A logistics return trip transport capacity intelligent matching system and method based on trip prediction

Abstract

A logistics return trip capacity intelligent matching system and method based on travel prediction, including, for each vehicle current collection period of multi-source heterogeneous data, capacity departure time window prediction, travel time estimation, space correction estimation, sudden factor correction estimation and bearable weight prediction, obtaining the capacity departure time window, the optimal return path, the return time and the estimated bearable weight of the target vehicle; based on the capacity departure time window, the optimal return path, the return time and the estimated bearable weight of the target vehicle, generate the capacity label, adapt the capacity label with the order information in multiple dimensions, obtain the comprehensive score of the capacity label, based on the comprehensive score of the capacity label, the order matching and state updating operation of the target vehicle, realize the accurate prediction of the capacity state, greatly reduce the invalid matching rate.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics scheduling technology, specifically a logistics return capacity intelligent matching system and method based on trip prediction. Background Technology

[0002] In the logistics industry, the issue of empty return trips for trucks after completing their outbound transportation tasks has long been a persistent pain point. Traditional logistics return trip capacity matching often relies on manual scheduling or a simple "cargo finds truck, truck finds cargo" model. Existing technologies generally suffer from core technical pain points in the logistics return trip capacity matching process, including vague understanding of capacity status, crude matching logic, and a lack of end-to-end control. Specifically, it is impossible to accurately predict the vehicle's return departure time window, actual carrying capacity, and route time. It relies heavily on manual experience or single-dimensional data for capacity screening, resulting in order and capacity matching only meeting the basic load and route adaptation needs. It is difficult to take into account multiple dimensions such as timeliness, cost, and safety. This not only easily leads to ineffective matching situations where "cargo waits for truck" and "truck waits for cargo," but also causes transportation risks due to insufficient consideration of unforeseen factors such as uphill sections and extreme weather. At the same time, the lack of a dynamic capacity status update mechanism restricts the improvement of overall logistics transportation efficiency and economic benefits. Summary of the Invention

[0003] To address the aforementioned technical problems, the present invention aims to provide an intelligent matching method for return freight capacity based on trip prediction, comprising the following steps:

[0004] Step s1: For each vehicle, perform capacity departure time window prediction, trip time estimation, spatial correction estimation, emergency factor correction estimation, and load-bearing weight prediction on the multi-source heterogeneous data of the current collection period to obtain the target vehicle's capacity departure time window, optimal return route, return time, and estimated load-bearing weight.

[0005] Step s2: Generate capacity tags based on the target vehicle's capacity departure time window, optimal return route, return time, and estimated carrying capacity. Adapt the capacity tags to order information in multiple dimensions to obtain a comprehensive score for the capacity tags. Based on the comprehensive score of the capacity tags, perform order matching and status update operations for the target vehicle.

[0006] Furthermore, a cloud-based monitoring platform is constructed, and data acquisition terminals are deployed on vehicles and within a preset range. The cloud-based monitoring platform communicates with each data acquisition terminal. The data acquisition terminals are used to transmit the collected real-time dynamic data to the cloud-based monitoring platform and mark the collection period. The cloud-based monitoring platform is equipped with a vehicle database, an order pool, and a capacity candidate pool. The vehicle database is used to store multi-source heterogeneous data for the current period and several historical collection periods. The multi-source heterogeneous data includes static basic data and real-time dynamic data.

[0007] Furthermore, the process of forecasting capacity departure time windows includes:

[0008] Extract the operating status of each vehicle from the multi-source heterogeneous data of the current collection period, mark the vehicles in the unloading state as target vehicles, and obtain the cargo source characteristics, remaining cargo weight, and the queuing vehicle type and the number of each type of queuing vehicle at the unloading point of the target vehicle.

[0009] Key feature analysis is performed on the multi-source heterogeneous data of the target vehicle within the historical collection period to obtain the historical average loading and unloading efficiency of the target vehicle for the current cargo characteristics.

[0010] Key feature analysis was performed on multi-source heterogeneous data of several vehicles in the vehicle database during the historical collection period to obtain the historical baseline duration of different vehicle types under the current cargo source characteristics.

[0011] The baseline unloading time of the target vehicle is obtained based on the remaining cargo weight of the target vehicle and the historical average loading and unloading efficiency. The queuing waiting time is obtained based on the vehicle type in the queue at the unloading point, the number of each type of vehicle type in the queue, and the historical baseline time.

[0012] The total unloading operation time is obtained based on the baseline unloading time and queuing waiting time. The historical dwelling behavior characteristics of the target vehicle are obtained. The capacity departure time window of the target vehicle is obtained based on the total unloading operation time, historical dwelling behavior characteristics and the current timestamp.

