectively. If there is a connection between two nodes, the corresponding element in the matrix is ​​1; otherwise, it is 0. The adjacency matrix of each path is used as a chromosome code to form the chromosome code set of the initial path population.

[0084] Step S330: Input the path topology structure in the initial path population into the cost evaluation unit, calculate the resource consumption correlation weight between path nodes through a graph neural network, and perform cost feature transfer learning based on historical cost data to generate a multidimensional cost feature distribution.

[0085] The initial path population is a group containing multiple path solutions obtained after validity verification and transformation. The path topology is the structural representation of the nodes and connections within a path. The cost assessment unit is a module in the large-scale transportation and logistics model used to evaluate path costs. Graph neural networks are neural networks specifically designed for processing graph-structured data, capable of learning the features of nodes and edges in a graph. The resource consumption correlation weights between path nodes refer to the degree of correlation in resource consumption between nodes in a path; calculating these weights reveals the resource consumption relationships between nodes. Historical cost data is cost information recorded from past transportation tasks, including time costs and resource costs. Cost feature transfer learning utilizes historical cost data to transfer existing cost feature knowledge to new path assessments, improving the accuracy of cost evaluation. Multidimensional cost feature distribution refers to the distribution of a path across multiple cost factors, such as time costs, resource costs, and efficiency costs.

[0086] Specifically, the path topology from the initial path population is input into the cost evaluation unit. The cost evaluation unit uses a graph neural network to process the path topology. The graph neural network learns the resource consumption correlation weights between path nodes through the transfer and aggregation of node and edge features. Simultaneously, cost feature transfer learning is performed based on historical cost data. Transfer learning algorithms, such as instance-based transfer learning and feature-based transfer learning, can be used. Through transfer learning, cost feature knowledge from historical cost data is applied to new path evaluations. For example, for a new path, the cost of that path is predicted based on the cost features of similar paths in historical cost data. Finally, a multi-dimensional cost feature distribution is generated, including time cost distribution, resource cost distribution, and efficiency cost distribution.

[0087] Step S340: The path generation unit and the cost evaluation unit interact and iterate. The path generation unit adjusts the path variation probability based on the multidimensional cost feature distribution fed back by the cost evaluation unit, and the cost evaluation unit optimizes the cost calculation parameters based on the updated topology structure of the path generation unit.

[0088] The route generation unit and the cost assessment unit are two crucial modules in the large-scale transportation and logistics model. The route generation unit is responsible for generating transportation routes, while the cost assessment unit is responsible for evaluating the costs of those routes. Interactive iteration is a process of continuous information exchange and adjustment between these two modules, gradually optimizing the route plan through multiple iterations. The multidimensional cost feature distribution represents the distribution of the route across multiple cost factors, such as time cost, resource cost, and efficiency cost. The route mutation probability is the probability of performing a mutation operation on the route in the evolutionary algorithm. The route generation unit adjusts the path mutation probability based on the multidimensional cost feature distribution feedback from the cost assessment unit to guide the evolutionary process towards a better outcome. The cost assessment unit optimizes the cost calculation parameters based on the updated topology from the route generation unit, making the cost assessment more accurate.

[0089] In one implementation, step S340 can be implemented as the following steps S341-S345:

[0090] Step S341: Extract multi-dimensional cost features from the topology of each path in the current path population through the cost evaluation unit, generate a cost feature distribution including time consumption, resource consumption and efficiency consumption, and convert it into a relative cost value within a preset numerical range through a normalization algorithm.

[0091] The cost assessment unit analyzes the path topology within the current path population. Multidimensional cost feature extraction extracts cost-related features from multiple dimensions during the path's transportation process. Time consumption refers to the time required to complete the transportation task on that path. Resource consumption encompasses various resources used during transportation, such as fuel and manpower. Efficiency consumption reflects efficiency losses during transportation, such as additional costs incurred due to waiting or congestion. Cost feature distribution describes the distribution of these cost features across different paths. The preset numerical range is typically [0,1] or [-1,1]. Converting cost features to relative cost values ​​facilitates subsequent comparisons and calculations.

[0092] Step S342: Receive relative cost values ​​through the path generation unit and adjust the probability. When the relative cost value exceeds the preset cost threshold, the mutation probability increases linearly with the increase of the difference by a preset ratio. When the relative cost value does not exceed the preset cost threshold, the mutation probability decreases linearly with the increase of the absolute value of the difference by a preset ratio. The adjustment range of the mutation probability is limited to a preset range by a dynamic boundary control mechanism.

[0093] The path generation unit receives relative cost values ​​from the cost evaluation unit and adjusts the mutation probability of the path based on these values. A preset cost threshold is a pre-defined cost limit used to determine if the path cost is too high. When the relative cost value exceeds the preset cost threshold, it indicates that the path cost is high, and the mutation probability needs to be increased to give it more opportunities to generate a better path through mutation. The mutation probability increases linearly with the difference between the relative cost value and the preset cost threshold by a preset ratio. This preset ratio can be adjusted according to actual conditions to control the growth rate of the mutation probability. When the relative cost value does not exceed the preset cost threshold, it indicates that the path cost is low. To maintain path stability, the mutation probability is reduced, and it decreases linearly with the absolute value of the difference by a preset ratio. The dynamic boundary control mechanism ensures that the mutation probability is within a reasonable range, avoiding excessively high or low mutation probabilities. The preset range is usually set based on the characteristics of the evolutionary algorithm and the needs of the actual problem.

[0094] Step S343: In the mutation operation phase of the evolutionary algorithm, the path chromosome is mutated by replacing nodes according to the adjusted mutation probability. A new path topology is generated by selecting road network nodes, and it is ensured that the mutated path is still a valid path.

[0095] Mutation in evolutionary algorithms involves randomly altering chromosomes (representing paths) to introduce new genes (path features) and increase population diversity. A path chromosome represents a path as a chromosome encoding, such as using an adjacency matrix or a sequence of nodes. Node replacement mutation is one form of mutation, where a node is randomly selected in the path and replaced with another suitable node from the network. Network node selection involves choosing a suitable node from the network to replace the original node, considering factors such as reachability and connectivity with other nodes. It is crucial to ensure that the mutated path remains a valid path, meaning it is free of loops and meets resource constraints.

[0096] Specifically, in the mutation phase of the evolutionary algorithm, for each path chromosome, the decision to mutate is made based on the adjusted mutation probability. If mutation is required, a node in the path is randomly selected, and a suitable node in the road network is chosen to replace it.

[0097] Step S344: Input the new path topology generated by the mutation into the cost evaluation unit. Based on the updated topology, the cost evaluation unit recalculates the multidimensional cost feature distribution and feeds the result back to the path generation unit to complete one interactive iteration.

[0098] The newly generated path topology is a different path representation obtained after the mutation operation, differing from the original path in topology. Upon receiving the new path topology, the cost assessment unit recalculates the multidimensional cost feature distribution based on the updated topology. Because the path's topology has changed, its cost characteristics, such as time consumption, resource consumption, and efficiency consumption, also change accordingly. The cost assessment unit uses the previously mentioned cost calculation method, combined with historical data and real-time information, to recalculate the multidimensional cost features of the new path. The recalculated multidimensional cost feature distribution is then fed back to the path generation unit, which adjusts parameters such as the path's mutation probability based on this feedback, completing one interactive iteration.

[0099] Specifically, after the path generation unit completes the mutation operation of the path chromosome, it inputs the new path topology into the cost evaluation unit. Based on the new topology, the cost evaluation unit re-analyzes cost factors such as travel time and fuel consumption of each segment on the path, and calculates a new multidimensional cost feature distribution.

[0100] Step S345: During the iteration process, the evolution trajectory of the path topology and the changing trend of the cost feature distribution are saved through a recording mechanism as a reference for subsequent iteration optimization until the preset number of iterations or the cost feature distribution converges.

[0101] The recording mechanism is a method used to save important information during the iteration process. It can record the evolutionary trajectory of the path topology and the changing trend of the cost feature distribution. The evolutionary trajectory of the path topology records the changes in the path in each iteration, including operations such as adding, deleting, and replacing nodes. The changing trend of the cost feature distribution reflects the changes in the path's cost during the iteration process, such as whether the time cost gradually decreases or increases, and the fluctuation of resource costs. Saving this information as a reference for subsequent iterations helps the algorithm better understand the path optimization process and avoids repeated and ineffective operations. The preset number of iterations is a pre-set upper limit for iterations. When the preset number of iterations is reached, the algorithm stops iterating. Cost feature distribution convergence means that the cost feature distribution no longer changes significantly after multiple iterations, indicating that the path has reached a relatively stable state, at which point iteration can also stop.

