Intelligent car dispatching method, system and equipment based on large model dynamic parameters

Abstract

The invention discloses an intelligent vehicle dispatching method, system and device based on large model dynamic parameters, and relates to the technical field of vehicle dispatching and path planning. Large model parameters are deeply bound with business data, the large model parameters are automatically generated through scene recognition based on real-time dynamic data and historical task data, dynamic adjustment of the large model parameters is achieved through a parameter rule self-optimization mechanism, a model is generated to generate a vehicle dispatching and path planning scheme, and a detection model is rechecked and automatically corrected. According to the method, scene specificity is adapted through dynamic parameter adjustment, the hallusion rate of a large model is reduced through double-model cooperation, compliance is improved through multi-dimensional verification, and the method is suitable for logistics self-distribution business of commerce wholesale and manufacturing enterprises.

Description

Technical Field

[0001] This invention relates to the field of vehicle dispatching and route planning technology, specifically applied to the logistics self-delivery business scenarios of wholesale and manufacturing enterprises. It is particularly suitable for the delivery of goods from fixed warehouses to the outskirts of cities, as well as the dispatching management and route planning of enterprise-owned vehicles, enabling precise matching of tasks and vehicles, dynamic route optimization, and compliant dispatching. Background Technology

[0002] In the daily operations of wholesale and manufacturing enterprises, companies typically rely on their own warehouses (such as retail forward warehouses and manufacturing factory warehouses) to deliver goods to multiple distribution points (including stores, distributors, and end users) within a radius of 1-60 kilometers. These company-owned trucks exhibit significant differences in physical attributes and operational limitations, specifically in load capacity variations of 1-5 tons, driving range limitations of 5-30 kilometers, and the differentiation of various vehicle types such as vans, tricycles, and refrigerated trucks. Precise matching of warehouse-to-multiple distribution point vehicle dispatch and route planning, delivery task demand, and vehicle carrying capacity becomes crucial to ensuring the efficient operation of the supply chain.

[0003] Existing intelligent vehicle dispatching technologies in logistics suffer from the following technical problems: First, the task matching efficiency of manual dispatching is low. In complex scenarios with multiple vehicles and multiple delivery tasks, relying on human experience to match and allocate vehicles and tasks is often time-consuming and difficult to adapt to dynamic changes in delivery tasks. Human error can easily lead to resource waste or task delays. Second, traditional algorithms have significant limitations in dispatching and route planning. They have weak dynamic adaptability and cannot respond in real time to dynamic factors such as changes in road conditions and adjustments to delivery task priorities. Their ability to balance multiple constraints is insufficient. Under multi-dimensional constraints such as load limits, time window requirements, and delivery cost control, they are prone to getting trapped in local optima or experiencing solution space explosion, exhibiting poor robustness and impacting business operations. The model is overly sensitive to abnormal data in its operations; third, when a single large model is applied to intelligent vehicle dispatch, there is a result illusion problem. In medium-to-large-scale scenarios with more than 20 delivery tasks, the stability and accuracy of the model output results drop sharply, which can lead to problems such as task conflicts where the same delivery task is assigned to two vehicles, fake task data that is not in the input task set, and abnormal phenomena such as the total load of the vehicle exceeding the rated limit. There is a lack of effective constraint closed-loop verification mechanism, and large models with fixed parameters are difficult to adapt to the industry specificity of logistics self-delivery business data. There are significant differences in delivery needs, vehicle type classification, vehicle restriction standards, and task demand levels in different industries, which leads to instability in the final calculation results of the model.

[0004] Therefore, there is an urgent need for an intelligent scheduling technology that can adapt to dynamic scenarios, balance multi-dimensional constraints, and ensure reliable results, in order to improve the operational efficiency and compliance of enterprises' self-delivery logistics business. Summary of the Invention

[0005] To reduce the probability of large model illusions and improve the accuracy of computational results, this application provides an intelligent vehicle dispatching method, system, and device based on the dynamic parameters of a large model.