[0013] Furthermore, the process of estimating travel time and making spatial corrections includes:

[0014] Obtain the unloading point and return destination of the target vehicle, construct a road transport network containing several return routes based on the unloading point and return destination, define the node types in the road transport network and the road segments connecting adjacent nodes, and add road segment length, real-time traffic speed and congestion index to the road segments connecting adjacent nodes based on the real-time dynamic data of the current collection period.

[0015] The travel time data of the target vehicle in several historical collection periods is extracted from the vehicle database. Core time-series features and time-derived features are extracted from the travel time data. The core time-series features and time-derived features are marked as training data to construct a travel time prediction model. The travel time prediction model is trained using the training data to obtain a completed travel time prediction model.

[0016] Input the return route between the unloading point and the return destination in the road transport network into the travel time prediction model, and output the basic travel time prediction value of each return route based on the travel time prediction model.

[0017] The topology of several return routes in the road transport network is represented in graph form. The graph attention network is used to learn the graph representation of the topology. The training data, as well as the real-time traffic speed and congestion index of each road segment, are imported into the graph attention network. The spatial correction time of each return route is output through the graph attention network after the learning is completed.

[0018] Furthermore, the process of revising the forecast for unforeseen factors includes:

[0019] Extract travel time data of target vehicles containing sudden factors within several historical collection periods from the vehicle database as sample data. Perform travel time prediction and spatial correction prediction on each sample data to obtain the basic travel time prediction value and the spatial correction travel time. Compare the sum of the basic travel time prediction value and the spatial correction travel time with the actual travel time in the sample data to obtain the travel time difference value. Bind the travel time difference value to the label value of the sample data.

[0020] Construct a sudden factor prediction model, train the sudden factor prediction model using sample data with completed label value binding, obtain the completed sudden factor prediction model, extract the sudden factor features contained in each return path from the real-time dynamic data of the target vehicle in the current collection period, input the sudden factor features into the sudden factor prediction model, and output the time correction amount of each return path according to the sudden factor prediction model.

[0021] The return time of each return path is obtained by spatially correcting the return time and the time correction amount based on the basic time prediction value of each return path, and the return path with the shortest return time is selected and marked as the optimal return path.

[0022] Furthermore, the process of predicting the load-bearing capacity includes:

[0023] Geographic information data, environmental interference data, and real-time status data of the target vehicle for the optimal return route are extracted from the multi-source heterogeneous data of the current collection period. Feature quantization modeling is performed on the geographic information data, environmental interference data, and real-time status data to obtain route geographic quantization features and vehicle status quantization features. Based on the route geographic quantization features and vehicle status quantization features, the estimated load-bearing weight of the target vehicle is obtained.

[0024] Furthermore, the process of adapting capacity tags to order information in multiple dimensions includes:

[0025] When the order pool receives order information, it parses the order information for demand parameters and generates order demand tags. At the same time, it converts the estimated load-bearing weight of the target vehicle, the optimal return route, the return time of the optimal return route, the capacity departure time window and static basic data generated in the current collection period into capacity tags to form a capacity candidate pool.

[0026] Based on the order demand tags, the capacity tags in the capacity candidate pool are filtered by hard constraints, and capacity tags that do not meet the requirements of the order demand tags are removed from the capacity candidate pool.

[0027] After completing the hard constraint filtering, the capacity tags in the capacity candidate pool are matched with the order demand tags to obtain the multi-dimensional fit of each capacity tag, and the comprehensive score of each capacity tag is obtained based on the multi-dimensional fit.

[0028] Furthermore, the process of matching orders for target vehicles based on the comprehensive score of capacity tags includes:

[0029] The top k capacity tags with the highest overall scores are selected, and the target vehicles corresponding to the top k capacity tags are marked as candidate capacity vehicles. The top k candidate capacity vehicles are sorted in ascending order according to their overall scores to generate an order push list.

[0030] Set a confirmation time window, push order information to the first candidate vehicle in the order push list. If the candidate vehicle accepts the order within the confirmation time window, update the status of the candidate vehicle and generate an electronic waybill to send to the candidate vehicle. If the candidate vehicle does not confirm the order within the window period, push order information to the next candidate vehicle in the order push list.

[0031] Furthermore, the process of performing a state update operation includes:

[0032] A preset carrying capacity threshold is set. After a candidate vehicle accepts an order, the remaining carrying weight percentage and remaining carrying volume percentage of the candidate vehicle are obtained. The remaining carrying weight percentage and remaining carrying volume percentage are compared with the carrying capacity threshold. If both the remaining carrying weight percentage and remaining carrying volume percentage are greater than or equal to the carrying capacity threshold, the carrying capacity label of the candidate vehicle in the carrying capacity candidate pool is updated according to the remaining carrying weight percentage and remaining carrying volume percentage.

[0033] If the remaining carryable weight percentage or the remaining carryable volume percentage is less than the carryable capacity threshold, the capacity label of the candidate capacity vehicle in the capacity candidate pool will be removed.

