A vehicle scheduling method of a network freight platform combined with a multi-type transport capacity pool
By constructing multi-type capacity pools and a multi-modal vehicle vector scheduling method, the problems of hierarchical organization of capacity pools and inaccurate route evaluation in vehicle scheduling of online freight platforms are solved, achieving high-accuracy and low-latency vehicle matching and scheduling, and improving the system's response efficiency and data accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Filing Date
- 2026-04-29
- Publication Date
- 2026-07-14
Smart Images
Figure CN122390599A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent logistics and transportation scheduling technology, specifically to a vehicle scheduling method and system based on trajectory data mining, temporal feature encoding and vector space retrieval, applicable to collaborative scenarios of order management system (OMS) and transportation management system (TMS) of online freight platforms. Background Technology
[0002] One of the core tasks of online freight platforms is to achieve efficient matching of waybills and transport capacity. Current mainstream vehicle scheduling methods suffer from the following technical limitations: the capacity pool often uses a single public pool, lacking hierarchical organizational logic and making it difficult to differentiate the characteristics of different types of capacity; the evaluation of route capacity often relies on static historical frequency statistics, failing to effectively reflect the business pattern of "recent activity being higher than long-term activity"; the determination of the permanent location is easily affected by temporary locations such as highway service areas and gas stations, leading to inaccurate geographical features; and the matching and scheduling methods employ multi-round rule filtering and weighted scoring, resulting in high response latency with millions of candidate capacity slots, making it difficult to balance recall rate and retrieval efficiency. Therefore, how to construct a vehicle scheduling method that balances high accuracy and low latency is a pressing technical problem to be solved in this field. Summary of the Invention
[0003] The technical problem to be solved by this invention is to provide a vehicle scheduling method that balances recall accuracy and retrieval efficiency in high-concurrency scenarios of online freight platforms. Specifically, this includes addressing issues such as heterogeneous data fusion, temporal feature modeling, geographic determination robustness, retrieval efficiency, and model interpretability.
[0004] To address the aforementioned technical problems, this invention provides a vehicle scheduling method for a network freight platform that combines multiple types of transportation capacity pools, comprising the following steps: S1. Construct multiple types of capacity pools, including: Based on vehicle trajectory data, a density clustering algorithm is used to identify long-term stopping points, and temporary locations are filtered out by point of interest (POI) type to construct an urban transportation capacity pool. The route capacity pool is constructed based on a 16-level time priority rule. The time level is divided into one level every 10 days based on the current date, and there are 16 levels in total, going back 160 days. The route priority is calculated using an exponential decay weight. Based on the vehicle master data and qualification information uploaded by enterprises, a private transportation capacity pool for enterprises is constructed. S2. Create vehicle multimodal vectors: Create private capacity pool sub-vectors, route capacity pool vectors, city capacity pool vectors, and real-time status sub-vectors respectively. Concatenate the four sub-vectors and normalize them to obtain a 256-dimensional vehicle vector. S3. Waybill-Vehicle Matching and Scheduling: Receive waybill requests and generate waybill vectors; perform initial screening of candidate vehicles based on hard constraint masks; in the initial screening results, use an approximate nearest neighbor ANN to retrieve the Top K vehicle vectors most similar to the waybill vectors; perform real-time location distance correction on the recall results, and output the scheduling results in order of the corrected similarity scores.
[0005] Furthermore, the step S1 of filtering out temporary locations by POI type specifically includes: The identified long-term stop points are matched with POI data for spatial proximity. If the matched POI type belongs to highway service area, gas station or toll station, the long-term stop point is removed. The city with the longest cumulative stay among the remaining long-term stops will be identified as the vehicle's permanent city and included in the corresponding city's transportation capacity pool.
[0006] Furthermore, the formula for calculating the exponential decay weight in step S1 is as follows:
[0007] Where i is the time level number, i=1 represents the most recent 10 days, i=16 represents days 151-160, and β is the decay coefficient, with a value range of 0.05 to 0.3.