[0102] Specifically, a recording mechanism such as a database or file system is used to save information on the path topology and cost feature distribution after each iteration. For example, a database can be used to record the node sequence and multidimensional cost feature values ​​of the path in each iteration. By analyzing these records, an evolutionary trajectory diagram of the path topology and a trend diagram of the cost feature distribution can be plotted. During the iteration process, the changes in the cost feature distribution are monitored in real time. If the change in the cost feature distribution is less than a preset threshold in several consecutive iterations, the cost feature distribution is considered to have converged. The iteration stops when the preset number of iterations is reached or the cost feature distribution converges.

[0103] Step S350: After a preset number of iterations, collect the path topology structure that satisfies the preset constraints on the multidimensional cost feature distribution, generate a data set containing path node sequences, cost composition associations and scenario adaptation scores through scenario adaptability evaluation, and form it into a candidate path scheme population. The number of schemes in the population is kept within a preset range through a screening mechanism.

[0104] The preset number of iterations is a pre-defined upper limit for the evolutionary algorithm's iterations. After multiple iterations, the path scheme will be continuously optimized. A path topology that satisfies preset constraints in terms of multi-dimensional cost feature distribution refers to paths that meet preset constraints in multiple aspects such as time cost, resource cost, and efficiency cost. For example, time cost does not exceed a certain threshold, and resource cost is within a certain range. Scenario adaptability assessment evaluates the adaptability of the path scheme to a preset transportation scenario, considering factors such as the spatial adaptability of the path to the transportation scenario and the degree to which it meets the needs of cargo transportation. A dataset containing path node sequences, cost composition relationships, and scenario adaptability scores is generated. The path node sequences clarify the specific direction of the path, the cost composition relationships illustrate the relationship between different cost factors, and the scenario adaptability score quantifies the adaptability of the path scheme to the transportation scenario. The candidate path scheme population is a set of path schemes that meet the conditions. A screening mechanism is used to control the number of schemes in the population, keeping it within a preset range to avoid excessive population size leading to reduced computational efficiency.

[0105] Step S400: Based on the preset optimization goal, the candidate route scheme population is subjected to multiple rounds of evolutionary screening, and the comprehensive performance evaluation of each scheme is carried out to determine the optimal route scheme that meets the transportation requirements. The evaluation process includes real-time feedback and iterative optimization of the scheme performance.

[0106] The preset optimization objective is a pre-defined optimization direction based on the specific needs of the transportation task, such as minimizing transportation time, minimizing transportation costs, and maximizing transportation efficiency. The candidate route population is a set of multiple candidate route solutions obtained after the previous screening and evaluation steps. Multi-round evolutionary screening is a process of repeatedly screening and optimizing the candidate route population, similar to selection, crossover, and mutation operations in evolutionary algorithms, gradually eliminating unsuitable solutions and retaining better ones. Comprehensive performance evaluation assesses multiple aspects of each candidate route solution, considering factors such as cost, time, and resource utilization efficiency to comprehensively measure the merits of each solution. Determining the optimal route solution that meets the transportation requirements involves selecting the route that best meets the needs of the transportation task from all candidate solutions. Real-time feedback and iterative optimization of solution performance involves providing real-time feedback on the performance of each solution during the evaluation process and iteratively optimizing the solutions based on the feedback information to continuously improve the quality of the solutions.

[0107] In one implementation, step S400 can be implemented as the following steps S410-S470:

[0108] Step S410: Construct a multi-objective optimization function by analyzing the correlation between the set of scenario influencing factors and historical optimization results. The construction process includes adapting the objective weights and adjusting the constraints. The objective weights are optimized in real time according to the urgency of transportation demand.

[0109] The scenario influencing factor set is a collection of factors that influence transportation route planning, considering various factors such as road congestion, weather conditions, and cargo transportation demand. Historical optimization results represent the optimal route solutions obtained from past transportation tasks, along with related cost and time information. Correlation analysis examines the relationship between the scenario influencing factor set and historical optimization results to identify key factors affecting route optimization. The multi-objective optimization function is used to comprehensively consider multiple optimization objectives, including transportation time, transportation cost, and resource utilization efficiency. Objective weight adaptation assigns appropriate weights to each objective based on its importance. Constraint adjustment adjusts constraints during the optimization process according to the actual situation of the transportation task, such as time limits and cargo weight limits. Transportation demand urgency refers to the urgency of the cargo transportation task; real-time optimization of objective weights based on transportation demand urgency makes the multi-objective optimization function more aligned with actual transportation needs.

[0110] Step S420: Initialize the evolutionary screening parameters, set the maximum number of evolutionary rounds and the convergence judgment threshold. The initial value of the evolutionary round counter is 0, and the convergence judgment threshold is determined by the complexity of the objective function.

[0111] Evolutionary selection parameters control the evolutionary selection process, including the maximum number of evolutionary iterations and the convergence threshold. The maximum number of evolutionary iterations is the maximum number of iterations allowed during the evolutionary selection process; the process stops when this maximum number of iterations is reached. The convergence threshold is the criterion used to determine whether the evolutionary selection process has converged. When the change in the objective function value during evolution is less than the convergence threshold, the evolutionary selection process is considered convergent, and it can be stopped at this point. The evolutionary iteration counter records the number of iterations in the evolutionary selection process, with an initial value of 0. Objective function complexity refers to the complexity of a multi-objective optimization function, which is related to the number of optimization objectives and the relationships between them. Determining the convergence threshold based on the objective function complexity can make the convergence judgment more accurate.

[0112] Specifically, the maximum number of evolutionary rounds is set based on the scale and complexity of the transportation task. If the transportation task is large and complex, the maximum number of evolutionary rounds can be set larger; if the transportation task is small and less complex, the maximum number of evolutionary rounds can be appropriately reduced. For the convergence threshold, the complexity of the multi-objective optimization function is analyzed. If the objective function is complex and the relationships between objectives are intricate, the convergence threshold can be set smaller; if the objective function is less complex, the convergence threshold can be appropriately increased. The evolutionary round counter is initialized to 0, and incremented by 1 after each evolutionary selection round.

[0113] Step S430: Calculate the multi-objective optimization function value for each scheme in the candidate path scheme population, and generate a performance evaluation matrix containing the objective function values. The number of rows in the matrix is ​​the number of schemes, and the number of columns is the number of objective functions.

[0114] The candidate path population is a set of multiple candidate path schemes obtained after the previous screening and evaluation steps. The multi-objective optimization function value calculation involves calculating the function value of each candidate path scheme on each optimization objective, based on the multi-objective optimization function. The performance evaluation matrix is ​​a matrix used to record the function values ​​of each candidate path scheme on each optimization objective; the number of rows in the matrix equals the number of candidate path schemes, and the number of columns equals the number of objective functions. Through the performance evaluation matrix, the performance of each candidate path scheme on different optimization objectives can be intuitively compared.

[0115] Specifically, for each candidate route, relevant information such as the sequence of route nodes and cost composition is input into a multi-objective optimization function for calculation. For example, the multi-objective optimization function includes objectives for transportation time, transportation cost, and resource utilization efficiency. For each candidate route, the function values ​​for transportation time, transportation cost, and resource utilization efficiency are calculated. The calculated function values ​​are then filled into a performance evaluation matrix, where each row corresponds to a candidate route and each column corresponds to an objective function.

[0116] It is understandable that, in order to avoid the problem of dimensions, those skilled in the art can standardize the objective function values ​​in the performance evaluation matrix, converting objective function values ​​of different dimensions into standardized values ​​within a preset numerical range before proceeding with subsequent steps. This will not be elaborated here.

[0117] Step S440: The path schemes are classified into different levels using a non-dominated sorting algorithm. The dominance relationship between the schemes is determined by the performance evaluation matrix. Different non-dominated levels are divided, and the schemes within the same level are sorted by congestion distance.