[0034] A logistics return capacity intelligent matching system based on trip prediction includes a cloud monitoring platform, which is communicatively connected to a data analysis module and a capacity matching module.

[0035] The data analysis module is used to predict the capacity departure time window, travel time, spatial correction, emergency factor correction, and carrying capacity weight of the multi-source heterogeneous data of each vehicle in the current collection period, and obtain the capacity departure time window, optimal return route, return time, and estimated carrying capacity weight of the target vehicle.

[0036] The capacity matching module generates capacity tags based on the target vehicle's capacity departure time window, optimal return route, return time, and estimated carrying weight. It then matches the capacity tags with order information in multiple dimensions to obtain a comprehensive score for the capacity tags. Based on this comprehensive score, it performs order matching and status update operations for the target vehicle.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] 1. Unlike the traditional "passive response" capacity matching model, this system uses in-depth analysis of multi-source heterogeneous vehicle data to accurately predict departure time windows, optimal return routes, travel time, and load-bearing weight, clarifying the spatiotemporal characteristics and carrying capacity boundaries of capacity from the source. Based on the prediction results, standardized capacity tags are generated and precisely matched with order demand tags, avoiding invalid matching problems such as "goods waiting for vehicles," "vehicles waiting for goods," and "inability to load after matching" caused by ambiguous capacity status. This ensures the effectiveness of order-capacity matching, improves the overall synergy of logistics transportation, achieves accurate prediction of capacity status, and significantly reduces the invalid matching rate.

[0039] 2. The push strategy based on confirmation time windows can quickly respond to drivers' willingness to accept orders, promptly handle issues such as order rejection and timeouts during the order matching process, ensure the timeliness of order fulfillment, and dynamically update and remove rules for the remaining capacity after an order is accepted, making full use of capacity and improving drivers' income. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating an intelligent matching method for return logistics capacity based on trip prediction, as described in an embodiment of this application.

[0041] Figure 2 This is a flowchart of a logistics return capacity intelligent matching system based on trip prediction, which is an embodiment of this application. Detailed Implementation

[0042] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0043] like Figure 1 As shown, a method for intelligent matching of return freight capacity based on trip prediction includes the following steps:

[0044] Step s1: For each vehicle, perform capacity departure time window prediction, trip time estimation, spatial correction estimation, emergency factor correction estimation, and load-bearing weight prediction on the multi-source heterogeneous data of the current collection period to obtain the target vehicle's capacity departure time window, optimal return route, return time, and estimated load-bearing weight.

[0045] Step s2: Generate capacity tags based on the target vehicle's capacity departure time window, optimal return route, return time, and estimated carrying capacity. Adapt the capacity tags to order information in multiple dimensions to obtain a comprehensive score for the capacity tags. Based on the comprehensive score of the capacity tags, perform order matching and status update operations for the target vehicle.

[0046] It should be further explained that, in the specific implementation process, a cloud monitoring platform is built, and data acquisition terminals (such as smartphone APP + external plug-and-play sensors) are deployed on vehicles and within a preset range (logistics park, unloading location). The cloud monitoring platform communicates with each data acquisition terminal. The data acquisition terminal is used to transmit the collected real-time dynamic data to the cloud monitoring platform and mark the collection period. The cloud monitoring platform is equipped with a vehicle database, an order pool, and a capacity candidate pool. The vehicle database is used to store multi-source heterogeneous data for the current period and several historical collection periods. The multi-source heterogeneous data includes static basic data and real-time dynamic data.

[0047] Static basic data includes: vehicle load capacity, carrying capacity, driving range, transportation qualifications (such as cold chain transportation qualifications, hazardous chemical transportation permits), historical trip time data, driver's historical stop characteristics (such as checking vehicle condition after unloading, handling park departure procedures, resting, etc., usually 10-30 minutes), driver's historical behavior data (historical unloading time distribution at various unloading points, the driver's historical operating efficiency (average unloading time, on-time rate), average loading and unloading time for various types of goods, historical cargo damage rate, and on-time performance rate), and the driver's expected freight rate; and cargo volume, weight, fragility level, temperature control requirements (such as 0-5℃ cold chain requirements), and timeliness requirements (such as delivery within 24 hours) obtained through cargo owner order placement.

[0048] Real-time dynamic data includes: real-time vehicle latitude and longitude and driving speed obtained through GPS / BeiDou positioning modules; real-time traffic conditions (congestion index, traffic control characteristics, special scene characteristics) obtained by calling Gaode and Baidu Maps open interfaces; extreme weather data such as rainstorms and icing along the transportation route obtained through meteorological platform APIs; vehicle models and numbers queuing at loading and unloading points collected through IoT devices deployed in logistics parks (such as cameras, weighbridges, and queuing systems); and vehicle status data such as real-time fuel consumption, battery level, tire pressure, and remaining cargo weight collected through the vehicle OBD interface.

