Intelligent carpooling management system based on deep learning

Abstract

The invention relates to the technical field of carpooling scheduling, in particular to an intelligent carpooling management system based on deep learning, which is used for calculating the path similarity between a request order and a to-be-selected carpooling order based on starting and ending point information of the request order and real-time road network topology, obtaining path overlap ratio data and providing a basic set for path splitting. The method comprises the steps of dividing a request order path into a plurality of sub-paths with transfer feasibility by combining time window constraints, outputting a corresponding transfer node sequence, and further matching each section of connection vehicles according to the transfer node sequence and vehicle real-time positioning information to generate a scheduling scheme of multi-vehicle splicing. And finally, a scheduling instruction is issued to a related vehicle terminal, and execution control is completed in a closed loop. The carpooling matching rate and the resource utilization efficiency are effectively improved, and the problems of resource waste and matching failure caused by a whole order strategy are relieved.

Description

Technical Field

[0001] The present invention relates to the technical field of carpooling scheduling, and in particular to an intelligent carpooling management system based on deep learning. Background Art

[0002] Ride-sharing has been widely used as an effective way to alleviate urban traffic congestion and improve vehicle utilization. However, existing ride-sharing systems generally employ a direct, full-order order matching strategy. This strategy uses the passenger's complete travel path (from origin to destination) as the basic scheduling unit. By comparing the route similarity and time window overlap between the requested order and the platform's pending orders, a vehicle with a highly matched route and time window is selected as the carrier. While this approach offers high scheduling efficiency in a few short-distance scenarios where routes completely overlap and boarding and alighting locations are close, it suffers from significant limitations in most real-world travel scenarios. These limitations include: Passenger order paths in cities are often spatially dispersed with limited route intersections, resulting in a significant number of orders overlapping only on certain sections. Because the full-order strategy only considers the similarity of complete routes, if the overlap is insufficient, even segments with potential for ride-sharing will be excluded, resulting in missed scheduling opportunities. When the carpooling system cannot find a complete vehicle that can carry a certain order, it often directly abandons the carpooling matching of the order and switches to ordinary exclusive order allocation, resulting in a limited overall carpooling success rate of the system, a high vehicle idle rate and a high passenger travel cost. Summary of the Invention

[0003] To solve the above problems, the present invention provides an intelligent carpooling management system based on deep learning.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] An intelligent carpooling management system based on deep learning, comprising:

[0006] The route matching module is used to calculate the similarity between the driving paths of the requested order and the selected carpooling order based on the starting and ending point information of the requested order and the real-time road network topology, and obtain the route overlap data;

[0007] A transfer node splitting module is used to split the request order based on the route overlap data and the time window constraints of each order to obtain target transfer node sequence data;

[0008] A vehicle matching module is used to match the target transfer node sequence data with the vehicle real-time positioning data based on a reinforcement learning model to obtain vehicle matching data;

[0009] The carpooling dispatch module outputs dispatch instructions based on vehicle matching data and sends them to the corresponding vehicle driver terminal.

[0010] Furthermore, the route matching module is used to perform the following steps:

[0011] Based on the starting point, end point, and historical trajectory samples of the requested order, a pre-trained convolutional neural network is used to extract features from the candidate driving paths of the requested order to obtain a path feature vector of the requested path.

[0012] Based on the historical driving path data of the carpooling orders to be selected, the convolutional neural network is used to perform feature extraction on each candidate carpooling path to obtain a path feature vector for each carpooling order;

[0013] A cosine similarity calculation is performed on the path feature vector of the requested path and the path feature vectors of each carpooling order to obtain route overlap data between the requested order and each carpooling order to be selected.

[0014] Furthermore, the training steps of the convolutional neural network are as follows:

[0015] Based on a path dataset containing historical order trajectories, we extract the path sequence and travel time corresponding to each order, construct path sequence sample pairs, and annotate the actual overlap between each pair of paths.

[0016] Convert the path sequence into a fixed-length segment embedding vector sequence, input the sequence into a convolutional neural network for feature extraction, and obtain a path representation vector;

[0017] Calculate the cosine similarity between the path representation vectors as the output value of the network, and calculate the loss value between the output value of the network and the actual coincidence label using the mean square error loss function;

[0018] With the goal of minimizing the loss value, the batch gradient descent method is used to iteratively train the parameters of the convolutional neural network until convergence reaches the preset threshold.

[0019] Furthermore, the transfer node splitting module is used to perform the following steps:

[0020] Based on the route overlap data, combined with the time window constraints of the requested order and the road segment nodes, an input tensor including path overlap, node travel time and transfer waiting tolerance is constructed;

[0021] Input the input tensor into the Transformer model, perform node analysis on the request order path, and obtain a probability value vector of each node as a transfer point;

[0022] Based on the probability value vector of each node as a transfer point, a number of high-confidence transfer nodes that meet the preset time constraints and spatial segmentation rationality are screened to generate target transfer node sequence data.

[0023] Furthermore, the road section nodes include road section intersection points and temporary parking areas.

[0024] Furthermore, the transfer waiting tolerance is calculated and generated by statistical results of actual waiting time of passengers at each node in historical orders.