[0118] Non-dominated ranking algorithms are used for ranking in multi-objective optimization problems. They rank multiple candidate solutions according to their non-dominance relationships. The performance evaluation matrix records the function values ​​of each candidate path solution on each optimization objective. By comparing the elements in the performance evaluation matrix, the dominance relationships between solutions can be determined. A dominance relationship means that in multi-objective optimization, if one solution is not inferior to another solution on all optimization objectives and is superior to another solution on at least one optimization objective, then that solution is said to dominate the other. Different non-dominated levels are determined based on the dominance relationships between solutions; solutions in the same non-dominated level do not have a dominance relationship. Crowding distance is an indicator used to measure the distribution density of solutions in the objective space. Solutions within the same level are ranked in descending order of crowding distance. A larger distance indicates a sparser distribution of solutions in the objective space, thus ensuring good diversity in the selected solutions.

[0119] In one implementation, step S440 can be implemented as the following steps S441-S446:

[0120] Step S441: Initialize the dominance count of each path scheme to 0, and the dominance solution set to an empty set. Traverse each pair of schemes i and j in the performance evaluation matrix through a traversal mechanism.

[0121] The dominance count of a path path records the number of times it is dominated by other paths; an initial value of 0 indicates that the path is initially not dominated by any path. The dominance solution set is a collection that records the other paths dominated by a given path. The traversal mechanism refers to the method of comparing each pair of paths in the performance evaluation matrix in a specific order. The performance evaluation matrix records the function values ​​of each path path on various optimization objectives; by traversing each pair of paths in the matrix, the dominance relationships between paths can be determined.

[0122] Specifically, a loop structure is used to implement the traversal mechanism, traversing each pair of solutions i and j in the performance evaluation matrix. For example, two nested for loops are used, with the outer loop traversing solution i and the inner loop traversing solution j. Before the traversal begins, the dominance count of each path solution is initialized to 0, and the dominance solution set is initialized to an empty set.

[0123] Step S442: Compare all objective function values ​​of path schemes i and j. If every objective function value of scheme i is not inferior to that of scheme j, and at least one objective function value is superior to that of scheme j, then it is determined that scheme i dominates scheme j.

[0124] In multi-objective optimization problems, determining the dominance relationship between alternatives requires comparing their function values ​​across all optimization objectives. "Not inferior" means that the function value of alternative i on any optimization objective is greater than or equal to the function value of alternative j, while "superior" means that the function value of alternative i on any optimization objective is greater than the function value of alternative j. If alternative i is not inferior to alternative j on all optimization objectives and is superior to alternative j on at least one optimization objective, then alternative i can be considered to dominate alternative j.

[0125] Specifically, the function values ​​of scheme i and scheme j on each optimization objective are obtained from the performance evaluation matrix and compared one by one. For example, the performance evaluation matrix has three optimization objectives: transportation time, transportation cost, and resource utilization efficiency. By comparing the function values ​​of scheme i and scheme j on these three objectives, if the transportation time of scheme i is less than or equal to the transportation time of scheme j, the transportation cost of scheme i is less than or equal to the transportation cost of scheme j, the resource utilization efficiency of scheme i is greater than or equal to the resource utilization efficiency of scheme j, and scheme i is strictly superior to scheme j on at least one objective, then scheme i is determined to dominate scheme j.

[0126] Step S443: If solution i dominates solution j, add solution j to the dominance solution set of solution i, and increment the dominance count of solution j by 1. The counting process is carried out through an accumulation mechanism.

[0127] If solution i dominates solution j, it means that solution i is better than solution j in multi-objective optimization. Add solution j to the dominance set of solution i and record the solutions dominated by solution i. Increment the dominance count of solution j by 1. The dominance count is used to record the number of times solution j is dominated by other solutions. The counting process is carried out by an accumulation mechanism, that is, each time a new solution dominates solution j, the dominance count is increased by 1.

[0128] Specifically, when it is determined that solution i dominates solution j, solution j is added to the dominance solution set of solution i. For example, a list or set data structure can be used to store the dominance solution set of solution i, and the identifier of solution j can be added to that list or set. Simultaneously, the dominance count of solution j is incremented. A variable can be used to store the dominance count of solution j, and the value of this variable is incremented by 1 each time a new dominance relationship occurs.

[0129] Step S444: After comparing all scheme pairs, the scheme with a dominance count of 0 is classified as the first non-dominant level and temporarily removed from the population through a marking mechanism.

[0130] After comparing all pairs of schemes, the dominance count for each scheme is determined. Schemes with a dominance count of 0 are not dominated by any other schemes; these schemes are the best in the current population and are classified as the first non-dominated level. The marking mechanism is used to temporarily remove these schemes from the population. By marking these schemes, they can be temporarily ignored in subsequent processing, allowing for further ranking of the remaining schemes. Specifically, iterate through the dominance counts of all schemes and find those with a dominance count of 0. Mark these schemes as the first non-dominated level. An array or list can be used to record the level information of each scheme, setting the level of schemes with a dominance count of 0 to 1. Simultaneously, the marking mechanism is used to temporarily remove these schemes from the population. For example, a boolean array can be used to mark whether a scheme has been removed, setting the boolean value corresponding to the scheme at the first non-dominated level to true.

[0131] Step S445: Recalculate the dominance count for the remaining schemes, and classify the schemes with a new dominance count of 0 into the second non-dominated level. Repeat this process until all schemes are classified into the corresponding non-dominated level.

[0132] After temporarily removing schemes from the population at the first non-dominant level, the dominance count is recalculated for the remaining schemes. Since the removal of some schemes may alter the dominance relationships, a recalculation of the dominance count is necessary. Schemes with a new dominance count of 0 represent those not dominated by other schemes among the remaining schemes and are classified as the second non-dominant level. This process is repeated, continuously calculating dominance counts and classifying the remaining schemes, until all schemes are assigned to their respective non-dominant levels.

[0133] Specifically, the dominance count is recalculated for the remaining schemes using the same method as in steps S441-S443. Schemes with a new dominance count of 0 are identified and classified as the second non-dominated level. The scheme level information is updated, and schemes in the second non-dominated level are temporarily removed from the remaining schemes using a marking mechanism. The remaining schemes are processed again, repeating the above process, until all schemes are classified into the corresponding non-dominated level.

[0134] Step S446: Calculate the congestion distance for each non-dominated level. The congestion distance reflects the distribution density of the scheme in the target space. The larger the distance, the sparser the distribution of the scheme in the target space. Schemes in the same non-dominated level are sorted in descending order of congestion distance. When the congestion distances are the same, the order is determined by comparing the path lengths to form a complete level classification result.

[0135] Crowding distance is an indicator used to measure the density of solutions in the target space. It helps select solutions that are more dispersed in the target space to ensure solution diversity. Within each undominated level, the crowding distance of solutions is calculated. For each solution, the difference between its boundary values ​​on each objective function and its neighboring solutions is calculated, and these differences are summed to obtain the crowding distance. Solutions within the same undominated level are sorted in descending order of crowding distance; a larger distance indicates a sparser distribution of the solution in the target space, and prioritizing these solutions ensures better solution diversity. When crowding distances are the same, the ranking is determined by comparing path lengths; solutions with shorter path lengths are generally better.

[0136] Specifically, for each scheme within a non-dominated level, the congestion distance is calculated using the congestion distance calculation formula. For example, for a multi-objective optimization problem with three optimization objectives, the difference between the boundary value of each scheme and its adjacent schemes at each optimization objective is calculated, and these three differences are added together to obtain the congestion distance. Schemes within the same non-dominated level are sorted in descending order of congestion distance. If the congestion distances are the same, the path lengths of the schemes are compared, and the scheme with the shorter path length is ranked first. In this way, a complete level division result is formed.

[0137] Step S450: Select the scheme with high non-dominance level and large crowding distance as the parent generation, generate offspring schemes through crossover and mutation operations, merge the offspring and parent schemes, re-divide and sort the levels, and retain the preset number of optimal schemes to form a new population.

[0138] A scheme with a high non-dominated level indicates a better scheme in multi-objective optimization, while a scheme with a large crowding distance indicates a sparser distribution in the objective space. Selecting a scheme with both a high non-dominated level and a large crowding distance as the parent scheme ensures that the selected scheme has both good optimization performance and good diversity. Crossover involves exchanging some genes between two parent schemes to generate new offspring schemes; mutation involves randomly changing the genes of the schemes to increase their diversity. After merging the offspring and parent schemes, the schemes are re-divided and sorted according to the methods described in steps S440-S446, assigning all schemes to different non-dominated levels and sorting them by crowding distance within each level. A preset number of optimal schemes are retained to form a new population. The preset number can be set according to actual conditions to control the population size.