[0049] It should be further explained that, in the specific implementation process, the process of predicting the departure time window for transport capacity includes:

[0050] Extract the operating status of each vehicle from the multi-source heterogeneous data of the current collection period. The operating status includes transportation status and unloading status (including unloading in progress and waiting to unload). Vehicles in the unloading status are marked as target vehicles. Obtain the cargo source characteristics (cargo type and unloading point type), remaining cargo weight, and the queuing vehicle type (the vehicle type of other queuing vehicles in the unloading queue that are ranked before the target vehicle) and the number of each type of queuing vehicle type of the target vehicle from the multi-source heterogeneous data of the current collection period.

[0051] Key feature analysis is performed on the multi-source heterogeneous data of the target vehicle within the historical collection period to obtain the historical average loading and unloading efficiency of the target vehicle for the current cargo characteristics.

[0052] Key feature analysis was performed on multi-source heterogeneous data of several vehicles in the vehicle database during the historical collection period to obtain the historical baseline time for different vehicle types under the current cargo characteristics (the average unloading time calculated based on the combination of "vehicle type-cargo type-unloading point" based on historical data, denoted as...). );

[0053] The baseline unloading time for the target vehicle is obtained based on its remaining cargo weight and historical average loading and unloading efficiency (baseline unloading time = remaining cargo weight / historical average loading and unloading efficiency). The queuing waiting time is then obtained based on the vehicle type queuing at the unloading point, the number of each type of vehicle in the queue, and the historical baseline time (queuing waiting time = ...). Where z represents the vehicle type in the queue. For the number of car models z in the queue, (This represents the historical baseline duration of queued vehicle type z under the current cargo characteristics).

[0054] The total unloading operation time is obtained based on the baseline unloading time and the queuing waiting time (total unloading operation time = baseline unloading time + queuing waiting time). The historical dwelling behavior characteristics of the target vehicle are obtained. Based on the total unloading operation time, historical dwelling behavior characteristics and the current timestamp, the capacity departure time window of the target vehicle is obtained (capacity departure time window = unloading completion timestamp + historical dwelling behavior characteristics. For example, if the unloading completion timestamp is determined to be 13:50 based on the total unloading operation time and the current timestamp, and the historical dwelling behavior characteristics are 10-30 minutes, then the capacity departure time window for the truck to depart from the unloading location is 14:00-14:20).

[0055] It should be further explained that, in the specific implementation process, the process of estimating travel time and making spatial correction estimates includes:

[0056] Obtain the unloading point and return destination of the target vehicle, construct a road transport network containing several return routes based on the unloading point and return destination, define the node types in the road transport network and the road segments connecting adjacent nodes, and add road segment length, real-time traffic speed and congestion index to the road segments connecting adjacent nodes based on the real-time dynamic data of the current collection period.

[0057] The travel time data of the target vehicle in several historical collection periods is extracted from the vehicle database (the return route of the same unloading point and return destination in the past 6 months, and the travel time data of different departure times are selected, covering different time scenarios such as weekdays / weekends, off-peak periods, and peak periods). Core time-series features (historical time-series (sorted by timestamp, such as one data point every 5 minutes), average daily time-series, and standard deviation of time-series (reflecting the degree of fluctuation)) and time-derived features (weekday type (weekdays / weekends, one-hot encoding), time period labels (morning peak / noon / evening peak / night), and holiday markers (such as special periods such as Spring Festival, 1 indicates holidays, and 0 indicates non-holidays)) are extracted. The core time-series features and time-derived features are marked as training data. An LSTM model is used to build a travel time prediction model. The travel time prediction model is trained using the training data to obtain the completed travel time prediction model.

[0058] The input layer of the trip time prediction model has a time step of 24 (i.e., predicting the time of the next node based on the time data of the previous 24 time nodes); hidden layers consist of two LSTM layers with 64 neurons each, using the ReLU activation function; the output layer is a fully connected layer that outputs the predicted time value for a single time step. Training strategies include: loss function: mean squared error (MSE) to minimize the deviation between predicted and actual time; optimizer: Adam optimizer with a learning rate of 0.001; regularization: a Dropout layer (dropout=0.2) is added to prevent overfitting. The core output outputs the basic predicted time values ​​for each return path based on historical time series patterns, while also capturing time fluctuation characteristics (such as Monday morning rush hour time being 20% ​​higher than off-peak time, and holiday time being 30% higher than weekday time).

[0059] The return routes between unloading points and return destinations in the road transport network are input into the travel time prediction model, and the model outputs the basic travel time prediction values ​​for each return route. ;

[0060] The topology of several return routes in the road transport network is represented in graph form. A graph attention network is used to learn this graph representation. Training data, along with real-time traffic speeds and congestion indices for each road segment, are input into the graph attention network. The learned graph attention network then outputs the spatial correction time for each return route. ;

[0061] Traditional LSTM cannot depict the spatial relationships of road networks (such as congestion at highway entrances leading to traffic backlog on subsequent road segments), while GraphTransformer can abstract the transportation network into a graph structure, accurately modeling the spatial chain effects between road segments, and spatially correcting the basic time consumption of LSTM.