[0008] Furthermore, the line priority in step S1 is specifically calculated in the following manner: Statistically count the frequency of vehicle departures from a vehicle to a specified route within the i-th time window. Calculate the cumulative priority total score And calculate the normalization priority. ; When the vehicle When the vehicle is in use, it will be included in the capacity pool of the corresponding route.
[0009] Furthermore, the line capacity pool sub-vector in step S2 is created in the following way: Extract the departure frequency sequence of vehicles on a certain route across 16 time levels. The sequence is input into a gated recurrent unit (GRU) for encoding. The hidden state of the last time step is taken as the line timing feature, and then mapped into a 64-dimensional sub-vector through a fully connected layer.
[0010] Further, in step S2, the private capacity pool sub-vector, the city capacity pool sub-vector, and the real-time state sub-vector are obtained through corresponding feature extraction and neural network mapping, respectively, wherein: The private capacity pool vector is generated based on the cooperative relationship features between vehicles and shipping companies; The city capacity pool vector is generated based on the vehicle's permanent city geocoding, the coordinates of core long-term parking points, and the distance features from the waybill's origin. The real-time state sub-vector is generated based on the vehicle's real-time location, idle status, and distance to the waybill loading location.
[0011] Furthermore, the hard constraint mask in step S3 is implemented using bitmap indexing, and the constraint conditions include at least: vehicle type matching, origin matching (coverage of city capacity pool or route capacity pool), vehicle idle status, and qualification compliance; by performing a bitwise AND operation on the bitmap of each constraint condition, a set of candidate vehicles is obtained.
[0012] Furthermore, the calculation formula for the distance correction in step S3 is as follows:
[0013] in, Let cosine similarity be the vector of the single shipment and the vector of the vehicle. This represents the spherical distance between the vehicle's current location and the origin of the waybill. This is the preset maximum matching distance.
[0014] This invention also provides a network freight platform vehicle dispatching system that integrates multiple types of transport capacity pools, comprising: A capacity pool construction module is used to perform the capacity pool construction operation described in step S1 of claim 1. A vector generation module is used to perform the vehicle vector creation operation described in step S2 of claim 1; The matching and scheduling module is used to perform the waybill-vehicle matching and scheduling operation described in step S3 of claim 1.
[0015] Furthermore, the vector generation module further includes: Private sub-vector unit, used to generate private capacity pool sub-vectors; Line sub-vector unit, used to encode 16-level time-frequency sequences using GRU to generate line capacity pool sub-vectors; Urban sub-vector units are used to generate urban capacity pool sub-vectors based on geographic raster encoding and embedding. Real-time sub-vector unit, used to generate real-time state sub-vectors based on real-time state features; The fusion unit is used to concatenate and normalize the four sub-vectors to obtain a 256-dimensional vehicle vector.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention achieves structured organization of heterogeneous data through the standardized construction of a three-layer capacity pool, increasing the efficiency of the line capacity pool to 92.3% and the data accuracy of the city capacity pool to 91.5%. By using 16-level time windows, exponential decay weights, and GRU time-series coding, the time priority of line capacity is accurately characterized, improving the matching accuracy by 17 percentage points.
[0017] 2. This invention reduces the scheduling response latency P95 from 1.2s to 0.15s, a reduction of 87.5%, by combining hard constraint masks with ANN for retrieval. The POI filtering mechanism eliminates interference from temporary locations, significantly improving the robustness of geographical determination. GRU-gated weight visualization and similarity decomposition ensure the model's interpretability. Attached Figure Description
[0018] Figure 1 This is a system architecture diagram of an embodiment of the present invention; Figure 2 This is a flowchart illustrating the multi-type capacity pool update process according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the vehicle vector construction process according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating the combined retrieval and sorting process according to an embodiment of the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0020] The core of this invention lies in resolving the technical contradiction in high-concurrency scenarios for online freight platforms, where vehicle dispatching must be both "accurate" (high matching accuracy) and "fast" (low response latency). It is not an improvement on a single algorithm, but rather an end-to-end, systematic engineering optimization solution, implemented through a three-pronged architecture of "layering, encoding, and retrieval."