[0139] Specifically, from the ranking results, schemes with high non-dominated levels and large crowding distances are selected as parents. For example, the top few schemes with large crowding distances in the first non-dominated level are selected as parents. Crossover and mutation operations are performed on the parent schemes to generate offspring schemes. In the crossover operation, two parent schemes are randomly selected, and some of their path node sequences are exchanged; in the mutation operation, a path node in the scheme is randomly changed. The offspring schemes are merged with the parent schemes, and the ranking and sorting are performed again. According to a preset number, schemes with high levels and large crowding distances are selected to form a new population.

[0140] Step S460: Increment the evolution round counter by 1, and repeat the performance evaluation, ranking and selection operations until the maximum evolution round or the change in the optimal solution of the population in multiple consecutive rounds is less than the convergence judgment threshold.

[0141] An evolutionary round counter records the number of evolutionary selection rounds, incrementing by 1 after each round. Performance evaluation involves calculating the multi-objective optimization function value for each scheme, generating a performance evaluation matrix. Rank division uses a non-dominated sorting algorithm to divide schemes into different non-dominated ranks, and sorts them by crowding distance within each rank. Selection involves choosing schemes with high non-dominated ranks and large crowding distances as parents, generating offspring schemes through crossover and mutation operations, and merging them to form a new population. The maximum number of evolutionary rounds is a pre-set maximum number of iterations for evolutionary selection; the process stops when this maximum number of rounds is reached. If the change in the optimal solution of the population is less than the convergence threshold for multiple consecutive rounds, the evolutionary selection process has converged, and can be stopped at this point.

[0142] Specifically, after each round of evolutionary selection, the evolutionary round counter is incremented by 1. The effectiveness of each scheme in the new population is evaluated, generating a new effectiveness evaluation matrix. A non-dominated sorting algorithm is used to rank the schemes, sorting them according to their crowding distance within the same rank. Schemes with high non-dominated ranks and large crowding distances are selected as parents, and offspring schemes are generated through crossover and mutation operations, then merged to form a new population. The change in the optimal solution of the population over multiple consecutive rounds is monitored in real time. If the change is less than the convergence threshold, or if the evolutionary round counter reaches the maximum evolutionary round, the evolutionary selection process stops.

[0143] Step S470: Extract the set of schemes with the highest non-dominant level from the population obtained by the final evolution, calculate the comprehensive effectiveness value of each scheme through the comprehensive effectiveness evaluation algorithm, and select the scheme with the best comprehensive effectiveness value as the optimal path scheme. The comprehensive effectiveness value is generated by weighted summation of the multi-objective function values.

[0144] The final evolved population is a set of multiple candidate path solutions obtained after multiple rounds of evolutionary selection. The set of solutions with the highest non-dominant level is the best set of solutions in the population, and these solutions have good performance in multi-objective optimization. The comprehensive performance evaluation algorithm is used to calculate the comprehensive performance value of the solutions. The comprehensive performance value is generated by weighted summation of the multi-objective function values, and the weighting coefficients can be set according to the importance of different optimization objectives. The solution with the optimal comprehensive performance value is selected as the optimal path solution, which has the best performance considering multiple optimization objectives.

[0145] In one implementation, step S470 can be implemented as the following steps S471-S475:

[0146] Step S471: Identify the highest non-dominated level in the final population, and obtain all the schemes within this level through an extraction mechanism to form a candidate optimal solution set. The candidate optimal solution set includes the path topology, multidimensional cost feature distribution, and the values ​​of each objective function.

[0147] The final population is a set of candidate path solutions obtained after multiple rounds of evolutionary selection. The highest non-dominated level is the level of the best solutions in the population, which exhibit good performance in multi-objective optimization. The extraction mechanism is used to obtain all solutions within the highest non-dominated level from the final population. The candidate optimal solution set is a set containing all solutions within the highest non-dominated level. This set includes important information such as the path topology, multi-dimensional cost feature distribution, and objective function values ​​of each solution. This information is crucial for subsequent performance evaluation and optimal solution selection.

[0148] Specifically, by examining the rank information of the solutions, the highest non-dominated rank in the final population is identified. For example, in the rank classification results, the rank with the smallest rank value is the highest non-dominated rank. Using an extraction mechanism, such as traversing the final population, solutions with the highest non-dominated rank are selected, and these solutions are grouped into a candidate optimal solution set. For each solution, its path topology is recorded, such as node sequence and adjacency matrix; multi-dimensional cost feature distribution is recorded, such as time cost distribution and resource cost distribution; and the values ​​of various objective functions are recorded, such as the function values ​​of transportation time, transportation cost, and resource utilization efficiency.

[0149] Step S472: Determine the weight coefficients of each objective function based on the transportation demand vector in the core element information. The weight coefficients are generated by the demand urgency matching algorithm, and the objective function corresponding to the demand with high urgency is assigned a higher weight.

[0150] The transportation demand vector within the core element information contains various demand information for freight transportation tasks, such as transportation time requirements and the impact of cargo characteristics on transportation costs. The weight coefficients of each objective function are used to measure the importance of different objective functions when calculating the overall efficiency value. The demand urgency matching algorithm generates weight coefficients based on the urgency of transportation demand; higher urgency demands are assigned higher weights to their corresponding objective functions, thus making the calculation of the overall efficiency value more consistent with actual transportation needs.

[0151] Step S473: Standardize the objective function values ​​of each scheme in the candidate optimal solution set, convert the objective function values ​​of different dimensions into standardized values ​​within a preset numerical range, and adjust the normalization parameters according to the type of objective function.

[0152] The objective function values ​​of each solution in the candidate optimal solution set typically have different units; for example, the unit of transportation time might be hours, the unit of transportation cost might be yuan, and the unit of resource utilization efficiency might be a percentage. Preset numerical ranges are, for example, [0,1] or [-1,1]. Normalization parameters are adjusted according to the type of objective function; different types of objective functions may require different normalization methods. For example, for a transportation time objective, a minimum-maximum normalization method can be used; for a resource utilization efficiency objective, a Z-score normalization method can be used.

[0153] Step S474: Calculate the overall effectiveness value by summing the standardized values ​​of each objective function with their corresponding weight coefficients. The sum of the weight coefficients is 1, and the result is taken as the overall effectiveness value of the scheme.

[0154] The overall performance value is an indicator used to comprehensively measure the performance of a scheme in multi-objective optimization. It is obtained by weighted summation of the standardized values ​​of each objective function and their corresponding weight coefficients. The standardized values ​​of each objective function are obtained after standardization in step S473, and the corresponding weight coefficients are determined based on the transportation demand vector in step S472. The sum of the weight coefficients is 1 to ensure the reasonableness and comparability of the weighted summation result.

[0155] Step S475: Compare the overall performance values ​​of all options and select the option with the highest overall performance value as the optimal path option. If there are multiple options with the same overall performance value, the final option is determined through a random selection mechanism.

[0156] Comparing the overall performance values ​​of all solutions aims to identify the optimal solution in multi-objective optimization. The solution with the highest overall performance value indicates that it has the best performance when considering multiple optimization objectives. If multiple solutions have the same overall performance value, it means that these solutions perform similarly in multi-objective optimization. In this case, a random selection mechanism is used to determine the final solution to ensure fairness and randomness in the selection process.

[0157] Specifically, iterate through all solutions in the candidate optimal solution set and compare their combined performance values. Record the identifier of the solution with the largest combined performance value. If multiple solutions have the same combined performance value and are all the maximum, use a random selection mechanism, such as a random number generator, to randomly select one of these solutions as the final optimal path solution.

[0158] Step S500: Convert the optimal route plan into instructions, generate a transportation scheduling instruction set, and push it to the transportation execution terminal through a real-time data interaction interface. The instruction conversion process includes instruction format adaptation and terminal interaction response verification.

[0159] The optimal route plan is the best transportation route obtained after multiple rounds of screening and evaluation, containing important information such as route topology and multi-dimensional cost characteristic distribution. Instruction conversion is the process of converting the optimal route plan into instructions that the transportation execution terminal can understand and execute. The transportation scheduling instruction set is a collection of various instructions for transportation tasks, such as route information, time control information, and resource allocation information. The real-time data interaction interface is used for real-time data exchange with the transportation execution terminal, through which the transportation scheduling instruction set can be pushed to the terminal. Instruction format adaptation converts instructions into a format that conforms to the requirements of the transportation execution terminal; different terminals may have different instruction format requirements. Terminal interaction response verification checks whether the transportation execution terminal has successfully received and responded to the instruction after it has been pushed, ensuring the effective execution of the instruction.