[0025] Furthermore, the construction of the Transformer model includes the following steps:

[0026] Constructing a path node sequence sample based on historical carpooling order data, wherein the path node sequence sample includes the path node sequence corresponding to each order, the time window constraint, the actual transfer node location, and the transfer success label;

[0027] Encoding the path node sequence, using position coding to mark the relative order of the nodes in the path, and constructing a node feature vector sequence based on the path overlap, travel time and transfer feasibility of each node;

[0028] The node feature vector sequence is input into the Transformer encoder to capture the interaction between nodes in terms of time constraints, overlap, and traffic efficiency, and output each node as a representation vector of the transfer node;

[0029] The representation vector is input into the fully connected prediction layer to output the probability value vector of the transfer point and compared with the historical transfer labels, and the model parameters are optimized through the cross entropy loss function.

[0030] Furthermore, the reinforcement learning model is trained by the following steps:

[0031] Constructing a state space based on historical dispatch records, which includes the target transfer node sequence, the time window parameters of the requested order, the real-time location and empty status of candidate vehicles, the predicted time when the vehicle can reach the transfer node, and the travel time between transfer nodes;

[0032] Based on the state space, determining an action space, the action space comprising selecting a candidate vehicle from a set of candidate vehicles for each transfer node;

[0033] Constructing a reward function based on the idle distance of vehicles in chain scheduling and the difference between the total travel time and the direct travel time of passengers;

[0034] Based on the state space, action space and reward function, a deep Q network algorithm is used for training, which takes the state and action as input, outputs the corresponding Q value, and updates the network parameters by minimizing the mean square error between the current Q value and the target Q value.

[0035] Furthermore, the reward function is as follows:

[0036] R=R max -(α·(T chain -T direct )+β·D idle );

[0037] Among them, R is the reward value; R max is the initial excitation value; T direct The theoretical direct access time; T chain The total time required to complete the entire trip by matching the vehicle solution corresponding to the requested order; D idle is the cumulative empty driving distance of the vehicle before executing the task in the vehicle matching scheme; α is the time difference coefficient; β is the empty driving distance coefficient.

[0038] Furthermore, the carpooling scheduling module is used to perform the following steps:

[0039] Based on the vehicle matching data, the vehicle number, connection sequence and estimated arrival time of each connecting route are extracted, and the route travel time and traffic status information are combined to generate a dispatch instruction including the vehicle number, starting and ending node locations, and task time period;

[0040] The dispatch instruction is sent to the corresponding vehicle driver terminal.

[0041] The beneficial effects of the present invention are as follows: By combining the origin and destination information of a requested order with the real-time road network topology, the present invention calculates the path similarity between the requested order and the selected carpooling orders, obtaining accurate route overlap data, which provides a structured basis for subsequent splitting and matching. Subsequently, based on the route overlap data and the time window constraints of each order, the original requested order path is split into several sub-paths with transfer feasibility, and the corresponding target transfer node sequence data is output, thus breaking through the rigid scheduling method of traditional carpooling systems based on entire orders. Furthermore, the vehicle matching module further performs matching calculations between the segmented paths and candidate vehicles based on the split node sequence data and the real-time positioning information of the vehicles, and outputs the vehicle matching results for the multi-vehicle splicing. Finally, the carpooling scheduling module generates scheduling instructions based on the matching results and issues them to the corresponding vehicle driver terminals, ensuring closed-loop control of scheduling execution. The present invention implements dynamic splitting of request order paths and multi-vehicle collaborative matching, improving the ability to utilize orders with partially overlapping paths, significantly enhancing the scheduling flexibility and resource utilization efficiency of the carpooling system, and effectively alleviating the resource waste and matching failure problems caused by the direct access restrictions of entire orders. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a structural diagram of an intelligent carpooling management system based on deep learning in the present invention.

[0043] Figure 2It is a flow chart of the training steps of the reinforcement learning model in the present invention. DETAILED DESCRIPTION

[0044] See also Figure 1-Figure 2 As shown, the present invention relates to an intelligent carpooling management system based on deep learning, comprising:

[0045] The route matching module is used to calculate the similarity between the driving paths of the requested order and the selected carpooling order based on the starting and ending point information of the requested order and the real-time road network topology, and obtain the route overlap data;

[0046] A transfer node splitting module is used to split the request order based on the route overlap data and the time window constraints of each order to obtain target transfer node sequence data;

[0047] A vehicle matching module is used to match the target transfer node sequence data with the vehicle real-time positioning data based on a reinforcement learning model to obtain vehicle matching data;

[0048] The carpooling dispatch module outputs dispatch instructions based on vehicle matching data and sends them to the corresponding vehicle driver terminal.