[0006] To achieve the above-mentioned objectives, the present invention provides the following technical solution: A smart vehicle dispatching method based on large model dynamic parameters includes the following steps: S1. Collect and standardize the business data required for vehicle dispatch to form standardized business data. Based on the standardized business data, construct a prompt word template and generate a complete prompt word. S11. Collect and standardize the business data required for vehicle dispatch to form standardized business data, which includes warehouse data, truck data set, task data set and route data; S12, construct a prompt word template based on the standardized business data, fill the standardized business data into the prompt word template, and generate a complete prompt word; S2, construct a scenario-based parameter template library, collect real-time dynamic data of the current task, extract historical task data, automatically generate large model parameters based on the real-time dynamic data and the historical task data, dynamically adjust the large model parameters using a parameter rule self-optimization mechanism, and update the scenario-based parameter template library; S21, construct a scenario-based parameter template library, which includes at least scenario labels and corresponding large model parameter baseline values, trigger thresholds, standard scenario feature vectors, feature sensitivity weights, and business weight coefficients for each scenario label. S22, collect real-time dynamic data of the current task, extract historical task data, extract dynamic scene data and task feature data from the real-time dynamic data, and form key feature data for scene recognition; S23, a scene recognition algorithm is used to identify the scene of the current task. The scene recognition algorithm adopts a dual mechanism of threshold determination and similarity matching. S231, Set a trigger threshold determination function for each scenario, substitute the real-time dynamic data into the trigger threshold determination function to determine the threshold, filter out candidate scenarios that meet the threshold conditions, and form a candidate scenario set. S232, Match scene tags according to the candidate scene set. If the candidate scene set contains only one candidate scene, then use its corresponding tag as the matching scene tag for the current task and call the corresponding parameter baseline value. If the candidate scene set is an empty set or contains multiple conflicting scene tags, then perform similarity matching. S233, based on the key feature data, obtain a 7-dimensional feature vector, call the standard scene feature vector, calculate the cosine similarity between the 7-dimensional feature vector and the standard scene feature vector, select the scene label corresponding to the scene with the highest feature vector similarity as the matching scene label for the current task, and call the parameter benchmark value corresponding to the scene label. The cosine similarity calculation formula is: In the formula, A represents the 7-dimensional feature vector extracted from real-time dynamic data based on the current task data, and B represents the standard scene feature vector in the scenario-based parameter template library. It is the dot product of two vectors. It is the product of the magnitudes of two vectors; S24. Based on the scene recognition results, the basic parameter generation formula is used to calculate the final parameters and output them. The final parameter calculation formula is as follows: In the formula, The parameter baseline value is obtained from the scenario-based parameter template library. Let i be the feature sensitivity weight of the i-th dimension. Initial values ​​are pre-built into the template library based on experience, and are subsequently updated via S252. This represents the deviation between the feature vector of the i-th dimension in the current task and the feature vector of the standard scene. S25, based on the execution results of the historical task data, execute the parameter rule self-optimization mechanism to correct the feature sensitivity weight; S251, Based on the historical task data, obtain the actual dispatch effect data after the vehicle dispatch plan is executed, and calculate the comprehensive effect score of the current task based on the actual dispatch effect data. The comprehensive effect score formula is: In the formula, For compliance indicators, As an indicator of on-time performance, The full load rate is an indicator. For excess cost indicators, These are preset business weighting coefficients; S252, extract the average of the historical task comprehensive performance scores for each scenario as the historical baseline score for that scenario. Using the incremental update rule in reinforcement learning, compare the current comprehensive performance score with the historical baseline score for that scenario, and calculate the new feature sensitivity weights, which are used for the next parameter calculation. The feature sensitivity weight correction formula is as follows: In the formula, The currently used feature sensitivity weights, For learning rate, To score the overall effect of this test, Score based on historical benchmarks. As a reward signal, This represents the deviation between the feature vector of the i-th dimension in the current task and the feature vector of the standard scene. S26, perform cluster analysis on the monthly new business data to generate new template vectors, or when the actual vehicle dispatch effect data of a certain scenario is lower than the preset threshold for one consecutive month, trigger S25 to correct the feature sensitivity weight and update the scenario-based parameter template library. S3, Based on the complete prompt words and the large model parameters, use the generative model to generate a preliminary vehicle dispatch and route planning scheme; S31, Call the large language model API interface, input the complete prompt word constructed in S1, the generation model controls the inference process of the model based on the final parameters calculated in step S24, and outputs the result; S32 extracts each data block of the output result through regular expressions, parses it into standard structured data objects, and forms a preliminary vehicle dispatching and route planning scheme; S4. Based on the complete prompt words, the detection model is used to review and automatically correct the preliminary vehicle dispatch and route planning scheme; S41, the preliminary vehicle dispatch and route planning scheme is used as the data to be checked, and a verification prompt word is constructed by combining the constraints of the complete prompt word; S42, the detection model operates under low temperature parameters and low Top-p parameters, and compares the preliminary vehicle dispatch and route planning scheme with the constraints. S43, the detection model outputs a structured feedback report based on the comparison results; S44. If the comparison result is fully compliant, a structured report with a status of "passed" will be output, and the process will proceed to the next step, S5. S45, if the comparison result is non-compliant, a structured report with an error status is output. The structured report with an error status includes the error status, error code, target vehicle identifier, current load, load limit and correction suggestions, and triggers an automatic correction mechanism. S46, parse the correction suggestion field in the report, remove the non-compliant task from the corresponding target vehicle, and reassign the task; S5, verify the output results of the generation model and the detection model; S6, supplement other data of the preliminary vehicle dispatch and route planning scheme, and generate and output intelligent vehicle dispatch and route planning scheme.

[0007] This invention also provides an intelligent vehicle dispatching system based on large model dynamic parameters, comprising: The business data assembly layer is used to collect and standardize the business data required for vehicle dispatch, forming standardized business data. Based on the standardized business data, a prompt word template is constructed to generate a complete prompt word. The large model parameter intelligent generation layer is used to build a scenario-based parameter template library, collect real-time dynamic data of the current task, extract historical task data, automatically generate large model parameters based on the real-time dynamic data and the historical task data, and update the scenario-based parameter template library using a parameter rule self-optimization mechanism. The dual-model collaborative computing layer includes a generation model and a detection model. The generation model generates a preliminary vehicle dispatch and route planning scheme based on the complete prompt words and the large model parameters. The detection model reviews and automatically corrects the preliminary vehicle dispatch and route planning scheme based on the prompt word template. The rule processing layer is used to verify the output results of the dual-model collaborative computation layer; The results output layer is used to complete the data of the preliminary vehicle dispatch and route planning scheme, and generate and output the intelligent vehicle dispatch and route planning scheme.

[0008] The intelligent generation layer for large model parameters includes: The key data extraction module is used to collect real-time dynamic data of the current task and extract historical task data, and extract dynamic scene data and task feature data from the real-time dynamic data to form key feature data for scene recognition. The feature extraction and scene recognition unit extracts the feature vector of the current task based on the key feature data and uses a scene recognition algorithm to match scene labels. The parameter generation rule engine is used to calculate the parameters of the large model based on the scene recognition results and the final parameter calculation formula. It also uses the execution results of the historical task data to execute the parameter rule self-optimization mechanism and correct the feature sensitivity weights. The parameter output verification module is used to calculate the final parameters based on the scene recognition results using the basic parameter generation formula, and correct the scene template feature sensitivity weights through a feedback correction mechanism. The scenario-based parameter template library module is used to build a scenario-based parameter template library and update the scenario-based parameter template library based on new business data or actual vehicle dispatch effect data.