[0062] The graph consists of a set of nodes and a set of edges. The set of nodes contains key nodes in the road network, such as intersections, toll stations, highway entrances and exits, loading and unloading points, etc. The set of edges consists of road segments connecting adjacent nodes. Each edge has a weight attribute, including the length of the road segment, the real-time traffic speed, and the current congestion index (1-10 points, the higher the score, the more severe the congestion).

[0063] Input features of graph attention network:

[0064] Node features: Node type; Edge features: Segment length, real-time traffic speed, current congestion index; Global features: Location coordinates of the start point (unloading point) and end point (return destination) of the return route.

[0065] The specific process of outputting the spatial correction time of each return path through the attention mechanism includes:

[0066] The attention weights between any two nodes are calculated using a multi-head attention mechanism to capture the spatial chain reaction coefficients between road segments. For example, when the highway entrance node ( When the congestion index of a certain point is 8, its impact on subsequent highway nodes can be quantified through attention weighting. The congestion transmission coefficient (e.g., the spatial chain reaction coefficient) is 0.7, i.e. ( Congestion can lead to ( ) (Traffic efficiency decreases by 70%); Graph convolution operation: updates the feature vector of each node based on attention weights, outputs the spatial congestion propagation matrix of the entire road network, and calculates the change in time consumption caused by spatial chain effects based on the congestion propagation matrix. —If there is congestion transmission between road sections A positive value (increased time consumption); if the road network is unobstructed, Approaching 0.

[0067] It should be further explained that, in the specific implementation process, the process of making corrective predictions for unforeseen factors includes:

[0068] The travel time data of target vehicles containing characteristics of sudden factors within several historical collection periods are extracted from the vehicle database as sample data. Sudden factor characteristics include special weather characteristics, traffic control characteristics, and special scenario characteristics. Among them: special weather characteristics: weather type (heavy rain / fog / snowstorm, unique thermal encoding), visibility, rainfall (mm), wind force level; traffic control characteristics: temporary restricted road section markings (1 indicates restricted, 0 indicates no restriction), restricted duration, location of road construction sections; special scenario characteristics: road closure markings for large-scale events (such as marathons, exhibitions), number of temporarily closed highway toll stations). For each sample data, travel time is estimated and spatially corrected is estimated to obtain the basic travel time prediction value and the spatially corrected travel time. The sum of the basic travel time prediction value and the spatially corrected travel time is compared with the actual travel time in the sample data to obtain the travel time difference value. The travel time difference value is then bound to the label value of the sample data.

[0069] An XGBoost model is used to construct an emergency factor prediction model. The model is trained using sample data with completed label binding. Trip records containing emergency factor characteristics (heavy rain, traffic control, etc.) from the past six months are collected. The difference between the predicted time (LSTM+GraphTransformer) and the actual time is used as the label value. The emergency factor feature set is used as input features to obtain the trained emergency factor prediction model. Emergency factor features are extracted from real-time dynamic data of the target vehicle during the current collection period for each return route. These features are then input into the emergency factor prediction model, which outputs the time correction amount for each return route. ;

[0070] The time corrections for each return path, based on the prediction model for unforeseen factors, are as follows:

[0071] The model for predicting sudden factors uses a regression loss function to minimize the difference between the prediction error and the actual error; and optimizes the tree depth (set to 6), learning rate (set to 0.01), and subsample ratio (set to 0.8) through grid search; feature importance ranking: outputs the influence weight of each sudden factor on the time consumption (e.g., the influence weight of rainstorm is 0.3, and the influence weight of temporary traffic restriction is 0.25).

[0072] Correction Calculation: Input real-time characteristics of sudden factors; output the time correction amount caused by sudden factors in the sudden factor prediction model. For example, during heavy rain. For 30 minutes, under temporary traffic restrictions It lasts for 20 minutes;

[0073] Based on the predicted base travel time for each return route Space correction time and time-consuming correction amount Obtain the return time for each return path , = + + The return path with the shortest return time is selected and marked as the optimal return path.

[0074] It should be further explained that, in the specific implementation process, the process of predicting the load-bearing capacity includes:

[0075] Geographic information data, environmental interference data, and real-time status data of the target vehicle for the optimal return route are extracted from the multi-source heterogeneous data of the current collection period. Feature quantization modeling is performed on the geographic information data, environmental interference data, and real-time status data to obtain route geographic quantization features and vehicle status quantization features. Based on the route geographic quantization features and vehicle status quantization features, the estimated load-bearing weight of the target vehicle is obtained.