[0049] It should be noted that the present invention provides a deep learning-based intelligent carpooling management system that can achieve efficient and dynamic carpooling scheduling optimization in a multi-source heterogeneous traffic data environment. Taking the morning rush hour commuting scenario as an example, assume that user A submits a trip request in the system, starting from a residential area in the southern part of the city and ending at an industrial park in the northern part, with a travel time window between 7:50 and 8:20 a.m. The system first invokes the route matching module to obtain user A's starting and ending points. Combined with the current real-time road network topology, the system generates multiple feasible driving routes using a predefined path search algorithm. The system then enhances the current set of candidate paths using historical order trajectories and regional high-frequency path data. The system then invokes a trained deep path representation model to calculate the path feature similarity between user A's requested path and other potential carpooling orders on the platform. This outputs a set of structured route overlap data, which characterizes the degree of overlap between the paths and their spatial topological consistency. Based on this overlap data, the transfer node segmentation module logically segments user A's path, taking into account the boarding and alighting time windows of each order. The system automatically analyzes route nodes with high overlap and relatively loose time tolerances and identifies two high-confidence transfer nodes, located at the intersection of the central city business district and the subway. Through this node-level splitting mechanism, user A's original point-to-point travel request is dynamically divided into three segments: the first segment is connected by vehicle B from the residential area to the business district, the second segment is connected by vehicle C from the business district to the subway entrance, and the third segment is completed by vehicle D to the destination. Compared to the existing integrated matching method based on the entire path, this invention for the first time links the selection of transfer nodes with the calculation of path overlap and filters and confirms the scheduling feasibility of transfer nodes using time window parameters, avoiding scheduling failures caused by partial path overlap but time conflicts. After the path segmentation is completed, the vehicle matching module is activated. The system uses the real-time location, unloaded status, expected driving trajectory, and accessibility indicators of all vehicles currently to be dispatched to construct the state space of the reinforcement learning model, and simultaneously considers each transfer node combination as a candidate action. Guided by the policy network, the system evaluates the contribution of each possible vehicle segment scheduling combination to overall travel efficiency and resource utilization, thereby selecting the optimal chain carpooling combination and outputting the vehicle matching results. Compared with traditional greedy algorithms based on shortest distance or optimal arrival time, this reinforcement learning matching method has adaptive learning capabilities and policy generalization capabilities, and can maintain scheduling robustness under complex traffic conditions. Ultimately, the carpooling scheduling module receives the chain matching results and generates a scheduling instruction set, which clearly specifies the pick-up and drop-off time, boarding and alighting nodes, and task priority for each vehicle. The instructions are issued through the system to the corresponding vehicle driver terminals, and the drivers perform the path connection tasks according to the scheduling interface.On the one hand, this system uses deep learning modeling to achieve fine-grained expression of path overlap, avoiding the problem that traditional path encoding methods cannot characterize local structural differences; on the other hand, it introduces a dynamic splitting strategy for transfer nodes to reconstruct originally single and unmatched orders into multiple segments, thereby maximizing the efficiency of carpooling utilization; combined with reinforcement learning-driven strategy optimization, it realizes cross-vehicle, multi-stage, and real-time adjustable carpooling scheduling path planning.

[0050] Furthermore, the route matching module is used to perform the following steps:

[0051] Based on the starting point, end point, and historical trajectory samples of the requested order, a pre-trained convolutional neural network is used to extract features from the candidate driving paths of the requested order to obtain a path feature vector of the requested path.

[0052] Based on the historical driving path data of the carpooling orders to be selected, the convolutional neural network is used to perform feature extraction on each candidate carpooling path to obtain a path feature vector for each carpooling order;

[0053] A cosine similarity calculation is performed on the path feature vector of the requested path and the path feature vectors of each carpooling order to obtain route overlap data between the requested order and each carpooling order to be selected.

[0054] In some embodiments, suppose user A submits a travel request starting from Community A and ending at Office Building B. When the system invokes the route matching module, it first generates several candidate routes within the road network based on the user's historical travel history and current starting and ending locations. Each route is encoded as a structured sequence consisting of a segment number, travel direction, and travel time, and fed as input into a pre-trained convolutional neural network model. This convolutional neural network uses a multi-layer one-dimensional convolutional structure to perform sliding window feature extraction on the path sequence. This extraction process not only considers the spatial topological characteristics of the segments within the path, but also captures the adjacency between segments, travel speed patterns, and historical high-frequency co-occurrence structures. By setting multiple convolution kernel sizes to adapt to the characteristics of path segments of varying lengths, the network automatically learns the local structural variations and overall morphological distribution of the path, and then constructs a high-dimensional path representation vector in the intermediate layer as an embedded representation of the requested route. In parallel, the system also invokes the aforementioned convolutional neural network model for available rideshare orders on the platform, encoding each historical route and extracting its corresponding path feature vector. These path vectors are uniformly mapped to the same embedding space to facilitate subsequent similarity calculations. After completing the embedding extraction of the requested path and each candidate path, the system performs cosine similarity calculations on the feature vectors of each path, that is, by calculating the cosine value of the angle between the vectors to characterize the degree of overlap between the paths, and outputs the route overlap score between the requested order and each carpooling order to be selected. This route overlap score not only reflects the similarity of the paths in spatial distribution, but also characterizes the consistency of the paths in historical traffic behavior patterns through deep feature abstraction capabilities, which is significantly better than traditional path comparison algorithms based on string matching or point set intersection. In actual applications, the system sorts according to the similarity score and retains several candidate carpooling paths with high overlap as input to the subsequent transfer node splitting and vehicle matching modules, effectively improving the accuracy of carpooling path identification and scheduling efficiency.