[0009] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor; wherein, when the processor executes the computer program stored in the memory, it implements the aforementioned intelligent vehicle dispatching method based on large model dynamic parameters.

[0010] Compared with the prior art, the present invention has the following significant advantages: This invention adapts to specific scenarios by dynamically adjusting the parameters of a large model, enabling it to respond in real time to dynamic factors such as changes in road conditions and adjustments to delivery task priorities. It effectively solves the problem that fixed parameter modes in existing technologies cannot cope with complex dynamic scenarios, significantly improves the dynamic adaptability of logistics dispatching systems, and ensures the real-time effectiveness of dispatching schemes. This invention adopts a dual-model collaborative architecture of generative model and detection model, which can solve the problem of decreased stability and accuracy of output results of a single large model in medium and large-scale scenarios. It can effectively reduce the illusion rate of large models and effectively avoid the occurrence of abnormal phenomena such as task conflicts, falsified task data, and overload. It provides double protection for the reliability of vehicle dispatching scheme. This invention constructs a multi-dimensional verification system. By verifying the output results of the generation model and the detection model, and with the addition of a parameter rule self-optimization mechanism, it can ensure that the scheduling results strictly meet the multi-dimensional constraints such as load limits, time window requirements, and delivery cost control, thereby comprehensively improving the compliance of logistics scheduling and reducing the operational risks and economic losses caused by non-compliant scheduling. This invention achieves deep binding between large model parameters and business data. By identifying and matching scenarios through current task data and calling the corresponding scenario-based parameter template library data, it automatically generates large model parameters and makes dynamic adjustments. This effectively solves the technical pain point that large models with fixed parameters are difficult to adapt to the specificities of different industries, significantly improving the stability and adaptability of model calculation results, and can flexibly adapt to the logistics scheduling needs of different industries. This invention is applicable to self-delivery logistics operations in various scenarios such as wholesale trade and manufacturing enterprises. It can achieve precise matching of tasks and vehicles, dynamic route optimization, and compliant scheduling, effectively improving the operational efficiency of enterprise self-delivery logistics operations, reducing logistics operating costs, and possessing high industrial application value. Attached Figure Description

[0011] Figure 1 This is an architecture diagram of an intelligent vehicle dispatching system based on large model dynamic parameters provided by the present invention; Detailed Implementation

[0012] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Example 1: This example provides an intelligent vehicle dispatching method based on large model dynamic parameters, including the following steps: S1. Collect and standardize the business data required for vehicle dispatch to form standardized business data. Based on the standardized business data, construct a prompt word template and generate a complete prompt word. The specific steps are as follows: S11. Collect and standardize the business data required for vehicle dispatch to form standardized business data. The standardized business data includes warehouse data, truck data set, task data set, and route data. The warehouse data includes warehouse ID and latitude and longitude coordinates. The truck data set includes vehicle ID, attributes, maximum load capacity, latitude and longitude coordinates, and special restrictions (such as delivery within a 5-kilometer radius). The task data set includes task ID, customer ID, cargo load capacity, cargo type, latitude and longitude coordinates of the delivery point, time requirement, and mileage from the warehouse to the delivery point. The route data includes a set of routes already traveled. Based on the customer ID in the task, query the set of routes already traveled related to this customer. Each route is a set of customer IDs.

[0014] In a preferred embodiment of the present invention, the mileage from the warehouse to the distribution point is obtained in real time through the open interface of the map platform, and the map platform will plan the route based on the real-time distance and return the mileage data.

[0015] In a preferred embodiment of the present invention, the set of routes already traveled is obtained by analyzing the set of customers on historical actual travel routes, and is used as reference data for the current task.

[0016] S12, construct a prompt word template based on the standardized business data, fill the standardized business data into the prompt word template, and generate a complete prompt word.

[0017] The prompt word template adopts a four-layer structure of "role-data-logic-format", the specific content of which is as follows: #Intelligent Dispatch Task for Logistics and Delivery (ReqID:{requestId}) ## I. Roles and Goals You are a senior logistics dispatch expert with 20 years of experience. Your core objective is to deliver a delivery plan that minimizes total mileage, vehicle usage, and load factor while meeting stringent constraints.

[0018] ## II. Input Data 1. Warehouse data |Warehouse ID|Warehouse Name|Geographic Coordinates| |---|---|---| |{warehouse[id]}|{warehouse[name]}|{warehouse[location]}| 2. Vehicle Resource Pool |Vehicle ID|License Plate|Load Limit|Special Restrictions| |---|---|---|---| {vehicle_info} 3. Task pool to be delivered Note: Only the {task_count} task IDs in the following list are valid; data outside the list is strictly prohibited.

[0019] |Task ID|Customer ID|Requirement Quantity|Geographic Coordinates|Warehouse Distance (km)| |---|---|---|---|---| {distributiontask_info} 4. Preset grouping constraints (existing customer routes) If the customer IDs in the task belong to the following groups, they should be prioritized for being assigned to the same train: {customergroups_str} ##III. Core Scheduling Logic (Thinking Chain) Step 1: Task Classification and Preprocessing Please scan the "Task Pool" first and mark the tasks into the following three categories: 1. Return task: Tasks with demand < 0. Logic: Path needs to be included, and demand is algebraically accumulated.