[0160] As one implementation method, step S500 can be implemented as the following steps S510-S560:

[0161] Step S510: Analyze the path topology in the optimal path scheme, obtain the path node sequence, the connection relationship between nodes and the resource consumption correlation weight through the node extraction mechanism, and generate a path execution flow model. Each node in the model contains location correlation and time window features.

[0162] The optimal route plan contains the best transportation route information, where the route topology describes the nodes and connections of the route. Parsing the route topology is the process of analyzing and extracting the topological information of the route. The node extraction mechanism is a method used to obtain the route node sequence, the connections between nodes, and the resource consumption correlation weights from the route topology. The route node sequence is the sequential arrangement of the nodes in the route; the connections between nodes describe the connectivity between nodes; and the resource consumption correlation weights represent the degree of correlation between nodes in terms of resource consumption. The route execution flow model is a model generated based on the route node sequence, the connections between nodes, and the resource consumption correlation weights. This model contains detailed information about route execution; each node includes location correlation and time window features. Location correlation describes the node's location in geographic space, and the time window feature specifies the node's arrival time range.

[0163] Specifically, analytical algorithms are used to analyze the path topology in the optimal path scheme. For example, if the path topology is represented by an adjacency matrix, the path node sequence and inter-node connections are obtained by traversing the adjacency matrix. Resource consumption correlation weights can be obtained from the multi-dimensional cost feature distribution. A node extraction mechanism, such as a depth-first search algorithm, is used to traverse the path topology and extract the path node sequence, inter-node connections, and resource consumption correlation weights. Based on this information, a path execution flow model is generated. For each node, its location correlations, such as latitude and longitude coordinates, and time window features, such as earliest arrival time and latest arrival time, are recorded.

[0164] Step S520: Based on the path execution process model and multi-dimensional cost feature distribution, generate a data set that includes time control association, resource allocation association and exception handling association, and the data format is adapted through the terminal protocol.

[0165] The path execution process model includes detailed information about path execution, such as the path node sequence, inter-node connections, and time window characteristics. The multidimensional cost feature distribution describes the distribution of time and resource costs within the path. Time control correlations provide time control information for each node and segment of the path, such as node dwell time and segment travel time. Resource allocation correlations provide information on resource allocation across nodes and segments, such as the allocation of manpower, fuel, and equipment. Anomaly handling correlations are strategies for handling potential anomalies during transportation, such as delay and fault handling. The dataset is a collection containing information related to time control, resource allocation, and anomaly handling. Terminal protocol adaptation converts the dataset format to a format that conforms to the requirements of the transportation execution terminal protocol; different transportation execution terminals may have different protocol format requirements.

[0166] As one implementation method, step S520 can be implemented as the following steps S521-S525:

[0167] Step S521: Traverse the path to obtain the node sequence of the process model. Based on the resource consumption correlation weight between nodes and the driving characteristics of the transportation vehicle, generate the estimated travel time between adjacent nodes through a time calculation algorithm to form a node arrival time window. The node arrival time window includes the correlation between the earliest and latest arrival times.

[0168] The node sequence in the path execution flow model describes the sequential arrangement of nodes in the path. The resource consumption correlation weight between nodes represents the degree of correlation in resource consumption between nodes, which is closely related to the travel status of the transportation vehicles. Transportation vehicle travel characteristics refer to the performance characteristics of the transportation vehicles, such as speed and acceleration. The time calculation algorithm is used to calculate the estimated travel time between adjacent nodes, considering factors such as the resource consumption correlation weight between nodes and the travel characteristics of the transportation vehicles. The node arrival time window refers to the range of arrival times for a node, including the correlation between the earliest and latest arrival times, providing time constraints for transportation scheduling.

[0169] Specifically, the node sequence of the path execution flow model is traversed. For two adjacent nodes, the estimated travel time is calculated using a time calculation algorithm based on the resource consumption correlation weight between the nodes and the vehicle's driving characteristics. For example, if the resource consumption correlation weight between nodes is large, it indicates that there may be congestion or other factors affecting travel, and the estimated travel time is increased accordingly. Vehicle driving characteristics, such as speed, can be used as the basic parameters for calculating the estimated travel time. Based on the calculated estimated travel time and the path's start time, a node arrival time window is formed. The earliest arrival time can be calculated based on the fastest travel scenario, and the latest arrival time can be calculated based on the slowest travel scenario.

[0170] Step S522: Based on the time consumption characteristics in the multidimensional cost feature distribution, calculate the dwell time and travel speed of each node through the time control parameter generation algorithm to form time control correlation data. The parameters are kept consistent through the time coupling mechanism.

[0171] The time consumption feature in the multidimensional cost feature distribution describes the time consumption of a path at different nodes and road segments, reflecting the time cost during transportation. The time control parameter generation algorithm is used to calculate the dwell time at each node and the travel speed on the road segment based on the time consumption feature. The dwell time at each node refers to the time the vehicle spends at the node, and the travel speed on the road segment refers to the speed at which the vehicle travels on the road segment. The time control associated data is a dataset containing information such as the dwell time at each node and the travel speed on the road segment. The time coupling mechanism is used to ensure consistency between various time control parameters; for example, the dwell time at a node and the travel speed on the road segment need to be coordinated to ensure the time rationality of the entire transportation process.

[0172] Specifically, a time-controlled parameter generation algorithm is used to calculate the dwell time of each node and the travel speed of each route segment based on the time consumption characteristics in the multidimensional cost feature distribution. For example, for a logistics delivery route, the time consumption characteristics indicate that a certain route segment has a high time consumption, possibly due to traffic congestion or a long distance. In this case, the travel speed of the route segment can be appropriately reduced, and the dwell time of each node can be reasonably arranged to balance the time cost. During the calculation process, a time coupling mechanism is used to ensure the consistency between the parameters. For example, if the dwell time of a node increases, in order to ensure that the entire transportation task is completed within the specified time, it may be necessary to increase the travel speed of subsequent route segments accordingly.

[0173] Step S523: Based on the resource consumption characteristics in the multidimensional cost feature distribution, generate resource allocation schemes for each node and path segment through a resource allocation algorithm, including the allocation associations of manpower, fuel and equipment, and the allocation schemes satisfy resource constraints.

[0174] The resource consumption characteristics in the multidimensional cost feature distribution reflect the consumption of various resources, such as manpower, fuel, and equipment, at different nodes and road segments. The resource allocation algorithm is based on these resource consumption characteristics and aims to optimize resource utilization while meeting transportation task requirements. The resource allocation scheme for each node and road segment clarifies the quantity and method of manpower, fuel, and equipment to be allocated at each node and road segment. Resource constraints refer to the limitations that must be followed during resource allocation, such as vehicle load limits and fuel storage capacity limits. The allocation scheme must meet these constraints to ensure the feasibility and safety of the transportation task.

[0175] Specifically, the resource consumption characteristics in the multidimensional cost feature distribution are analyzed first to determine the resource requirements of each node and path segment. For example, for long-distance transportation routes, some segments may require more fuel due to complex road conditions, while some nodes may require more manpower due to large cargo loading and unloading volumes. Then, resource allocation algorithms are used for resource allocation. Available resource allocation algorithms include greedy algorithms and dynamic programming algorithms. Taking the greedy algorithm as an example, at each step, the seemingly optimal resource allocation scheme is selected, gradually completing the resource allocation for the entire path. During the allocation process, resource constraints are strictly checked.

[0176] Step S524: Combining the risk characteristics in the set of scenario impact factors, formulate rules for handling abnormal situations through an anomaly handling strategy generation algorithm, including associated strategies for delay handling, fault handling, and path adjustment, and adapt the strategies to the scenario risk level.

[0177] The risk characteristics in the scenario impact factor set reflect various risk situations that may be encountered during transportation, such as weather changes, road construction, and vehicle breakdowns. The anomaly handling strategy generation algorithm is an algorithm that formulates rules for dealing with anomalies based on these risk characteristics. These anomaly handling rules cover related strategies for delay handling, fault handling, and route adjustment. Delay handling strategies are used to address time delays during transportation, such as rescheduling schedules and adjusting the priority of subsequent tasks; fault handling strategies address vehicle or equipment malfunctions, such as arranging repairs or replacing vehicles; route adjustment strategies are used to replan transportation routes when encountering road closures or severe congestion. Strategy adaptation to scenario risk levels refers to formulating corresponding strength and complexity of response strategies based on different risk levels; the higher the risk level, the more stringent and detailed the response strategy.