[0055] Furthermore, the training steps of the convolutional neural network are as follows:

[0056] Based on a path dataset containing historical order trajectories, we extract the path sequence and travel time corresponding to each order, construct path sequence sample pairs, and annotate the actual overlap between each pair of paths.

[0057] Convert the path sequence into a fixed-length segment embedding vector sequence, input the sequence into a convolutional neural network for feature extraction, and obtain a path representation vector;

[0058] Calculate the cosine similarity between the path representation vectors as the output value of the network, and calculate the loss value between the output value of the network and the actual coincidence label using the mean square error loss function;

[0059] With the goal of minimizing the loss value, the batch gradient descent method is used to iteratively train the parameters of the convolutional neural network until convergence reaches the preset threshold.

[0060] In some embodiments, during implementation, convolutional neural network training aims to construct an embedding representation model that accurately captures the spatial structural similarities and common characteristics of traffic behavior between paths. The system first constructs a path training dataset derived from a collection of trajectories accumulated from historical ridesharing orders. Each historical trajectory is represented as a sequence of paths from a starting point to a destination, consisting of consecutively numbered segments, along with travel time, speed category, and direction information for each segment. To construct training pairs, the system constructs positive and negative sample sets for any two path combinations. Positive pairs consist of paths that partially overlap or have a high degree of spatial proximity, while negative pairs consist of paths that are completely different or have significant regional distribution differences. Furthermore, by measuring the actual proportion of overlapping segments between the paths, the system assigns an overlap label between 0 and 1 to each pair of paths as a training supervisory signal. During the data preprocessing phase, the system uses an embedding coding mechanism to vectorize the segment numbers in the path sequence. Specifically, a trainable embedding vector is assigned to each unique segment number, which is then combined with its travel time and other auxiliary attributes to form a unified segment feature representation. To handle variable-length paths, all path sequences are zero-padded or truncated to a fixed length L, forming an input tensor of the form [L, d], where d is the feature dimension of a single path segment. This tensor is fed into a multi-layer one-dimensional convolutional neural network. The convolutional layers extract local path segment features using sliding windows of varying sizes. The subsequent fully connected layers incorporate global structural information, ultimately outputting a fixed-dimensional path embedding vector representing the semantic characteristics of the entire path. During training, the corresponding path embedding vector is calculated for each path pair, and the model-predicted overlap value is obtained based on the cosine similarity formula. This predicted value is compared with the true overlap label to construct a loss function. The system minimizes this loss function and performs backpropagation updates of the network parameters using the Adam optimizer or SGD (stochastic gradient descent). After each round of training, the model's mean squared error performance is monitored on a validation set until the loss converges to a preset threshold or there is no significant improvement in validation performance. Through this training process, the convolutional neural network model effectively captures spatial patterns, local repetitive structures, and differences in traffic behavior within path sequences, establishing measurable similarity relationships between paths in a high-dimensional vector space. Compared to traditional path similarity methods based on segment intersection, the Jaccard coefficient, or heuristics, this implementation demonstrates significant adaptability and generalization capabilities. It is independent of rule-based design and can adapt to path representation learning tasks in diverse cities and road network structures.

[0061] In a specific embodiment, path A represents a user starting from "Jinyu Garden, Xicheng District", passing "Zhongshan North Road", and arriving at "People's Square"; path B represents another user starting from the same starting point, passing "Zhongshan North Road" and "Yan'an Middle Road", and finally arriving at "Nanjing West Road". These two paths have significant local overlap characteristics in space, especially in the "Zhongshan North Road" section, while the latter half go to different areas. Based on the proportion of shared sections and historical overlapping traffic records, the system marks these two path pairs as having an actual overlap of 0.65 as part of the training data. Before inputting into the model, the system first converts each path into a path sequence consisting of consecutive road segment numbers, and assigns a trainable embedding vector to each road segment. In addition to the spatial number, the embedding vector also includes dynamic attributes such as travel time, average speed and congestion level. Taking path A as an example, its path sequence might be encoded as [r101, r203, r305], while path B is encoded as [r101, r203, r407, r501]. After transformation through the embedding layer, each path forms a fixed-length segment feature tensor, uniformly padded to the same dimension to accommodate the network structure. These processed path sequences are then fed into a shared-parameter convolutional neural network, which includes multiple one-dimensional convolution kernels of different sizes to extract local segment features at different scales, such as the spatial pattern of a triplet or quintuplet. After the convolution operation, the feature map undergoes max pooling and fully connected layers, outputting a uniformly dimensional path representation vector. The system takes the vectors of path A and path B as input and calculates the cosine similarity between them as the model's predicted output, which measures the degree of spatial overlap between the path pairs. During training, this predicted value is compared with the pre-annotated true overlap labels to construct a mean squared error loss function. For example, when the network initially predicts a degree of overlap of 0.61 between two paths, which differs from the true label of 0.65, the system calculates a loss of (0.61-0.65)² and feeds this error back into the network structure through backpropagation, updating the convolution kernel parameters and embedding matrix. The entire training process is repeated using batch gradient descent, iterating and optimizing samples across multiple paths until the error on the validation set converges to a preset threshold.