[0076] Among them, the vehicle condition quantification characteristics include the tire pressure safety factor. and dynamic performance coefficient , ;in, For tire pressure, Standard tire pressure for the target vehicle model; ;

[0077] The geographical quantitative characteristics of the route include: slope influence coefficient. Altitude Influence Coefficient Wind force influence coefficient , ,in, For the maximum slope, This represents the total length of the uphill section. This represents the total length of the return route. For example, when the maximum gradient is 10° and the uphill section accounts for 20%, This means that a 20% load margin needs to be reserved; , ;in, The highest altitude is indicated by F, and the wind speed rating is indicated by F.

[0078] First, calculate the foundation bearing capacity considering the effects of slope and altitude. That is, the maximum safe load when climbing a hill:

[0079] ;in, The rated load capacity indicated on the vehicle registration certificate of the target vehicle;

[0080] Subsequently, considering factors such as tire pressure, power performance, and wind force, the load-bearing capacity of the basic climbing structure was assessed. Make corrections to obtain the estimated load-bearing weight of the target vehicle. : .

[0081] It should be further explained that the remaining load capacity within the vehicle's departure time window is obtained. ,like > Then take = (The load must not exceed the remaining load within the vehicle's departure time window).

[0082] It should be further explained that, in the specific implementation process, the multi-dimensional adaptation of capacity tags and order information includes:

[0083] When the order pool receives order information from a shipper, it parses the order information for demand parameters and generates order demand tags (including cargo attributes (weight, volume, type), spatiotemporal constraints (loading time window, loading location, unloading location, timeliness requirements), and cost and service constraints (acceptable freight rates, cargo damage compensation requirements)). Simultaneously, it converts the estimated carrying capacity of the target vehicle, optimal return route, return time of the optimal return route, capacity departure time window, and static basic data generated within the current collection period into capacity tags (including: time tags, capacity departure time window, return time; spatial tags: return route nodes, connection time from loading point to the vehicle's current location; carrying capacity tags: estimated carrying weight, carrying capacity, transportation qualifications (cold chain / hazardous materials); cost tags: route tolls, driver's expected freight rate), forming a capacity candidate pool.

[0084] Based on the order demand tags, the capacity tags in the capacity candidate pool are filtered by hard constraints, and capacity tags that do not meet the requirements of the order demand tags are removed from the capacity candidate pool.

[0085] Among them, the label requirements for order requirements include:

[0086] Load capacity matching: (estimated load capacity weight ≥ order weight) and (load capacity volume ≥ order volume); cold chain goods must be matched with temperature-controlled vehicles, and hazardous chemicals must be matched with vehicles with corresponding qualifications;

[0087] Time matching: The overlap between the capacity departure time window and the order loading time window is ≥50%; the total transportation time (connection time + return time) ≤ the order's timeliness requirement;

[0088] Spatial matching: The order unloading location is on the return route of the transportation capacity (or the detour distance is ≤10 kilometers);

[0089] Cost matching: Driver's expected freight rate ≤ Order freight rate ceiling;

[0090] After completing the hard constraint filtering, the capacity tags in the capacity candidate pool are matched with the order demand tags to obtain the multi-dimensional fit of each capacity tag, and the comprehensive score of each capacity tag is obtained based on the multi-dimensional fit.

[0091] It should be further explained that multi-dimensional adaptability includes time adaptability, capacity adaptability, path adaptability, service adaptability, and cost adaptability, as shown in Table 1 below:

[0092] Table 1

[0093]

[0094] Overall Score: ,in For dimension weights, Score the dimension.

[0095] It should be further explained that, in the specific implementation process, the process of matching target vehicle orders based on the comprehensive score of capacity tags includes:

[0096] Select the top k (k=10) capacity tags with the highest comprehensive scores, mark the target vehicles corresponding to the top k capacity tags as candidate capacity vehicles, sort the top k candidate capacity vehicles in ascending order according to their comprehensive scores, and generate an order push list (the higher the comprehensive score, the higher the ranking).

[0097] Set a confirmation time window (5 minutes), push order information to the first candidate vehicle in the order push list. If the candidate vehicle accepts the order within the confirmation time window, update the status of the candidate vehicle and generate an electronic waybill to send to the candidate vehicle. If the candidate vehicle does not confirm the order within the window period, push order information to the next candidate vehicle in the order push list.

[0098] It should be further explained that, in the specific implementation process, the state update operation includes:

[0099] A preset carrying capacity threshold (50%) is set. After a candidate vehicle accepts an order, the remaining carrying weight percentage ((estimated carrying weight - order weight) / estimated carrying weight) and the remaining carrying volume percentage ((carrying volume - order volume) / carrying volume) of the candidate vehicle are obtained. The remaining carrying weight percentage and the remaining carrying volume percentage are compared with the carrying capacity threshold. If both the remaining carrying weight percentage and the remaining carrying volume percentage are greater than or equal to the carrying capacity threshold, the carrying capacity labels of the candidate vehicles in the carrying capacity candidate pool are updated according to the remaining carrying weight percentage and the remaining carrying volume percentage (replacing the original estimated carrying weight with the remaining carrying weight corresponding to (estimated carrying weight - order weight), and replacing the original carrying volume with the remaining carrying volume corresponding to (carrying volume - order volume)).