[0178] Step S525: Couple the time control associated data, resource allocation associated scheme and exception handling associated strategy through a data fusion algorithm to generate a basic instruction data set containing multi-dimensional associated features. The field associations of the data set are calibrated through the terminal data protocol.

[0179] The time control associated data includes information on time control aspects such as dwell time at each node and travel speed along the route; the resource allocation associated scheme clarifies the allocation of resources such as manpower, fuel, and equipment at each node and route segment; the anomaly handling associated strategy formulates rules for handling anomalies such as delays, failures, and route adjustments. The data fusion algorithm integrates and couples these three different types of data, aiming to generate a dataset containing multi-dimensional associated features, enabling these data to be interconnected and synergistic. The instruction base dataset is a dataset rich in information obtained after data fusion, which can be used to generate transportation scheduling instructions. Field association calibration via terminal data protocol refers to adjusting the association relationships of each field in the dataset according to the protocol requirements of the transportation execution terminal, ensuring that the data can be correctly identified and processed by the terminal.

[0180] As one implementation method, step S525 can be implemented as the following steps S5251-S5255:

[0181] Step S5251: Align the time control associated data with the time axis, convert the time window features of different nodes into time markers on a unified time axis, and determine the time axis accuracy through the transportation task cycle.

[0182] Time control-related data includes time window characteristics of different nodes, such as the earliest arrival time, latest arrival time, and dwell time. These time window characteristics may be based on different time references or have different levels of precision. Time axis alignment is the process of converting these different node time window characteristics into time markers on a unified time axis, ensuring consistency and comparability of time information. Time axis precision is determined by the transportation task cycle; when the transportation task cycle is long, the time axis precision can be relatively low; when the transportation task cycle is short, the time axis precision needs to be high to ensure the accuracy of time control.

[0183] Step S5252: Normalize the resource allocation association scheme according to the resource dimension, and convert the allocation data of different types of resources into allocation weights of the same resource dimension. The total weights are 1. The allocation weights are generated through resource importance assessment.

[0184] The resource allocation and correlation scheme includes allocation data for different types of resources, such as manpower, fuel, and equipment. These resources have different properties and units of measurement. To facilitate data fusion and comprehensive analysis, resource dimension normalization is required. Resource dimension normalization is the process of converting allocation data for different types of resources into allocation weights for a unified resource dimension. This allows for comparison and processing of different resource allocations on the same dimension. The sum of the allocation weights is 1, ensuring the rationality and balance of resource allocation. Resource importance assessment is the process of evaluating the importance of different resources based on the needs of the transportation task and the characteristics of the resources. The allocation weights are generated through the resource importance assessment; that is, the more important the resource, the higher its allocation weight.

[0185] Step S5253: Prioritize the anomaly handling strategies. Determine the order of strategies based on the probability of anomaly occurrence and the degree of impact. Priority is calculated using a risk assessment algorithm.

[0186] The anomaly handling strategy includes response rules for different anomalies, such as delay handling, fault handling, and route adjustment. Different anomalies have different probabilities of occurrence and varying degrees of impact; therefore, these anomaly handling strategies need to be prioritized. Prioritization determines the order in which the corresponding response strategies should be executed when an anomaly occurs. The probability of an anomaly occurring refers to the likelihood of such an anomaly happening during transportation, while the degree of impact refers to the extent of loss or impact the anomaly causes to the transportation task. The risk assessment algorithm is an algorithm that calculates the priority of anomaly handling strategies based on the probability of anomaly occurrence and the degree of impact.

[0187] Step S5254: Use a multimodal fusion algorithm to correlate and couple the time control, resource allocation and anomaly handling data to generate a fusion feature vector containing multidimensional correlations of time, resources and risk. The vector dimension is optimized by a data compression algorithm.

[0188] Time control data includes time-related information such as time window characteristics, dwell time, and travel speed along the route segment for each node; resource allocation data clarifies the allocation of resources such as manpower, fuel, and equipment; and anomaly handling data outlines strategies for handling various abnormal situations. The multimodal fusion algorithm is an algorithm that associates and couples these three different types of data. By mining the inherent relationships between the data, it generates a fused feature vector containing multi-dimensional correlations of time, resources, and risk. This fused feature vector integrates information from multiple dimensions such as time, resources, and risk, providing a more comprehensive description of the transportation task. The vector dimension is optimized through a data compression algorithm to reduce data redundancy and dimensionality, improving the efficiency and accuracy of data processing.

[0189] Specifically, multimodal fusion algorithms, such as deep learning-based multimodal fusion networks, are used to correlate and couple time control, resource allocation, and anomaly handling data. This network can learn the relationships between different data types, fusing them into a single feature vector. For example, it correlates node arrival times in time control data with resource allocation times in resource allocation data, and correlates response strategies in anomaly handling data with time and resource adjustments. After generating the fused feature vector, data compression algorithms are used for dimensionality optimization. Suitable data compression algorithms include Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA).

[0190] Step S5255: Convert the fused feature vector into a structured data format, generate a basic data set corresponding to the fields of the instruction template through a field mapping mechanism, and record the relationship between the fields in the data set through an association matrix.

[0191] The fused feature vector, generated by a multimodal fusion algorithm, contains multidimensional correlations of time, resources, and risk, representing relevant information about a transportation task in a comprehensive way. Structured data formats, such as tables and database records, are data formats with clear structures and rules, facilitating storage, transmission, and processing. The field mapping mechanism maps elements in the fused feature vector to instruction template fields, which are predefined fields used to generate transportation scheduling instructions. Through the field mapping mechanism, the fused feature vector can be converted into a basic dataset that conforms to the requirements of the instruction template. The relationships between fields in the dataset are recorded through an association matrix, which clearly represents the relationships between fields, providing a reference for instruction generation and processing.

[0192] Specifically, the field structure and meaning of the instruction template are first determined. For example, the instruction template may contain fields such as path information, time control, resource allocation, and exception handling. Then, according to the field mapping mechanism, the elements in the fused feature vector are mapped to the corresponding instruction template fields. For example, time dimension elements in the fused feature vector are mapped to the time control field, and resource dimension elements are mapped to the resource allocation field. The mapped elements are then used to form a basic dataset. Simultaneously, an association matrix is ​​used to record the relationships between the fields in the dataset. The rows and columns of the association matrix correspond to different fields, and the elements in the matrix represent the strength or type of relationship between the fields.

[0193] Step S530: Call the instruction template generator to fill the corresponding fields of the template with the basic data set, and generate a structured scheduling instruction containing an instruction header, path information segment, resource allocation segment and control logic segment. The order of the fields is adjusted according to the terminal interaction habits.

[0194] The instruction template generator is a pre-designed tool for generating transportation scheduling instructions, containing the basic structure and format of the instructions. The basic data set is a structured dataset containing information such as time, resources, and exception handling, obtained after the previous steps. Filling the template with the corresponding fields of the basic data set involves inserting the values ​​of each field from the dataset into the corresponding positions in the instruction template, giving the template actual content. The structured scheduling instruction is a complete instruction generated, containing an instruction header, a path information segment, a resource allocation segment, and a control logic segment. The instruction header contains basic information such as the instruction number and sending time; the path information segment describes detailed information about the transportation path; the resource allocation segment describes the allocation of resources such as manpower, fuel, and equipment; and the control logic segment contains logical rules for time control, exception handling, etc. The field order is adjusted according to terminal interaction habits. Different transportation execution terminals may have different interaction habits; adjusting the field order makes the instructions easier for the terminal to receive and process.

[0195] Specifically, the instruction template generator is invoked; this generator can be a software module or program. The values ​​of each field in the basic dataset are filled into the corresponding fields of the instruction template according to the field mapping relationships. For example, the path information field value from the basic dataset is filled into the path information segment of the instruction template, and the resource allocation field value is filled into the resource allocation segment. After generating the structured scheduling instructions, the field order is adjusted according to the terminal's interaction habits. If the transport execution terminal typically receives path information first and then resource allocation information, then the path information segment is placed before the resource allocation segment.

[0196] Step S540: Verify the structured scheduling instructions by performing syntax verification through syntax rule matching and consistency checks through logical association analysis. Errors found during verification are corrected through an automatic repair mechanism.