[0062] Furthermore, the transfer node splitting module is used to perform the following steps:

[0063] Based on the route overlap data, combined with the time window constraints of the requested order and the road segment nodes, an input tensor including path overlap, node travel time and transfer waiting tolerance is constructed;

[0064] Input the input tensor into the Transformer model, perform node analysis on the request order path, and obtain a probability value vector of each node as a transfer point;

[0065] Based on the probability value vector of each node as a transfer point, a number of high-confidence transfer nodes that meet the preset time constraints and spatial segmentation rationality are screened to generate target transfer node sequence data.

[0066] In some specific embodiments, assume that user A's travel path starts at "Jinyu Garden" and ends at "Donghai Software Park." This complete path consists of several adjacent road segments, traversing multiple urban areas and passing through transportation hubs such as "Zhongshan North Road," "People's Square," and "Yan'an East Road." The system first calculates the overlap between this path and multiple candidate ridesharing orders at different node segments based on the overlap data output by the pre-configured route matching module. Simultaneously, the system retrieves the time window constraint for user A's request (e.g., 7:45–8:20) and, combined with historical data on the average travel time, travel time distribution, and node stop conditions for each road segment, constructs a comprehensive time constraint tensor for each node. Based on this, the system constructs a three-dimensional input tensor for each node in the path. Each column of the tensor represents the corresponding path overlap value, estimated travel time, and transfer wait tolerance. The transfer wait tolerance is statistically modeled based on the average acceptable wait time distribution of passengers at various nodes in historical data. This input tensor is fed into a pre-trained Transformer encoder model for time series modeling. Because the Transformer architecture features a global attention mechanism, it can capture cross-temporal dependencies between path nodes when processing path sequences. For example, it can model the transfer probability of a preceding node in a path in conjunction with the time tolerance characteristics of the following node. This overcomes the lack of structural awareness in global scheduling by traditional sliding window-based or local scoring strategies. The output layer of the Transformer encoder generates a set of feature embeddings for each node, which then pass through a fully connected prediction network to output the probability of each node being a transfer point. For example, the model might identify the transfer probability of the "People's Square" node as 0.87, while that of the "Zhaojiabang Road" node is only 0.21. In practical applications, these probabilities represent the credibility of a node as a transfer point, taking into account a combination of factors such as meeting time constraints, high path overlap, and appropriate transfer tolerance. The system finally filters the predicted probability vector, selects a group of high-confidence nodes that exceed the preset threshold (such as 0.7) in probability and have sufficient path decoupling characteristics in space, such as "People's Square" and "Yan'an Middle Road Subway Exit", and combines them into target transfer node sequence data. This sequence serves as the basic structure for subsequent vehicle matching and chain scheduling, marking that the request order path is formally split into multiple independently schedulable path segments. Compared with the node division method based on rule settings or fixed geographical breakpoints commonly used in the prior art, the Transformer-based node analysis mechanism introduced in this embodiment does not rely on static settings, and has the ability to dynamically learn spatiotemporal patterns within the path and make decisions based on multi-dimensional behavioral characteristics. It can adapt to changes in transfer conditions between different urban road networks and different types of orders. Its core advantage lies in the realization of joint modeling and probabilistic expression of path structure and transfer scenarios from the algorithm level, which improves the accuracy of node splitting and scheduling feasibility.

[0067] Furthermore, the road section nodes include road section intersection points and temporary parking areas.

[0068] It should be noted that the platform, based on high-precision urban road network data, categorizes all candidate transfer nodes along a route based on their geometric structure and traffic function attributes. These nodes primarily fall into two categories: First, segment intersections—intersections, corners, or hubs where multiple road segments are geometrically connected. Examples include crossroads, T-junctions, and elevated ramps. These nodes possess inherent route switching potential and are the most common locations for segmentation. Second, temporary parking areas—spaces where vehicles can board and disembark for short periods without disrupting the main traffic flow—may include bus bays, sunken parking lanes in commercial areas, slow lanes at residential entrances, and designated smart carpooling spots. During the route splitting modeling process, the system prioritizes these structural nodes in the transfer candidate set and assigns a transfer weight to each node based on factors such as boarding and disembarking density, road grade, and temporary parking regulations as measured by historical dispatch data. Furthermore, a peak-period capacity assessment model can be incorporated to determine whether these nodes have the necessary traffic environment for transfers during specific periods. For example, when the traffic flow at a certain intersection point is saturated between 7:00 and 9:00 in the morning and there are no safe stopping conditions, even if the spatial geometry meets the transfer characteristics, it will be dynamically eliminated from the candidate set by the model.

[0069] Furthermore, the transfer waiting tolerance is calculated and generated by statistical results of actual waiting time of passengers at each node in historical orders.