[0021] Therefore, this embodiment provides a vehicle scheduling method for a network freight platform that combines multiple types of transport capacity pools, including the following steps: S1. Construct multiple types of capacity pools, including: Based on vehicle trajectory data, a density clustering algorithm is used to identify long-term stopping points, and temporary locations are filtered out by point of interest (POI) type to construct an urban transportation capacity pool. The route capacity pool is constructed based on a 16-level time priority rule. The time level is divided into one level every 10 days based on the current date, and there are 16 levels in total, going back 160 days. The route priority is calculated using an exponential decay weight. Based on the vehicle master data and qualification information uploaded by enterprises, a private transportation capacity pool for enterprises is constructed.
[0022] The step S1 of filtering out temporary locations by POI type specifically includes: The identified long-term stop points are matched with POI data for spatial proximity. If the matched POI type belongs to highway service area, gas station or toll station, the long-term stop point is removed. The city with the longest cumulative stay among the remaining long-term stops will be identified as the vehicle's permanent city and included in the corresponding city's transportation capacity pool.
[0023] Furthermore, the formula for calculating the exponential decay weight in step S1 is as follows:
[0024] Where i is the time level number, i=1 represents the most recent 10 days, i=16 represents days 151-160, and β is the decay coefficient, with a value range of 0.05 to 0.3.
[0025] The line priority in step S1 is calculated in the following manner: Statistically count the frequency of vehicle departures from a vehicle to a specified route within the i-th time window. Calculate the cumulative priority total score And calculate the normalization priority. ; When the vehicle When the vehicle is in use, it will be included in the capacity pool of the corresponding route.
[0026] S2. Create vehicle multimodal vectors: Create private capacity pool sub-vectors, route capacity pool vectors, city capacity pool vectors, and real-time status sub-vectors respectively. Concatenate the four sub-vectors and normalize them to obtain a 256-dimensional vehicle vector.
[0027] The route capacity pool sub-vector in step S2 is created in the following way: Extract the departure frequency sequence of vehicles on a certain route across 16 time levels. The sequence is input into a gated recurrent unit (GRU) for encoding. The hidden state of the last time step is taken as the line timing feature, and then mapped into a 64-dimensional sub-vector through a fully connected layer.
[0028] In step S2, the private capacity pool sub-vector, the city capacity pool sub-vector, and the real-time status sub-vector are obtained through corresponding feature extraction and neural network mapping, respectively, where: The private capacity pool vector is generated based on the cooperative relationship features between vehicles and shipping companies; The city capacity pool vector is generated based on the vehicle's permanent city geocoding, the coordinates of core long-term parking points, and the distance features from the waybill's origin. The real-time state sub-vector is generated based on the vehicle's real-time location, idle status, and distance to the waybill loading location.
[0029] S3. Waybill-Vehicle Matching and Scheduling: Receive waybill requests and generate waybill vectors; perform initial screening of candidate vehicles based on hard constraint masks; in the initial screening results, use an approximate nearest neighbor ANN to retrieve the Top K vehicle vectors most similar to the waybill vectors; perform real-time location distance correction on the recall results, and output the scheduling results in order of the corrected similarity scores.
[0030] The hard constraint mask in step S3 is implemented using bitmap indexing. The constraints include at least: vehicle type matching, origin matching (coverage of city capacity pool or route capacity pool), vehicle idle status, and qualification compliance. A set of candidate vehicles is obtained by performing a bitwise AND operation on the bitmap of each constraint.
[0031] The formula for calculating the distance correction in step S3 is as follows:
[0032] in, Let cosine similarity be the vector of the single shipment and the vector of the vehicle. This represents the spherical distance between the vehicle's current location and the origin of the waybill. This is the preset maximum matching distance.
[0033] Example 1: Scheduling scenario of a manufacturing company in Hefei This embodiment uses an actual scheduling scenario of a manufacturing enterprise in Hefei as an example to illustrate the specific implementation process of the present invention.