[0020] 2. Special Case: Tasks where [Warehouse Distance to Point Mileage] is empty. Logic: No need to calculate the estimated mileage; prioritize assigning to the vehicle with the largest remaining load capacity.

[0021] 3. Regular tasks: Tasks other than the two categories mentioned above.

[0022] Step 2: Vehicle allocation and route planning (greedy strategy) 1. Priority: Prioritize fulfilling customers in the same vehicle for delivery as specified in the "Preset Grouping Constraints".

[0023] 2. Clustering and Loading: Clustering is based on [geographic coordinates] proximity. The total load of a single vehicle must be less than or equal to the vehicle's maximum load capacity; the loading rate of a single vehicle should ideally be greater than or equal to 80%.

[0024] 3. Route sorting: Follow the principle of "warehouse -> nearest delivery point -> ... -> warehouse". The mileage of the first point is taken as [distance from warehouse to point], and the point-to-point mileage is estimated based on coordinates.

[0025] ## IV. Hard Constraints (Violation results in task failure) 1. ID Integrity: The task IDs output in the results must strictly exist in the input "task pool to be delivered". Fabricating IDs is strictly prohibited.

[0026] 2. Uniqueness: Each valid task ID can only appear once in the final result.

[0027] 3. Single route: One vehicle can only correspond to one set of aggregated tasks.

[0028] ##V. Output Format Specifications Please strictly follow the following CSV format to output 3 data blocks, wrapping them with the specified tags: Data Block 1: Vehicle Allocation Results vehicleGroups vehicleId, id, distance, estimateDistance [Vehicle ID], [Task ID], [Warehouse Distance], [Estimated Mileage] vehicleGroups-end Data Block 2: Unable to deliver task undeliverableTasks id,unDeliverableReason undeliverableTasks-end Data Block 3: Idle Vehicles undeliverable Vehicles vehicleId undeliverable Vehicles - end S2, construct a scenario-based parameter template library, collect real-time dynamic data of the current task, extract historical task data, automatically generate large model parameters based on the real-time dynamic data and the historical task data, dynamically adjust the large model parameters using a parameter rule self-optimization mechanism, and update the scenario-based parameter template library. The specific steps are as follows: S21, construct a scenario-based parameter template library, which includes at least scenario labels and corresponding large model parameter baseline values, trigger thresholds, standard scenario feature vectors, feature sensitivity weights, and business weight coefficients for each scenario label. S22: Collect real-time dynamic data of the current task, extract historical task data, and use the real-time dynamic data and historical task data as key data for generating parameters. Extract dynamic scene data and task feature data from the real-time dynamic data to form key feature data for scene recognition. The key feature data includes road congestion index, cargo type, proportion of urgent orders, time requirements, task load, mileage, vehicle load, and delivery radius. In a preferred embodiment, S21 includes: (1) The real-time dynamic data includes dynamic scene data and task feature data. The dynamic scene data includes road congestion index, cargo type, proportion of emergency orders, and time requirements. The task feature data includes task load, mileage, vehicle load, and delivery radius.

[0029] (2) The historical task data includes business scenario feature data and parameter effect data obtained from the business system (ERP). The business scenario feature data includes cargo type (obtained from the product file), task load, mileage (obtained from the delivery task or order), time requirement (obtained from the order), vehicle load (obtained from the vehicle file), delivery radius (obtained from the vehicle file), road congestion index (obtained by real-time map open interface), and emergency order ratio (number of emergency orders in the delivery task / total number of orders). The parameter effect data includes actual vehicle dispatch effect data and large model parameter data.

[0030] S23, A scene recognition algorithm is used to identify the scene of the current task. The scene recognition algorithm adopts a dual mechanism of threshold determination and similarity matching. The specific steps are as follows: S231, Set a trigger threshold determination function for each scenario. The scenario-based parameter template library contains preset trigger thresholds for each scenario, and a global scenario set is defined based on the collected real-time dynamic data. The real-time dynamic data is substituted into the trigger threshold determination function. A threshold is determined to filter out candidate scenes that meet the threshold conditions, thus forming a candidate scene set. .

[0031] The decision logic is as follows: High congestion scenario determination : in, This is the real-time road congestion index (value 0-1). Set a preset congestion threshold (e.g., 0.6).

[0032] Major promotion / emergency scenario assessment : in, The percentage of urgent orders, , The preset percentage threshold (e.g., 0.3).

[0033] Cold chain fresh food scenario judgment : in, For the type of goods, Temperature requirements.

[0034] Determination of short-distance transport of dangerous goods : in, For delivery radius, This is the short-distance threshold (e.g., 20km).

[0035] S232, Match scene tags according to the candidate scene set. The candidate scene set output in step S231 is: ,like If there is only one candidate scene, the scene label is directly matched, and the parameter baseline value corresponding to that scene in the scene-based parameter template library is called; if or If the candidate scene set is empty or contains multiple conflicting scene labels, then a feature vector similarity matching step is performed to further complete accurate scene matching.

[0036] S233, Based on the key feature data, a 7-dimensional feature vector is obtained, and the standard scene feature vector is called. The standard scene feature vector is a feature vector stored in the scene parameter template library that can represent the typical business form of the scene. The standard scene feature vector is generated by cluster analysis of the task data corresponding to the best dispatch effect in the historical task data.

[0037] Based on the key feature data, extract feature vectors of 7 dimensions from the current task data. Calculate its relationship with the feature vectors of each standard scene in the scene-based parameter template library. The cosine similarity is used to select the scene with the highest feature vector similarity based on the calculation results, and its corresponding label is used as the matching scene label for the current task.