[0100] If the remaining carryable weight percentage or the remaining carryable volume percentage is less than the carryable capacity threshold, the capacity label of the candidate capacity vehicle in the capacity candidate pool will be removed.

[0101] like Figure 2 As shown, a logistics return capacity intelligent matching system based on trip prediction includes a cloud monitoring platform, which is communicatively connected to a data analysis module and a capacity matching module.

[0102] The data analysis module is used to predict the capacity departure time window, travel time, spatial correction, emergency factor correction, and carrying capacity weight of the multi-source heterogeneous data of each vehicle in the current collection period, and obtain the capacity departure time window, optimal return route, return time, and estimated carrying capacity weight of the target vehicle.

[0103] The capacity matching module generates capacity tags based on the target vehicle's capacity departure time window, optimal return route, return time, and estimated carrying weight. It then matches the capacity tags with order information in multiple dimensions to obtain a comprehensive score for the capacity tags. Based on this comprehensive score, it performs order matching and status update operations for the target vehicle.

[0104] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A logistics return trip transport capacity intelligent matching method based on trip prediction, characterized in that, Includes the following steps: Step s1: For each vehicle, perform capacity departure time window prediction, trip time estimation, spatial correction estimation, emergency factor correction estimation, and load-bearing weight prediction on the multi-source heterogeneous data of the current collection period to obtain the target vehicle's capacity departure time window, optimal return route, return time, and estimated load-bearing weight. The process of forecasting the departure time window for transport capacity includes: Extract the operating status of each vehicle from the multi-source heterogeneous data of the current collection period, mark the vehicles in the unloading state as target vehicles, and obtain the cargo source characteristics, remaining cargo weight, and the queuing vehicle type and the number of each type of queuing vehicle at the unloading point of the target vehicle. Key feature analysis is performed on the multi-source heterogeneous data of the target vehicle within the historical collection period to obtain the historical average loading and unloading efficiency of the target vehicle for the current cargo characteristics. Key feature analysis was performed on multi-source heterogeneous data of several vehicles in the vehicle database during the historical collection period to obtain the historical baseline duration of different vehicle types under the current cargo source characteristics. The baseline unloading time of the target vehicle is obtained based on the remaining cargo weight of the target vehicle and the historical average loading and unloading efficiency. The queuing waiting time is obtained based on the vehicle type in the queue at the unloading point, the number of each type of vehicle type in the queue, and the historical baseline time. The total unloading operation time is obtained based on the baseline unloading time and queuing waiting time. The historical dwelling behavior characteristics of the target vehicle are obtained. The capacity departure time window of the target vehicle is obtained based on the total unloading operation time, historical dwelling behavior characteristics and the current timestamp. The process of making travel time estimates and spatial correction estimates includes: Obtain the unloading point and return destination of the target vehicle, construct a road transport network containing several return routes based on the unloading point and return destination, define the node types in the road transport network and the road segments connecting adjacent nodes, and add road segment length, real-time traffic speed and congestion index to the road segments connecting adjacent nodes based on the real-time dynamic data of the current collection period. The travel time data of the target vehicle in several historical collection periods is extracted from the vehicle database. Core time-series features and time-derived features are extracted from the travel time data. The core time-series features and time-derived features are marked as training data to construct a travel time prediction model. The travel time prediction model is trained using the training data to obtain a completed travel time prediction model. Input the return route between the unloading point and the return destination in the road transport network into the travel time prediction model, and output the basic travel time prediction value of each return route based on the travel time prediction model. The topology of several return routes in the road transport network is represented in graph form. The graph attention network is used to learn the graph representation of the topology. The training data and the real-time traffic speed and congestion index of each road segment are imported into the graph attention network. The spatial correction time of each return route is output through the graph attention network after the learning is completed. The process of making revised predictions for unforeseen factors includes: Extract travel time data of target vehicles containing sudden factors within several historical collection periods from the vehicle database as sample data. Perform travel time prediction and spatial correction prediction on each sample data to obtain the basic travel time prediction value and the spatial correction travel time. Compare the sum of the basic travel time prediction value and the spatial correction travel time with the actual travel time in the sample data to obtain the travel time difference value. Bind the travel time difference value to the label value of the sample data. Construct a sudden factor prediction model, train the sudden factor prediction model using sample data with completed label value binding, obtain the completed sudden factor prediction model, extract the sudden factor features contained in each return path from the real-time dynamic data of the target vehicle in the current collection period, input the sudden factor features into the sudden factor prediction model, and output the time correction amount of each return path according to the sudden factor prediction model. The return time of each return path is obtained by spatially correcting the return time and the time correction amount based on the basic time prediction value of each return path, and the return path with the shortest return time is selected and marked as the optimal return path. The process of predicting load-bearing capacity includes: Geographic information data, environmental interference data, and real-time status data of the target vehicle for the optimal return route are extracted from the multi-source heterogeneous data of the current collection period. Feature quantization modeling is performed on the geographic information data, environmental interference data, and real-time status data to obtain route geographic quantization features and vehicle status quantization features. Based on the route geographic quantization features and vehicle status quantization features, the estimated load-bearing weight of the target vehicle is obtained. Step s2: Generate capacity tags based on the target vehicle's capacity departure time window, optimal return route, return time, and estimated carrying weight. Adapt the capacity tags to order information in multiple dimensions to obtain a comprehensive score for the capacity tags. Perform order matching and status update operations for the target vehicle based on the comprehensive score of the capacity tags. The process of adapting capacity tags to order information in multiple dimensions includes: When the order pool receives order information, it parses the order information for demand parameters and generates order demand tags. At the same time, it converts the estimated load-bearing weight of the target vehicle, the optimal return route, the return time of the optimal return route, the capacity departure time window and static basic data generated in the current collection period into capacity tags to form a capacity candidate pool. Based on the order demand tags, the capacity tags in the capacity candidate pool are filtered by hard constraints, and capacity tags that do not meet the requirements of the order demand tags are removed from the capacity candidate pool. After completing the hard constraint filtering, the capacity tags in the capacity candidate pool are matched with the order demand tags to obtain the multi-dimensional fit of each capacity tag, and the comprehensive score of each capacity tag is obtained based on the multi-dimensional fit.