[0034] Implementation Environment: - Platform-connected operational vehicles: 520,000 - Hardware Environment: 32-core CPU, 128GB RAM, NVIDIA T4 GPU (for training) - Vector Library: Milvus 2.3.0, HNSW index - Index Parameters: , ,
[0035] Implementation steps: Step 1: Constructing the Capacity Pool Private Capacity Pool: A manufacturing company in Hefei (Company ID: HF2024001) submitted the vehicle master data and qualification fields for 8 4.2-meter vans; - After passing the consistency and validity period verification, they were included in the company's private capacity pool upon initial binding; - A private capacity pool tag was generated: {Company ID: HF2024001, Vehicle ID: [V001-V008], Compliance Status: 1, Vehicle Code: 42XL, Binding Timestamp: 2024-01-15T09:23:17Z, Historical Cooperation Count: 0, Recent Cooperation Time: null}.
[0036] Urban transportation capacity pool: For the Anhui AXXXX heavy semi-trailer tractor (vehicle ID: V10086), collect 3 months of trajectory data (sampling period 30s, total 86,400 trajectory points); - Identify 15 candidate long-stop points (DBSCAN clustering). , - Filter out 10 temporary locations including highway service areas (3), gas stations (5), and toll stations (2); - Among the remaining 5 valid long-term stops, the core location with the longest cumulative stay is located in Yaohai District, Hefei City (117.32°E, 31.87°N), with a cumulative stay duration of 312 hours; - Determine the permanent city as Hefei (city code: 340100) and include it in the Hefei City Transportation Capacity Pool; - Generate the city transportation capacity pool label: {Permanent City Code: 340100, Core Long-Term Stop Geohash: wtfedr, Cumulative Stay Duration: 312h, Vehicle Type Code: 42XL, Determination Timestamp: 2024-01-15T10:00:00Z}.
[0037] Route capacity pool construction: For vehicle V10086, backtrack 160 days of waybill data, dividing it into 16 time levels based on 10-day increments; - Statistics on the departure frequency of the Hefei-Nanjing route (origin 340100 - destination 320100): [5,4,3,2,1,0,0,1,2,3,2,1,0,0,0,0] (Level 1 to Level 16); - Calculation of exponential decay weights ( ): - Calculate the cumulative priority score: - Normalization priority: - This line is included in the threshold. (quantiles) Since 0.201 ≥ 0.15, it is included in the Hefei-Nanjing route capacity pool; - Generate route capacity pool tags: {Route Code: 340100-320100, Cumulative Priority Score: 4.82, Normalized Priority: 0.201, Last Departure Time: 2024-01-10T14:30:00Z, 16-level Time Distribution Summary: [5,4,3,2,1,0,0,1,2,3,2,1,0,0,0,0], Update Cycle Identifier: 20240115}.
[0038] Step 2: Vehicle Vector Creation Private capacity pool vector : 2.3 times the module length of vehicles not belonging to this company; Line capacity pool vector The cosine similarity of the single vector with the Hefei-Nanjing route is 0.92. City capacity pool vector The cosine similarity between the single vector and the originating point Hefei is 0.88. Real-time state subvector The current location is 12km away from the loading site in Yaohai District, Hefei, and is currently unavailable. After concatenation and normalization, a 256-dimensional vehicle vector is generated. Write it into the Milvus vector library.
[0039] Step 3: Waybill Matching and Scheduling The platform receives a Hefei-Nanjing waybill submitted by Hefei manufacturing company HF2024001; - Waybill characteristics: Origin 340100, Destination 320100, Vehicle type 42XL, Loading time 2024-01-16T08:00:00Z, Cargo type: General cargo; - Generates a waybill vector. ; Hard constraint filtering: Vehicle type = 42XL, origin matching (city pool 340100 or route pool 340100-320100), status = idle, qualifications include general cargo transportation; - Hard constraint candidate set: 12,500 vehicles → 3,200 vehicles; ANN search: Top 50 similar vehicles recalled, with vehicles V001-V003 from company HF2024001's private capacity pool ranking in the top 3; - Distance correction: V10086's current location is 12km from the loading site. Correction factor - Final Score: ; Top 5 results: V001 (0.89), V002 (0.85), V003 (0.82), V10086 (0.70), V2056 (0.68); - Pushed to the shipper's end, confirmed transaction V001, the whole process took 12ms.