[0038] set up Let A be a feature vector with 7 dimensions extracted from the current task data. Let B be the standard scene feature vector of the i-th scene in the scene-based parameter template library. Where n=7 represents the 7 dimensions of the feature vector, and the 7 dimensions are as follows: (Goods type): e.g., General = 0.1, Fresh / Cold = 0.5, Dangerous Goods = 0.9. Based on the discrete mapping of delivery difficulty, a mapping dictionary is constructed: Example of a mapping dictionary: 0.1: General dried goods / daily necessities (no special requirements) 0.3: Fragile items / large appliances (handle with care to prevent drops) 0.5: Cold chain fresh food (requires refrigerated truck, 0-4℃) 0.8: Frozen food (requires refrigerated truck, -18℃) 0.9: Hazardous chemicals / medical devices (requires vehicles with special qualifications) (Average Task Load): The normalized value (0-1), mapped to the [0,1] interval using max-min normalization. in, The average weight of the current list of tasks to be assigned (calculated in real time); The minimum weight limit for a single delivery set for the system (e.g., 0 kg); This sets the maximum single delivery weight limit for the system (e.g., the maximum vehicle load capacity, such as 10,000 kg). If the real-time calculated result is greater than 1 (e.g., an oversized order), it will be forcibly truncated to 1. (Time window strictness): The percentage of orders with time requirements.

[0039] (Average Vehicle Load Capacity): Normalized average tonnage of currently available vehicles. in, The average rated load of currently available (idle and in good working order) vehicles; This is the rated load capacity of the largest vehicle in the company's fleet (e.g., if the company's largest vehicle is 15 tons, then the denominator is 15,000 kg). A calculated value close to 0.2 indicates that the currently available vehicles are mostly small vans / tricycles (lightweight transport capacity), while a value close to 0.9 indicates that the currently available vehicles are mostly heavy trucks (heavyweight transport capacity). This feature helps the model determine "whether large vehicle matching logic is needed".

[0040] (Delivery radius): Based on proportional normalization of the maximum service radius, used to measure the dispersion and distance span of delivery tasks. in, This represents the average navigation distance from all current task points to the warehouse. The maximum service radius covered by the vehicle (a hard boundary defined by the business, e.g., 60km for deliveries around the city). Specifically: (Assuming the vehicle's maximum service radius is 60km) A calculated value <0.3 (within approximately 18km) indicates short-distance intensive delivery, while a value >0.8 (beyond approximately 48km) indicates long-distance sparse delivery.

[0041] (Congestion Sensitivity Coefficient): Road Congestion Index.

[0042] (Percentage of urgent orders): Number of urgent orders / Total number of orders.

[0043] The formula for calculating cosine similarity is: In the formula, The dot product of two vectors represents the sum of the corresponding products of the two vectors in each dimension. It is the product of the magnitudes of two vectors (Euclidean Norm), used to eliminate the influence of magnitude and focus only on the direction of the vectors (i.e. the trend of the feature distribution).

[0044] Calculation results The value range is [-1, 1]. In the feature vector of the logistics scenario, since the value is usually non-negative, the result is usually between [0, 1]. The closer the value is to 1, the more similar the current task scenario is to the template scenario.

[0045] Iterate through all scene sets in the scenario-based parameter template library. Calculate the corresponding similarity set: Select The scene label corresponding to the largest value is used as the final recognition result, and the baseline values ​​of the large model temperature parameter and Top-p parameter in that scene are called.

[0046] S24, Based on the scene recognition results, use the basic parameter generation formula to calculate the final parameters and output them. The calculation formula is: In the formula, The parameter baseline value is obtained from the scenario-based parameter template library; The feature sensitivity weight for the i-th dimension (e.g., congestion sensitivity weight of 0.1). This represents the deviation between the feature vector of the i-th dimension in the current task and the feature vector of the standard scene. A specific calculation example uses temperature as an example: (Benchmark values ​​for temperature parameters in high-congestion scenarios) (Congestion-sensitive weight) (Current real-time congestion index) (Trigger threshold) Calculation result: 0.2 + 0.1 × (0.8 - 0.6) = 0.22.

[0047] The above calculation results imply that the more severe the congestion, the higher the temperature parameter should be adjusted, allowing the model to explore more detour options.

[0048] S25, based on the execution results of the historical task data, the parameter rule self-optimization mechanism is executed to correct the feature sensitivity weights, so as to achieve dynamic adjustment of the parameters of the large model. The specific steps are as follows: S251, Based on the historical task data, obtain the actual dispatch effect data after the vehicle dispatch plan is executed, and calculate the comprehensive effect score of the single dispatch task of the current task using a multi-dimensional weighted formula based on the actual dispatch effect data (such as non-violation rate, accuracy rate, vehicle occupancy rate, etc.). The formula for overall performance evaluation is: In the formula, This is a compliance indicator (binary variable: 0 or 1). If no physical restrictions (load capacity, traffic restrictions) are violated during actual implementation, the value is 1; otherwise, it is 0. The on-time delivery rate metric (continuous variable: 0.0-1.0) represents the proportion of orders delivered within the specified time window in the current task. The load factor index (continuous variable: 0.0-1.0) represents the ratio of the actual load of the dispatched vehicle to its rated load. The excess cost indicator (normalized value) is calculated as (actual mileage - theoretical minimum mileage) / theoretical minimum mileage. Preset business weighting coefficients (e.g.: This is used to balance the priorities of safety, timeliness, and cost.