2. The intelligent matching method for return freight capacity based on trip prediction according to claim 1, characterized in that, A cloud-based monitoring platform is constructed, and data acquisition terminals are deployed on vehicles and within a preset range. The cloud-based monitoring platform communicates with each data acquisition terminal. The data acquisition terminals are used to transmit the collected real-time dynamic data to the cloud-based monitoring platform and mark the collection period. The cloud-based monitoring platform is equipped with a vehicle database, an order pool, and a capacity candidate pool. The vehicle database is used to store multi-source heterogeneous data for the current period and several historical collection periods. The multi-source heterogeneous data includes static basic data and real-time dynamic data.

3. The intelligent matching method for return freight capacity based on trip prediction according to claim 2, characterized in that, The process of matching orders for target vehicles based on a comprehensive score of capacity tags includes: The top k capacity tags with the highest overall scores are selected, and the target vehicles corresponding to the top k capacity tags are marked as candidate capacity vehicles. The top k candidate capacity vehicles are sorted in ascending order according to their overall scores to generate an order push list. Set a confirmation time window, push order information to the first candidate vehicle in the order push list. If the candidate vehicle accepts the order within the confirmation time window, update the status of the candidate vehicle and generate an electronic waybill to send to the candidate vehicle. If the candidate vehicle does not confirm the order within the window period, push order information to the next candidate vehicle in the order push list.

4. The intelligent matching method for return freight capacity based on trip prediction according to claim 3, characterized in that, The process of performing a state update operation includes: A preset carrying capacity threshold is set. After a candidate vehicle accepts an order, the remaining carrying weight percentage and remaining carrying volume percentage of the candidate vehicle are obtained. The remaining carrying weight percentage and remaining carrying volume percentage are compared with the carrying capacity threshold. If both the remaining carrying weight percentage and remaining carrying volume percentage are greater than or equal to the carrying capacity threshold, the carrying capacity label of the candidate vehicle in the carrying capacity candidate pool is updated according to the remaining carrying weight percentage and remaining carrying volume percentage. If the remaining carryable weight percentage or the remaining carryable volume percentage is less than the carryable capacity threshold, the capacity label of the candidate capacity vehicle in the capacity candidate pool will be removed.

5. A logistics return capacity intelligent matching system based on trip prediction, specifically applied to the logistics return capacity intelligent matching method based on trip prediction as described in any one of claims 1 to 4, characterized in that, This includes a cloud-based monitoring platform, which has communication connections with a data analysis module and a capacity matching module. The data analysis module is used to predict the capacity departure time window, travel time, spatial correction, emergency factor correction, and carrying capacity weight of the multi-source heterogeneous data of each vehicle in the current collection period, and obtain the capacity departure time window, optimal return route, return time, and estimated carrying capacity weight of the target vehicle. The capacity matching module generates capacity tags based on the target vehicle's capacity departure time window, optimal return route, return time, and estimated carrying weight. It then matches the capacity tags with order information in multiple dimensions to obtain a comprehensive score for the capacity tags. Based on this comprehensive score, it performs order matching and status update operations for the target vehicle.

Images