[0040] Example 2: Large-scale concurrent stress test Test conditions: 500 concurrent requests per second (QPS), 1.2 million candidate vehicles, and a continuous test duration of 72 hours.
[0041] Test results: The scheduling response latency P95 was 142ms (target <200ms), P99 was 298ms (target <500ms), the system availability was 99.97%, the Top 5 matching accuracy was 81.3%, and the vector retrieval recall rate was 97.2%. All indicators met the standards.
[0042] Example 3: Ablation Experiment
[0043] Ablation experiments show that GRU timing coding contributes 8 percentage points to the accuracy improvement, POI filtering contributes 6 percentage points, and hard constraint mask reduces latency by 72%.
[0044] Recommended values for key parameters
[0045] AI Ethics and Compliance Statement This invention strictly adheres to data ethics guidelines during its implementation: Legality of data sources: Vehicle trajectory data, waybill transaction data, and enterprise qualification data are all authorized through user agreements or authorization letters, with clear authorization scope and withdrawal mechanisms.
[0046] Data anonymization: Vehicle identifiers are generated using SHA-256 hashing to create internal IDs; enterprise identifiers are encrypted using AES; random noise (±10m) is added to trajectory coordinates; timestamps are rounded down to the 5-minute level.
[0047] Interpretability: Provides a GRU-gated weighted heatmap, supports similarity decomposition into four sub-vector contributions, records the complete decision path, and meets audit requirements.
[0048] Data security: The vector library implements RBAC access control, and audit logs are retained for 3 years. This invention also provides a network freight platform vehicle dispatching system that integrates multiple types of transport capacity pools, comprising: A capacity pool construction module is used to perform the capacity pool construction operation described in step S1 of claim 1. A vector generation module is used to perform the vehicle vector creation operation described in step S2 of claim 1; The matching and scheduling module is used to perform the waybill-vehicle matching and scheduling operation described in step S3 of claim 1.
[0049] Furthermore, the vector generation module further includes: Private sub-vector unit, used to generate private capacity pool sub-vectors; Line sub-vector unit, used to encode 16-level time-frequency sequences using GRU to generate line capacity pool sub-vectors; Urban sub-vector units are used to generate urban capacity pool sub-vectors based on geographic raster encoding and embedding. Real-time sub-vector unit, used to generate real-time state sub-vectors based on real-time state features; The fusion unit is used to concatenate and normalize the four sub-vectors to obtain a 256-dimensional vehicle vector.
[0050] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.
Claims
1. A vehicle scheduling method for a network freight platform that combines multiple types of transport capacity pools, characterized in that, Includes the following steps: S1. Construct multiple types of capacity pools, including: Based on vehicle trajectory data, a density clustering algorithm is used to identify long-term stopping points, and temporary locations are filtered out by point of interest (POI) type to construct an urban transportation capacity pool. The route capacity pool is constructed based on a 16-level time priority rule. The time level is divided into one level every 10 days based on the current date, and there are 16 levels in total, going back 160 days. The route priority is calculated using an exponential decay weight. Based on the vehicle master data and qualification information uploaded by enterprises, a private transportation capacity pool for enterprises is constructed. S2. Create vehicle multimodal vectors: Create private capacity pool sub-vectors, route capacity pool vectors, city capacity pool vectors, and real-time status sub-vectors respectively. Concatenate the four sub-vectors and normalize them to obtain a 256-dimensional vehicle vector. S3. Waybill-Vehicle Matching and Scheduling: Receive waybill requests and generate waybill vectors; perform initial screening of candidate vehicles based on hard constraint masks; in the initial screening results, use an approximate nearest neighbor ANN to retrieve the Top K vehicle vectors most similar to the waybill vectors; perform real-time location distance correction on the recall results, and output the scheduling results in order of the corrected similarity scores.