[0049] S252, extract the average of the historical task comprehensive performance scores for each scenario as the historical baseline score for that scenario, and use the Delta Rule in reinforcement learning to update the current comprehensive performance score. Compared with the historical benchmark score of this scenario By comparing the results, a reward or penalty signal is generated, which in turn corrects the feature weights used in step S24 to calculate the dynamic adjustment coefficient. The formula for correcting feature sensitivity weights is: In the formula, The currently used feature sensitivity weights; The learning rate controls the step size for parameter adjustments (recommended value: 0.01-0.05) to prevent weight oscillations caused by a single instance of outlier data; if A positive difference indicates that the parameter adjustment has yielded better-than-expected results (positive reward). A negative difference indicates that the effect is worse than expected (negative reward); This represents the deviation between the feature vector of the i-th dimension in the current task and the feature vector of the standard scene.

[0050] Calculated The sensitivity weights of the modified i-th dimension features (such as the congestion sensitivity coefficient) are used for the next calculation of S24.

[0051] The self-optimization mechanism of the parameter rules establishes a logical connection between "dynamic adjustment coefficients" and "parameter self-optimization," encompassing both positive reinforcement and negative suppression. When the deviation of a certain feature vector (e.g.) This caused the system to increase the model temperature, and ultimately affected the vehicle dispatch score. When the value is above the baseline, the formula will increase the weight corresponding to that feature. This is the positive reinforcement mentioned above. If the adjustment leads to a decrease in the score (e.g., due to a violation or serious timeout), the formula will reduce the weight. This suppresses excessive parameter adjustment in similar scenarios in the future, thus converging the solution space, which is the negative suppression.

[0052] Table 1 Examples of Feature Vector-Parameter-Effect Mapping and Feedback for Business Scenarios S26, perform cluster analysis on the monthly new business data to generate new template vectors. The new business data includes new delivery routes, new vehicle models, etc.; or when the actual dispatch performance data for a certain scenario is lower than a preset threshold for one consecutive month (e.g., violation rate > 5%), trigger S25 to correct the feature sensitivity weights and update the scenario-based parameter template library. Table 2 shows a partial example of the scenario-based parameter template library.

[0053] Table 2, Examples of Some Contents in the Scenario-Based Parameter Template Library S3: Input the prompts generated in S1 and the dynamic parameters generated in S2 into the generation model to generate a vehicle dispatch plan and a delivery route planning plan. The specific steps are as follows: S31, Call the large language model API interface and input the complete prompt word constructed in S1. During this process, apply the final parameters calculated in step S24. (Temperature, Top-p) controls the inference process of the model: Input the structured prompts containing specific business data. The model uses a self-attention mechanism to analyze the spatial distribution of task points and the numerical relationship between vehicle load, searching the probability space for vehicle-task combinations that satisfy both "no overloading of individual vehicles" and "shortest total mileage". For example, a higher Temperature (e.g., 0.3) is used in "high congestion scenarios" to explore unconventional path combinations; a lower Temperature (e.g., 0.2) is used in "normal scenarios" to ensure output stability.

[0054] S32, Parse the output results. The output of the generated model is a structured text stream containing specific tags (such as vehicleGroups). The system extracts each data block through regular expressions and parses it into standard structured data objects (such as JSON or List objects) to form a preliminary vehicle dispatch and route planning scheme.

[0055] S4. Based on the complete prompt words, the detection model reviews and automatically corrects the result data of the generated model. The detection model here is not a simple rule code, but an LLM instance configured with a "code audit / logic verification" role through specific prompt words to handle implicit logical errors in unstructured semantics.

[0056] S41 uses the preliminary vehicle dispatch and route planning scheme output from S3 as the data to be checked. Combined with the constraints of the complete prompt words in S1, a validation prompt word is constructed. The prompt word focuses on "logical auditing". An example is shown below: "You are a meticulous data auditor. Please verify whether the following dispatch plan (enter A) violates the original vehicle restrictions (enter B). Verification steps:" 1. Calculate the total weight of the task assigned to each vehicle and determine if it exceeds the load limit.

[0057] 2. Check whether the output task ID set is completely consistent with the original task ID set (no omissions, no fictitious ones).

[0058] 3. If a violation is found, output {"status":"FAIL","error_code":"...","suggestion":"..."}; if fully compliant, output {"status":"PASS"}. S42, the detection model performs deterministic verification under extremely low temperature (Temperature=0.15) and low Top-p (0.5) parameters, reads the vehicle paths in vehicleGroups one by one, retrieves the weight and distance in the original task data, performs arithmetic summation again, and compares it with the original constraints of the prompt word template to ensure the uniqueness and rigor of the logical judgment.

[0059] S43, Output a structured feedback report: Case 1 (Pass): Output {"status":"PASS"}, and the process enters the S5 rule processing layer.

[0060] Case 2 (Failed): Output JSON containing the error location, for example: { "status":"FAIL", "error_code":"OVERWEIGHT", "target_vehicle":"899900", "current_load":5200, "limit_load":5000, "suggestion":"Remove Task-62332240" } S44. If the comparison result is fully compliant, a structured report with a status of "PASS" will be output, and the process will proceed to the next step, S5. S45, if the comparison result is non-compliant, a structured report with an error status (FAIL) is output. The structured report with an error status includes the error status, error code, target vehicle identifier, current load, load limit and correction suggestions, and triggers an automatic correction mechanism. S46, in the correction suggestion field of the parsing report, remove the non-compliant task (such as Task-62332240) from the corresponding target vehicle and mark it as "unassigned", and then reassign the task.