[0070] Specifically, the system first analyzes all route segments with transfers in historical ridesharing orders, extracting the boarding and alighting timestamps corresponding to each transfer node. The system then calculates the time difference between the time a passenger disembarks from the previous connecting vehicle and the time a passenger boards the next connecting vehicle at each transfer node, representing the actual wait time. The system then aggregates this time difference by node location, constructing a statistical distribution of wait times for each node, including its mean, variance, median, and quantiles (e.g., 90% confidence interval). During the transfer node screening process, the system uses this statistical distribution to generate a desired transfer wait tolerance for each candidate node. For example, if 90% of the transfer wait times for the "People's Square East Exit" node in historical orders are less than 6 minutes, the system can set the transfer tolerance for that node to 6 minutes, which serves as a key reference parameter for subsequent time window constraint calculations. Combined with the overall time budget of the requested order, the system can further determine whether the node has the scheduling flexibility to facilitate smooth transfers within that timeframe. In this way, the transfer waiting tolerance can be adaptively adjusted according to the transfer pressure, scheduling efficiency and environmental stability of the node itself in the historical operation process, significantly improving the feasibility of path splitting and multi-segment scheduling strategies and the ability to ensure passenger experience.

[0071] Furthermore, the construction of the Transformer model includes the following steps:

[0072] Constructing a path node sequence sample based on historical carpooling order data, wherein the path node sequence sample includes the path node sequence corresponding to each order, the time window constraint, the actual transfer node location, and the transfer success label;

[0073] Encoding the path node sequence, using position coding to mark the relative order of the nodes in the path, and constructing a node feature vector sequence based on the path overlap, travel time and transfer feasibility of each node;

[0074] The node feature vector sequence is input into the Transformer encoder to capture the interaction between nodes in terms of time constraints, overlap, and traffic efficiency, and output each node as a representation vector of the transfer node;

[0075] The representation vector is input into the fully connected prediction layer to output the probability value vector of the transfer point and compared with the historical transfer labels, and the model parameters are optimized through the cross entropy loss function.

[0076] In some embodiments, the system first parses the complete path of each order, extracts the node sequence of each route segment, and records the time window constraint information corresponding to each path. Furthermore, by retrospectively analyzing the actual order scheduling process, the system further annotates whether transfers occur along the path, as well as the node location and successful completion of the transfer, thereby constructing a sample set of path node sequences with supervisory labels. During the sample representation phase, the system encodes each path node sequence. To preserve the order of nodes in the path, a position encoding mechanism is introduced, attaching the relative position of each node in the sequence. Furthermore, combining data on the overlap between upstream and downstream ridesharing order paths, average travel time at nodes, and transfer waiting tolerance, the system constructs a multi-dimensional node feature vector. Each node's feature vector not only captures its own static properties but also incorporates its spatiotemporal contextual information within the entire path, forming a well-structured node feature tensor. After the node feature vector sequence is input into the Transformer encoder, the model uses a multi-head self-attention mechanism to model the interactions between all nodes in the path. Unlike traditional convolutional architectures, the Transformer is not limited to local adjacency within a fixed window size. Instead, it can identify distant but logically cohesive node pairs within a path based on global dependencies. This allows it to capture complex transfer conditions, such as high overlap at the beginning of a path but requiring delayed fulfillment at the end. The attention mechanism assigns dynamic attention weights to each node based on features such as time window overlap, travel time trends, and historical transfer success rates, making the final node embedding highly temporally and spatially aware. The output node representation vector is then fed into a fully connected prediction network, which uses a sigmoid activation function to output the probability of each node being a transfer point. During the training phase, the system compares the model's predictions with historically annotated transfer labels, constructing a cross-entropy loss function as the optimization objective. By minimizing the information discrepancy between the predicted probabilities and the true transfer labels, the model gradually optimizes the attention weights and representation layer parameters over multiple iterations, achieving refined scoring of transfer nodes. Compared to traditional rule-based node scoring methods, this embodiment introduces global modeling capabilities through the Transformer architecture, enabling the discovery of potential transfer nodes within a path based on multi-feature fusion. It not only improves the rationality of route splitting, but also enhances the flexibility of multi-segment carpooling scheduling.

[0077] Furthermore, the reinforcement learning model is trained by the following steps:

[0078] Constructing a state space based on historical dispatch records, which includes the target transfer node sequence, the time window parameters of the requested order, the real-time location and empty status of candidate vehicles, the predicted time when the vehicle can reach the transfer node, and the travel time between transfer nodes;

[0079] Based on the state space, determining an action space, the action space comprising selecting a candidate vehicle from a set of candidate vehicles for each transfer node;

[0080] Constructing a reward function based on the idle distance of vehicles in chain scheduling and the difference between the total travel time and the direct travel time of passengers;

[0081] Based on the state space, action space and reward function, a deep Q network algorithm is used for training, which takes the state and action as input, outputs the corresponding Q value, and updates the network parameters by minimizing the mean square error between the current Q value and the target Q value.