[0197] Structured scheduling instructions may contain syntax errors or logical inconsistencies during generation, necessitating validation. Syntax rule matching compares the structured scheduling instructions against predefined syntax rules to check for compliance. Syntax rules can include instructions' format, field types and lengths, keyword usage, and other related rules. Logical correlation analysis analyzes the logical relationships between the various fields in the instruction, checking for logical contradictions or irrationalities. For example, it checks for logical consistency between time control and resource allocation fields, and whether exception handling strategies match path information and time control. An automatic correction mechanism automatically corrects errors detected during validation, attempting to fix problems in the instruction based on error type and rules.

[0198] Specifically, a syntax rule matching tool is used to perform syntax validation on structured scheduling instructions. This tool can be a regular expression engine or a syntax parser, checking the format and field validity of the instructions according to predefined syntax rules. For example, it checks whether the instruction number in the instruction header conforms to the numbering rules, and whether the node number in the path information segment is a valid number. Simultaneously, logical correlation analysis is performed to analyze the logical relationships between the various fields in the instruction. For example, it checks whether the node arrival time in the time control field and the resource allocation time in the resource allocation field are consistent. If the node arrival time is too early, and the resource has not yet been allocated, there is a logical inconsistency. For errors found during validation, an automatic repair mechanism is used to correct them. For example, if syntax validation finds that the length of a field exceeds the specified range, the automatic repair mechanism can truncate the excess part; if logical correlation analysis finds inconsistencies between time and resource allocation, the automatic repair mechanism can adjust the time or resource allocation parameters.

[0199] Step S550: Convert the verified structured instructions into a binary instruction format supported by the terminal, add check codes and time stamp information, generate a transportation scheduling instruction set, and adapt the instruction set length through the terminal's receive buffer.

[0200] Structured instructions that pass verification are those that have been confirmed to be error-free after syntax and logical consistency checks. The binary instruction format supported by the terminal is the instruction format that the transportation execution terminal can understand and execute; different terminals may have different binary instruction format requirements. Converting structured instructions to binary instruction format is the process of converting text-based instructions into binary encoding, facilitating terminal processing. A checksum is a code used to verify whether errors occurred during instruction transmission. By calculating the checksum and adding it to the instruction, the terminal can verify the instruction upon receipt, ensuring its integrity. Timestamp information records the instruction's generation time, helping the terminal understand the instruction's timeliness. The transportation scheduling instruction set is a collection containing all processed instructions. The instruction set length is adapted through the terminal's receive buffer, a memory area used by the terminal to receive instructions. Adjusting the instruction set length ensures that the instruction set can be completely received by the terminal, avoiding reception errors due to an excessively long instruction set.

[0201] Specifically, a binary encoding tool is used to convert the verified structured instructions into a binary instruction format supported by the terminal. For example, if the terminal supports a certain binary protocol, the structured instructions are encoded according to the format of that protocol. The checksum of the instruction is calculated using an available checksum algorithm, such as the Cyclic Redundancy Check (CRC) algorithm. The checksum is appended to the end of the binary instruction. Simultaneously, a timestamp is added to record the time the instruction was generated. The instruction set length is adjusted according to the size of the terminal's receive buffer. If the instruction set length exceeds the size of the terminal's receive buffer, the instruction set can be split and sent to the terminal in multiple parts.

[0202] Step S560: Establish a secure communication connection with the transport execution terminal, push the scheduling instruction set to the terminal through the real-time data interaction interface, receive the instruction reception confirmation information returned by the terminal, and re-push through the retransmission mechanism if no confirmation is received.

[0203] Establishing a secure communication connection with the transportation execution terminal is a crucial step in ensuring the safe and reliable transmission of instructions. Secure communication connections can utilize encryption technologies, such as SSL / TLS protocols, to encrypt transmitted data and prevent theft or tampering. The real-time data interaction interface is used for real-time data exchange with the transportation execution terminal, through which transportation scheduling instruction sets can be pushed to the terminal. Receiving instruction reception confirmation information from the terminal confirms whether the terminal has successfully received the instructions. If no confirmation information is received, it indicates that the instructions may have been lost during transmission or that the terminal has encountered a problem. In this case, a retransmission mechanism is needed to re-push the instructions to ensure that they are correctly received and executed by the terminal.

[0204] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solutions of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes here.

[0205] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present invention, such as... Figure 2As shown, the hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can run on the processor 1001, and the processor 1001 executes the program to implement the steps in the method of any of the above embodiments.

[0206] The memory 1002 stores computer programs that can run on the processor. The memory 1002 is configured to store instructions and applications that can be executed by the processor 1001. It can also cache data to be processed or already processed (e.g., image data, audio data, voice communication data, and video communication data) of the processor 1001 and various modules in the computer system 1000. It can be implemented by flash memory or random access memory (RAM).

[0207] When the processor 1001 executes the program, it implements the steps of the logistics transportation optimization method based on the large-scale transportation and logistics model mentioned above. The processor 1001 typically controls the overall operation of the computer system 1000.

Claims

1. A logistics transportation optimization method based on a large-scale transportation and logistics model, characterized in that, The method includes: Receive basic scheduling instructions for logistics transportation tasks, extract elements from the basic scheduling instructions, and generate core element information including the spatial association of the origin and destination of the path and the cargo transportation requirements through semantic scene parsing; Based on the core element information, transportation scenarios are adapted, and a set of scenario influence factors required for transportation route planning is generated by combining the deep feature mining results of the historical transportation scenario database. Specifically, this includes: performing grid-based modeling of the spatial association between the origin and destination of the path in the core element information to generate a spatial grid index that matches the coordinate system of the historical transportation scenario database; extracting demand features from the core element information to generate a demand vector containing the urgency of transportation demand and the adaptability of goods, wherein the dimension of the demand vector is the same as the feature dimension of the scenario influence factors; calling the feature mining unit of the historical transportation scenario database to perform similar scenario retrieval based on the spatial grid index, and extracting the implicit association patterns between scenario features and transportation results through an association rule mining algorithm, wherein the association patterns are represented by a feature interaction weight matrix; inputting the implicit association patterns and the demand vector into the temporal convolutional network in the scenario situation coupling model for scenario state evolution analysis, generating an influence weight distribution containing time decay characteristics, wherein the time series length of the weight distribution matches the transportation task cycle; coupling the implicit association patterns, influence weight distribution, and spatial grid features through a feature interaction algorithm to generate a set of scenario influence factors with spatiotemporal characteristics, wherein the feature dimensions of each scenario influence factor in the set are consistent. The pre-trained large-scale transportation and logistics model is invoked to make collaborative path decisions on the set of influencing factors of the scenario. Through the interactive iteration of the path generation unit and the cost evaluation unit within the model, a population of candidate path solutions containing path topology and cost feature distribution is output. Based on the preset optimization objective, the candidate route scheme population is subjected to multiple rounds of evolutionary screening, and the comprehensive performance evaluation of each scheme is carried out to determine the optimal route scheme that meets the transportation requirements. The optimal route plan is converted into instructions to generate a transportation scheduling instruction set, which is then pushed to the transportation execution terminal through a real-time data interaction interface.

2. The method according to claim 1, characterized in that, The feature mining unit that calls the historical transportation scenario database performs similar scenario retrieval based on the spatial grid index and extracts implicit association patterns between scenario features and transportation results through an association rule mining algorithm, including: The spatial grid index is converted into an entity identifier sequence that can be recognized by the feature mining unit, and a query statement containing the origin and destination entities of the path and the cargo type entities is constructed. The entity relationship of the query statement is adjusted by time-series association weight. The feature mining unit performs a depth-first scenario traversal based on the query statement, retrieves historical transportation scenario nodes associated with the origin and destination entities, records the scenario feature attributes and transportation result feedback data of each historical transportation scenario node, and adapts the traversal depth according to the scenario complexity. The retrieved historical scene nodes are subjected to feature standardization processing, which converts scene feature attributes of different dimensions into standardized feature values ​​of a unified range. The association rule mining algorithm is used to perform association analysis between standardized feature values ​​and transportation result feedback data. Key association rules are filtered by support-confidence threshold, and a rule set containing antecedent feature items, consequent result items and association strength is generated. Redundant rules are pruned from the rule set, and semantically repetitive association rules are eliminated by calculating rule coverage.