[0061] S5, verify the output results of the generation model and the detection model; If the output of the dual-model collaborative computing layer still contains data anomalies, a unified fallback check and compatibility processing are performed to improve the system's stability. As a preferred implementation, the data anomalies include: multiple vehicles assigned to a single task, invalid task ID, and other unexpected data (not output in the format required by the prompt words, incorrect data type).

[0062] S6, complete the other data of the preliminary vehicle dispatch and route planning scheme, and generate a complete intelligent vehicle dispatch and route planning scheme; In a preferred embodiment, the other data includes: vehicle name, license plate number, number of tasks, customer name, number of customers, load factor, total mileage, etc.

[0063] Example 2: This invention also provides an intelligent vehicle dispatching system based on large model dynamic parameters, such as... Figure 1 As shown, it includes: The business data assembly layer collects and standardizes the business data required for vehicle dispatch, forming standardized business data. Based on the standardized business data, it constructs a prompt word template and generates a complete prompt word. The large model parameter intelligent generation layer constructs a scenario-based parameter template library, collects real-time dynamic data of the current task, extracts historical task data, and automatically generates large model parameters based on the real-time dynamic data and the historical task data. The scenario-based parameter template library is updated using a parameter rule self-optimization mechanism. The dual-model collaborative computing layer includes a generation model and a detection model. The generation model generates a preliminary vehicle dispatch and route planning scheme based on the complete prompt words and the large model parameters. The detection model reviews and automatically corrects the preliminary vehicle dispatch and route planning scheme based on the large model parameters and the prompt word template. The rule processing layer is used to verify the output results of the dual-model collaborative computation layer; The output layer completes the data for the preliminary vehicle dispatch and route planning scheme, and generates and outputs the intelligent vehicle dispatch and route planning scheme.

[0064] The intelligent generation layer for large model parameters further includes: The key data extraction module is used to collect real-time dynamic data of the current task and extract historical task data, and extract dynamic scene data and task feature data from the real-time dynamic data to form key feature data for scene recognition. The feature extraction and scene recognition unit extracts the feature vector of the current task based on the key feature data and uses a scene recognition algorithm to match scene labels. The parameter generation rule engine calculates the parameters of the large model through the final parameter calculation formula, and performs a parameter rule self-optimization mechanism based on the execution results of the historical task data to correct the feature sensitivity weights. The parameter output verification module is used to calculate the final parameters based on the scene recognition results using the basic parameter generation formula, and correct the scene template feature sensitivity weights through a feedback correction mechanism. The scenario-based parameter template library module is used to build a scenario-based parameter template library and dynamically update the scenario-based parameter template library based on new business data or actual vehicle dispatch effect data.

[0065] For more detailed information on the working process of each of the above modules, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0066] As a preferred embodiment, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor; wherein, when the processor executes the computer program stored in the memory, it implements the aforementioned intelligent vehicle dispatching method based on large model dynamic parameters.

Claims

1. An intelligent vehicle dispatching method based on large model dynamic parameters, characterized in that, Includes the following steps: S1. Collect and standardize the business data required for vehicle dispatch to form standardized business data. Based on the standardized business data, construct a prompt word template and generate a complete prompt word. S2, construct a scenario-based parameter template library, collect real-time dynamic data of the current task, extract historical task data, automatically generate large model parameters based on the real-time dynamic data and the historical task data, dynamically adjust the large model parameters using a parameter rule self-optimization mechanism, and update the scenario-based parameter template library; S3. Based on the complete prompt words and the large model parameters, use the generative model to generate a preliminary vehicle dispatch and route planning scheme; S4. Based on the complete prompt words, the detection model is used to review and automatically correct the preliminary vehicle dispatch and route planning scheme; S5, verify the output results of the generation model and the detection model; S6, supplement other data of the preliminary vehicle dispatch and route planning scheme, and generate and output intelligent vehicle dispatch and route planning scheme.

2. The intelligent vehicle dispatching method based on large model dynamic parameters according to claim 1, characterized in that, S1 includes: S11. Collect and standardize the business data required for vehicle dispatch to form standardized business data, which includes warehouse data, truck data set, task data set and route data; S12, construct a prompt word template based on the standardized business data, fill the standardized business data into the prompt word template, and generate a complete prompt word containing constraints.

3. The intelligent vehicle dispatching method based on large model dynamic parameters according to claim 1, characterized in that, S2 includes: S21, construct a scenario-based parameter template library, which includes at least scenario labels and corresponding large model parameter baseline values, trigger thresholds, standard scenario feature vectors, feature sensitivity weights, and business weight coefficients for each scenario label. S22, collect real-time dynamic data of the current task, extract historical task data, extract dynamic scene data and task feature data from the real-time dynamic data, and form key feature data for scene recognition; S23, a scene recognition algorithm is used to identify the scene of the current task. The scene recognition algorithm adopts a dual mechanism of threshold determination and similarity matching. S24. Based on the scene recognition results, the basic parameter generation formula is used to calculate the final parameters and output them. The final parameter calculation formula is as follows: In the formula, The parameter baseline value is obtained from the scenario-based parameter template library. Let i be the feature sensitivity weight of the i-th dimension. The deviation between the feature vector of the i-th dimension in the current task and the feature vector of the standard scene; S25, based on the execution results of the historical task data, execute the parameter rule self-optimization mechanism to correct the feature sensitivity weight; S26, perform cluster analysis on the monthly new business data to generate new template vectors, or when the actual vehicle dispatch effect data of a certain scenario is lower than the preset threshold for one consecutive month, trigger S25 to correct the feature sensitivity weight and update the scenario-based parameter template library.