2. The vehicle scheduling method for a network freight platform combining multiple types of transport capacity pools according to claim 1, characterized in that, The step S1 of filtering out temporary locations by POI type specifically includes: The identified long-term stop points are matched with POI data for spatial proximity. If the matched POI type belongs to highway service area, gas station or toll station, the long-term stop point is removed. The city with the longest cumulative stay among the remaining long-term stops will be identified as the vehicle's permanent city and included in the corresponding city's transportation capacity pool.
3. The vehicle scheduling method for a network freight platform combining multiple types of transport capacity pools according to claim 1, characterized in that, The formula for calculating the exponential decay weight in step S1 is as follows: Where i is the time level number, i=1 represents the most recent 10 days, i=16 represents days 151-160, and β is the decay coefficient, with a value range of 0.05 to 0.
3.
4. The vehicle scheduling method for a network freight platform combining multiple types of transport capacity pools according to claim 1, characterized in that, The line priority in step S1 is calculated in the following manner: Statistically count the frequency of vehicle departures from a vehicle to a specified route within the i-th time window. Calculate the cumulative priority total score And calculate the normalization priority. ; When the vehicle When the vehicle is in use, it will be included in the capacity pool of the corresponding route.
5. The vehicle scheduling method for a network freight platform combining multiple types of transport capacity pools according to claim 1, characterized in that, The route capacity pool sub-vector in step S2 is created in the following way: Extract the departure frequency sequence of vehicles on a certain route across 16 time levels. The sequence is input into a gated recurrent unit (GRU) for encoding. The hidden state of the last time step is taken as the line timing feature, and then mapped into a 64-dimensional sub-vector through a fully connected layer.
6. The vehicle scheduling method for a network freight platform combining multiple types of transport capacity pools according to claim 1, characterized in that, In step S2, the private capacity pool sub-vector, the city capacity pool sub-vector, and the real-time status sub-vector are obtained through corresponding feature extraction and neural network mapping, respectively, where: The private capacity pool vector is generated based on the cooperative relationship features between vehicles and shipping companies; The city capacity pool vector is generated based on the vehicle's permanent city geocoding, the coordinates of core long-term parking points, and the distance features from the waybill's origin. The real-time state sub-vector is generated based on the vehicle's real-time location, idle status, and distance to the waybill loading location.
7. The vehicle scheduling method for a network freight platform combining multiple types of transport capacity pools according to claim 6, characterized in that, The hard constraint mask in step S3 is implemented using bitmap indexing. The constraints include at least: vehicle model matching, origin matching, vehicle idle status, and qualification compliance. A set of candidate vehicles is obtained by performing a bitwise AND operation on the bitmap of each constraint.
8. The vehicle scheduling method for a network freight platform combining multiple types of transport capacity pools according to claim 7, characterized in that, The formula for calculating the distance correction in step S3 is as follows: in, Let cosine similarity be the vector of the single shipment and the vector of the vehicle. This represents the spherical distance between the vehicle's current location and the origin of the waybill. This is the preset maximum matching distance.
9. A vehicle dispatching system for a network freight platform that integrates multiple types of transport capacity pools, characterized in that, include: A capacity pool construction module is used to perform the capacity pool construction operation described in step S1 of claim 1. A vector generation module is used to perform the vehicle vector creation operation described in step S2 of claim 1; The matching and scheduling module is used to perform the waybill-vehicle matching and scheduling operation described in step S3 of claim 1.
10. The system according to claim 9, characterized in that, The vector generation module further includes: Private sub-vector unit, used to generate private capacity pool sub-vectors; Line sub-vector unit, used to encode 16-level time-frequency sequences using GRU to generate line capacity pool sub-vectors; Urban sub-vector units are used to generate urban capacity pool sub-vectors based on geographic raster encoding and embedding. Real-time sub-vector unit, used to generate real-time state sub-vectors based on real-time state features; The fusion unit is used to concatenate and normalize the four sub-vectors to obtain a 256-dimensional vehicle vector.