[0082] It should be noted that during the training data construction phase, the system extracts a large number of real-world dispatch samples based on historical dispatch records, each corresponding to a split carpool route. For each route, the system constructs a corresponding state space. State variables include: the time window parameters of the current request; the spatial location, estimated arrival time, and remaining time window of each node in the transfer node sequence; the real-time geographic location, empty status, and current mission status of all candidate vehicles on the platform; and the predicted travel time required for each candidate vehicle to reach each transfer node from its current location. Furthermore, the system calculates the travel time between adjacent transfer nodes to fully construct a high-dimensional state vector encompassing time, space, and capacity states, which is used as input to the reinforcement learning model. Within this state space, the system defines an action space as a set of dispatch selection actions, each representing the selection of a candidate vehicle for a connection at a specific transfer node. The size of the action set depends on the combination of the number of candidate nodes and the number of vehicles. During training, the model extracts an action from the current state at each time step, sequentially selecting the most suitable vehicle for each transfer node to achieve the optimal overall route combination. To guide policy optimization, the system designed a reward function tightly coupled to the dispatching objective. This function comprehensively considers two key metrics: the idle distance caused by dispatched vehicles that are not part of the original order but are scheduled for transfers; and the difference between the total passenger travel time on the current dispatched route and their theoretical minimum direct travel time. The system constructs a negatively correlated reward term based on this: the fewer idle trips and the lower the additional passenger travel time, the higher the reward, ensuring that the model strategy strikes a balance between matching success rate and passenger experience. Based on the aforementioned state space, action space, and reward function, the system employs a deep Q-network algorithm for model training. The network inputs a state-action pair, which is mapped through a multi-layer neural network and outputs the Q-value of the action, which is an estimate of the long-term benefit of taking that action under the current policy. During training, the model continuously constructs a mean squared error loss function based on the difference between the current Q-value and the target Q-value. The network parameters are iteratively optimized using backpropagation and gradient descent. The target Q-value is estimated based on the maximum Q-value of the next state, and policy improvement is achieved through temporal difference learning. After multiple rounds of training iterations, the reinforcement learning model can quickly make globally optimal or suboptimal vehicle allocation decisions when faced with new orders, changing vehicle locations and node status, and build an efficient travel route composed of multiple vehicles.

[0083] Furthermore, the reward function is as follows:

[0084] R=R max -(α·(T chain -T direct )+β·D idle );

[0085] Among them, R is the reward value; R max is the initial excitation value; T direct The theoretical direct access time; T chain The total time required to complete the entire trip by matching the vehicle solution corresponding to the requested order; D idle is the cumulative empty driving distance of the vehicle before executing the task in the vehicle matching scheme; α is the time difference coefficient; β is the empty driving distance coefficient.

[0086] It should be noted that the base reward value represents the initial incentive provided by the system when the dispatch plan is ideal. Ideal means that passengers require little to no waiting or transfers, and the matched vehicles complete the connection without detours or empty runs. Based on this, the system compares the actual total travel time for each dispatched route with the theoretically shortest direct travel time for that route to calculate the additional time. The larger the difference in time, the more transfers and waiting involved in the dispatched route, and the more significant the impact on the passenger experience. Therefore, the system multiplies the overall reward value by a certain time penalty factor. The system also accumulates the idle distance generated by all vehicles involved in the dispatch. This idle distance reflects the additional resources consumed by the platform to facilitate carpooling connections. If a chained dispatch route requires multiple vehicles to travel long distances to the connection point, this not only increases dispatch complexity but also consumes additional transportation resources. Therefore, the system further reduces the reward value by multiplying the idle distance by a idle penalty factor. This reward function not only considers the total time cost from a passenger's request to their final arrival, but also measures the platform's transportation costs to complete the rideshare. This ensures that the model automatically learns scheduling strategies that balance efficiency and economy during training. Compared to existing technologies, the reward function employed in this invention has dual optimization objectives, improving passenger satisfaction while effectively preventing frequent empty rides and inefficient scheduling within the platform system.

[0087] Furthermore, the carpooling scheduling module is used to perform the following steps:

[0088] Based on the vehicle matching data, the vehicle number, connection sequence and estimated arrival time of each connecting route are extracted, and the route travel time and traffic status information are combined to generate a dispatch instruction including the vehicle number, starting and ending node locations, and task time period;

[0089] The dispatch instruction is sent to the corresponding vehicle driver terminal.

[0090] In this embodiment, the route scheduling information for each segment contained in the vehicle matching data is first parsed. The vehicle number for each connecting task is extracted, along with the order of each vehicle's route segment within the entire carpool route, the starting and ending node locations, and the predicted arrival and departure times. The system then dynamically adjusts the actual travel time for each route segment by combining real-time urban traffic status data with a historical travel duration database. After integrating travel times, vehicle status, and node locations, the system generates a separate scheduling task for each connecting vehicle. The task includes information such as the vehicle number, the corresponding starting and destination nodes, the scheduling time window, and the task priority. These tasks are distributed in real time to the driver's terminal of the relevant vehicle via the onboard terminal management system. Drivers can view their current task's pickup location, estimated departure and arrival times, and any subsequent transfer connections through the terminal interface. If a vehicle's connecting task cannot be completed on time due to traffic disturbances or emergencies, the system automatically triggers a rescheduling mechanism to dynamically adjust the subsequent connection order or reselect the connecting vehicle to ensure an uninterrupted carpool route. This module realizes a complete closed loop of chain carpooling tasks from algorithm decision-making to actual execution, which not only improves the automation level of scheduling implementation, but also realizes the timing guarantee of task flow in the complex carpooling process with multiple vehicles and multiple sections.