4. The intelligent vehicle dispatching method based on large model dynamic parameters according to claim 3, characterized in that, S23 includes: S231, Set a trigger threshold determination function for each scenario, substitute the real-time dynamic data into the trigger threshold determination function to determine the threshold, filter out candidate scenarios that meet the threshold conditions, and form a candidate scenario set. S232, Match scene tags according to the candidate scene set. If the candidate scene set contains only one candidate scene, then use its corresponding tag as the matching scene tag for the current task and call the corresponding parameter baseline value. If the candidate scene set is an empty set or contains multiple conflicting scene tags, then perform similarity matching. S233, based on the key feature data, obtain a 7-dimensional feature vector, call the standard scene feature vector in the scene-based parameter template library, calculate the cosine similarity between the 7-dimensional feature vector and the standard scene feature vector, select the scene label corresponding to the scene with the highest feature vector similarity as the matching scene label for the current task, and call the parameter benchmark value corresponding to the scene label. The cosine similarity calculation formula is: In the formula, A represents the 7-dimensional feature vector extracted from the real-time dynamic data based on the current task data, and B represents the standard scene feature vector in the scene-based parameter template library. It is the dot product of two vectors. It is the product of the magnitudes of two vectors.

5. The intelligent vehicle dispatching method based on large model dynamic parameters according to claim 3, characterized in that, S25 includes: S251, Based on the historical task data, obtain the actual dispatch effect data after the vehicle dispatch plan is executed, and calculate the comprehensive effect score of the current task based on the actual dispatch effect data. The comprehensive effect score formula is: In the formula, For compliance indicators, As an indicator of on-time performance, The full load rate is an indicator. For excess cost indicators, These are preset business weighting coefficients; S252, extract the average of the historical task comprehensive performance scores for each scenario as the historical baseline score for that scenario. Using the incremental update rule in reinforcement learning, compare the current task's comprehensive performance score with the historical baseline score for that scenario to calculate a new feature sensitivity weight, which is used for the next parameter calculation. The feature sensitivity weight correction formula is as follows: In the formula, The currently used feature sensitivity weights, For learning rate, To score the overall effect of this test, Score based on historical benchmarks. As a reward signal, This represents the deviation between the feature vector of the i-th dimension in the current task and the feature vector of the standard scene.

6. The intelligent vehicle dispatching method based on large model dynamic parameters according to claim 1, characterized in that, S3 includes: S31, Call the large language model API interface, input the complete prompt word constructed in S1, the generation model controls the inference process of the model based on the final parameters calculated in step S24, and outputs the result; S32 extracts the data blocks of the output results using regular expressions, parses them into standard structured data objects, and forms a preliminary vehicle dispatching and route planning scheme.

7. The intelligent vehicle dispatching method based on large model dynamic parameters according to claim 1, characterized in that, S4 includes: S41, the preliminary vehicle dispatch and route planning scheme is used as the data to be checked, and the constraint conditions of the complete prompt word are combined to construct a verification prompt word; S42, the detection model operates under low temperature parameters and low Top-p parameters, and compares the preliminary vehicle dispatch and route planning scheme with the constraints. S43, the detection model outputs a structured feedback report based on the comparison results; S44. If the comparison result is fully compliant, a structured report with a status of "passed" will be output, and the process will proceed to the next step, S5. S45, if the comparison result is non-compliant, a structured report with an error status is output. The structured report with an error status includes the error status, error code, target vehicle identifier, current load, load limit and correction suggestions, and triggers an automatic correction mechanism. S46, in the correction suggestion field of the parsing report, remove the non-compliant task from the corresponding target vehicle and reassign the task.

8. An intelligent vehicle dispatching system based on large model dynamic parameters, comprising: The business data assembly layer is used to collect and standardize the business data required for vehicle dispatch, forming standardized business data. Based on the standardized business data, a prompt word template is constructed to generate a complete prompt word. The large model parameter intelligent generation layer is used to build a scenario-based parameter template library, collect real-time dynamic data of the current task, extract historical task data, automatically generate large model parameters based on the real-time dynamic data and the historical task data, and update the scenario-based parameter template library using a parameter rule self-optimization mechanism. The dual-model collaborative computing layer includes a generation model and a detection model. The generation model generates a preliminary vehicle dispatch and route planning scheme based on the complete prompt words and the large model parameters. The detection model reviews and automatically corrects the preliminary vehicle dispatch and route planning scheme based on the complete prompt words. The rule processing layer is used to verify the output results of the dual-model collaborative computation layer; The results output layer is used to complete the data of the preliminary vehicle dispatch and route planning scheme, and generate and output the intelligent vehicle dispatch and route planning scheme.

9. The intelligent vehicle dispatching system based on large model dynamic parameters according to claim 8, characterized in that, The intelligent generation layer for large model parameters includes: The key data extraction module is used to collect real-time dynamic data of the current task and extract historical task data, and extract dynamic scene data and task feature data from the real-time dynamic data to form key feature data for scene recognition. The feature extraction and scene recognition unit extracts the feature vector of the current task based on the key feature data and uses a scene recognition algorithm to match scene labels. The parameter generation rule engine is used to calculate the parameters of the large model based on the scene recognition results and the final parameter calculation formula. It also uses the execution results of the historical task data to execute the parameter rule self-optimization mechanism and correct the feature sensitivity weights. The parameter output verification module is used to calculate the final parameters based on the scene recognition results using the basic parameter generation formula, and correct the scene template feature sensitivity weights through a feedback correction mechanism. The scenario-based parameter template library module is used to build a scenario-based parameter template library and update the scenario-based parameter template library based on new business data or actual vehicle dispatch effect data.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent vehicle dispatching method based on large model dynamic parameters as described in any one of claims 1 to 7.

Images