3. The method according to claim 2, characterized in that, The correlation analysis between standardized feature values ​​and transportation result feedback data, and the screening of key correlation rules using a support-confidence threshold, includes: The standardized feature values ​​are discretized to generate a set of feature terms, and continuous feature values ​​are converted into discrete feature labels by dividing the feature intervals. The transportation result feedback data is categorized into result categories to generate result level labels, which are dynamically related to the degree of transportation demand satisfaction. Initialize the set of individual scene feature units as an empty set, scan the historical scene database through a sliding window to calculate the occurrence frequency of individual scene feature units, and filter significant individual scene feature units through a frequency threshold, wherein the frequency threshold is adaptively adjusted according to the number of scene features; The combined scene feature units containing multiple features are generated from significant single scene feature units through a layer-by-layer iterative approach. Combined scene feature units containing non-significant features are removed through pruning operations. The occurrence frequency of combined scene feature units is calculated and significant combined scene feature units are filtered out through a threshold. For each significant combination of scene feature units, all possible non-empty true subsets are generated as rule antecedents, and the remaining parts are used as rule consequents. Key association rules are filtered by confidence threshold to generate a rule set containing antecedent feature items, consequent result items, and association strength. The association strength is calculated by weighted fusion of occurrence frequency and confidence.

4. The method according to claim 1, characterized in that, The process involves calling a pre-trained large-scale transportation and logistics model to perform collaborative path decision-making on the set of influencing factors of the scenario. Through the interactive iteration between the path generation unit and the cost evaluation unit within the model, a population of candidate path solutions containing path topology and cost feature distribution is output, including: The set of scene influencing factors is input into the feature encoding layer of the transportation and logistics big model. The correlation weights between scene features are calculated through a multi-head self-attention interaction mechanism to generate a scene feature context vector. Initial path sampling is performed using the path generation unit of the large transportation and logistics model. An initial path population containing different topological structures is generated based on the scene feature context vector. Invalid paths are removed through path validity verification. The population size is adjusted according to the scene complexity. The path topology in the initial path population is input into the cost evaluation unit. The resource consumption correlation weights between path nodes are calculated through a graph neural network. Based on historical cost data, cost feature transfer learning is performed to generate a multidimensional cost feature distribution. The path generation unit and the cost evaluation unit interact and iterate. The path generation unit adjusts the path variation probability based on the multidimensional cost feature distribution fed back by the cost evaluation unit, and the cost evaluation unit optimizes the cost calculation parameters based on the updated topology structure of the path generation unit. After a preset number of iterations, the path topology structure that satisfies the preset constraints on the multidimensional cost feature distribution is collected. Through scenario adaptability evaluation, a data set containing path node sequences, cost composition associations and scenario adaptation scores is generated and formed into a candidate path scheme population.

5. The method according to claim 4, characterized in that, The initial path sampling is performed using the path generation unit of the large-scale transportation and logistics model, and an initial path population containing different topological structures is generated based on the scene feature context vector, including: The spatial association features in the scene feature context vector are analyzed to determine the location codes of the path start and end points in the road network topology. The location codes are adjusted by the road network node association weights. Initialize the evolutionary algorithm parameters, setting the population size, crossover probability, mutation probability, and maximum number of generations. The population size is adapted to the scene complexity level, which is determined by the number of dimensions of the spatially related features. An initial population is created through a path generation strategy. Valid paths connecting the origin and destination are generated by selecting road network nodes, ensuring that each path is free of loops and that the number of nodes is within a threshold range. The threshold is adaptively adjusted according to the road network density. The validity of each path in the initial population is verified, the constraint adaptability is evaluated based on the resource constraint characteristics in the set of scene influence factors, invalid paths that violate hard constraints are eliminated, and new paths are generated through a supplementary mechanism until the preset population size is reached. The validated initial path is converted into a standardized topological structure representation. The connection relationships between path nodes are recorded through an adjacency matrix, with the node order arranged according to the driving direction, forming a chromosome encoding set for the initial path population.

6. The method according to claim 5, characterized in that, The step involves interactive iteration between the path generation unit and the cost assessment unit. The path generation unit adjusts the path variation probability based on the multidimensional cost feature distribution fed back by the cost assessment unit, including: The cost assessment unit extracts multidimensional cost features from the topology of each path in the current path population, generates a cost feature distribution that includes time consumption, resource consumption and efficiency consumption, and converts it into a relative cost value within a preset range through a normalization algorithm. The path generation unit receives relative cost values ​​and performs probability adjustments. When the relative cost value exceeds a preset cost threshold, the mutation probability increases linearly with the increase of the difference by a preset ratio. When the relative cost value does not exceed the preset cost threshold, the mutation probability decreases linearly with the increase of the absolute value of the difference by a preset ratio. The adjustment range of the mutation probability is limited to a preset range by a dynamic boundary control mechanism. During the mutation phase of the evolutionary algorithm, the path chromosome is mutated by replacing nodes according to the adjusted mutation probability. A new path topology is generated by selecting road network nodes, and it is ensured that the mutated path is still a valid path. The new path topology generated by the mutation is input into the cost evaluation unit. The cost evaluation unit recalculates the multidimensional cost feature distribution based on the updated topology and feeds the result back to the path generation unit, completing one interactive iteration. During the iteration process, the evolution trajectory of the path topology and the changing trend of the cost feature distribution are saved through a recording mechanism, which serves as a reference for subsequent iteration optimization until the preset number of iterations or the cost feature distribution converges.

7. The method according to claim 1, characterized in that, The process of performing multiple rounds of evolutionary screening on the candidate route scheme population based on a preset optimization objective, and conducting a comprehensive performance evaluation of each scheme to determine the optimal route scheme that meets transportation requirements includes: A multi-objective optimization function is constructed by correlating the set of scenario influencing factors with historical optimization results. The construction process includes the adaptation of objective weights and the adjustment of constraints. The objective weights are optimized in real time according to the urgency of transportation demand. Initialize the evolutionary screening parameters, set the maximum number of evolutionary rounds and the convergence judgment threshold, the initial value of the evolutionary round counter is 0, and the convergence judgment threshold is determined by the objective function complexity. For each candidate path scheme in the population, multi-objective optimization function values ​​are calculated to generate a performance evaluation matrix containing the values ​​of each objective function. The number of rows in the matrix is ​​the number of schemes, and the number of columns is the number of objective functions. The path solutions are ranked using a non-dominated sorting algorithm. The dominance relationship between the solutions is determined by the efficiency evaluation matrix. Different non-dominated levels are divided, and the solutions within the same level are sorted by congestion distance. Select the scheme with high non-dominance level and large crowding distance as the parent generation, generate offspring schemes through crossover and mutation operations, merge the offspring and parent schemes, re-classify and sort them, and retain a preset number of optimal schemes to form a new population. The evolution round counter is incremented by 1, and the efficiency evaluation, ranking and selection operations are repeated until the maximum evolution round is reached or the change in the optimal solution of the population in multiple consecutive rounds is less than the convergence judgment threshold. The set of schemes with the highest non-dominant level is extracted from the population obtained from the final evolution. The comprehensive efficiency value of each scheme is calculated by a comprehensive efficiency evaluation algorithm. The scheme with the best comprehensive efficiency value is selected as the optimal path scheme. The comprehensive efficiency value is generated by weighted summation of multi-objective function values.

8. The method according to claim 7, characterized in that, The process of classifying path solutions into levels using a non-dominated ranking algorithm, determining the dominance relationship between solutions using a performance evaluation matrix, and classifying different non-dominated levels includes: Initialize the dominance count of each path scheme to 0, and the dominance solution set to an empty set. Then, traverse each pair of schemes i and j in the performance evaluation matrix using a traversal mechanism. Compare all objective function values ​​of path schemes i and j. If every objective function value of scheme i is not inferior to that of scheme j, and at least one objective function value is superior to that of scheme j, then scheme i is determined to dominate scheme j. If solution i dominates solution j, solution j is added to the dominant solution set of solution i, and the dominance count of solution j is incremented by 1. The counting process is carried out through an accumulation mechanism. After comparing all pairs of schemes, schemes with a dominance count of 0 are classified as the first non-dominant class and temporarily removed from the population through a marking mechanism. Recalculate the dominance count for the remaining schemes, and classify the schemes with a new dominance count of 0 into the second non-dominated level. Repeat this process until all schemes are classified into the corresponding non-dominated level. For each non-dominated level, the congestion distance is calculated. The congestion distance reflects the distribution density of the scheme in the target space. The larger the distance, the sparser the distribution of the scheme in the target space. Schemes in the same non-dominated level are sorted in descending order of congestion distance. When the congestion distances are the same, the order is determined by comparing the path lengths to form a complete level classification result.

9. A computer system comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

Images