[0091] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. An intelligent carpooling management system based on deep learning, characterized in that: include: The route matching module is used to calculate the similarity between the driving paths of the requested order and the selected carpooling order based on the starting and ending point information of the requested order and the real-time road network topology, and obtain the route overlap data; A transfer node splitting module is used to split the request order based on the route overlap data and the time window constraints of each order to obtain target transfer node sequence data; A vehicle matching module is used to match the target transfer node sequence data with the vehicle real-time positioning data based on a reinforcement learning model to obtain vehicle matching data; The carpooling dispatch module outputs dispatch instructions based on vehicle matching data and sends them to the corresponding vehicle driver terminal.

2. The intelligent carpooling management system based on deep learning according to claim 1 is characterized in that: The route matching module is used to perform the following steps: Based on the starting point, end point, and historical trajectory samples of the requested order, a pre-trained convolutional neural network is used to extract features from the candidate driving paths of the requested order to obtain a path feature vector of the requested path. Based on the historical driving path data of the carpooling orders to be selected, the convolutional neural network is used to perform feature extraction on each candidate carpooling path to obtain a path feature vector for each carpooling order; A cosine similarity calculation is performed on the path feature vector of the requested path and the path feature vectors of each carpooling order to obtain route overlap data between the requested order and each carpooling order to be selected.

3. The intelligent carpooling management system based on deep learning according to claim 2 is characterized in that: The training steps of the convolutional neural network are as follows: Based on a path dataset containing historical order trajectories, we extract the path sequence and travel time corresponding to each order, construct path sequence sample pairs, and annotate the actual overlap between each pair of paths. Convert the path sequence into a fixed-length segment embedding vector sequence, input the sequence into a convolutional neural network for feature extraction, and obtain a path representation vector; Calculate the cosine similarity between the path representation vectors as the output value of the network, and calculate the loss value between the output value of the network and the actual coincidence label using the mean square error loss function; With the goal of minimizing the loss value, the batch gradient descent method is used to iteratively train the parameters of the convolutional neural network until convergence reaches the preset threshold.

4. The intelligent carpooling management system based on deep learning according to claim 1 is characterized in that: The transfer node splitting module is used to perform the following steps: Based on the route overlap data, combined with the time window constraints of the requested order and the road segment nodes, an input tensor including path overlap, node travel time and transfer waiting tolerance is constructed; Input the input tensor into the Transformer model, perform node analysis on the request order path, and obtain a probability value vector of each node as a transfer point; Based on the probability value vector of each node as a transfer point, a number of high-confidence transfer nodes that meet the preset time constraints and spatial segmentation rationality are screened to generate target transfer node sequence data.

5. The deep learning-based intelligent carpooling management system according to claim 4 is characterized in that: The road section nodes include road section intersection points and temporary parking areas.

6. The deep learning-based intelligent carpooling management system according to claim 4, characterized in that: The transfer waiting tolerance is calculated and generated by statistical results of actual waiting time of passengers at each node in historical orders.

7. The intelligent carpooling management system based on deep learning according to claim 4 is characterized in that: The construction of the Transformer model includes the following steps: Constructing a path node sequence sample based on historical carpooling order data, wherein the path node sequence sample includes the path node sequence corresponding to each order, the time window constraint, the actual transfer node location, and the transfer success label; Encoding the path node sequence, using position coding to mark the relative order of the nodes in the path, and constructing a node feature vector sequence based on the path overlap, travel time and transfer feasibility of each node; The node feature vector sequence is input into the Transformer encoder to capture the interaction between nodes in terms of time constraints, overlap, and traffic efficiency, and output each node as a representation vector of the transfer node; The representation vector is input into the fully connected prediction layer to output the probability value vector of the transfer point and compared with the historical transfer labels, and the model parameters are optimized through the cross entropy loss function.

8. The intelligent carpooling management system based on deep learning according to claim 1 is characterized in that: The reinforcement learning model is trained by the following steps: Constructing a state space based on historical dispatch records, which includes the target transfer node sequence, the time window parameters of the requested order, the real-time location and empty status of candidate vehicles, the predicted time when the vehicle can reach the transfer node, and the travel time between transfer nodes; Based on the state space, determining an action space, the action space comprising selecting a candidate vehicle from a set of candidate vehicles for each transfer node; Constructing a reward function based on the idle distance of vehicles in chain scheduling and the difference between the total travel time and the direct travel time of passengers; Based on the state space, action space and reward function, a deep Q network algorithm is used for training, which takes the state and action as input, outputs the corresponding Q value, and updates the network parameters by minimizing the mean square error between the current Q value and the target Q value.

9. The deep learning-based intelligent carpooling management system according to claim 8, characterized in that: The reward function is as follows: R=R max -(a·(T chain -T direct )+β·D idle ); Among them, R is the reward value; R max is the initial excitation value; T direct The theoretical direct access time; T chain The total time required to complete the entire trip by matching the vehicle solution corresponding to the requested order; D idle is the cumulative empty driving distance of the vehicle before executing the task in the vehicle matching scheme; α is the time difference coefficient; β is the empty driving distance coefficient.

10. The intelligent carpooling management system based on deep learning according to claim 1, characterized in that: The carpooling scheduling module is used to perform the following steps: Based on the vehicle matching data, the vehicle number, connection sequence and estimated arrival time of each connecting route are extracted, and the route travel time and traffic status information are combined to generate a dispatch instruction including the vehicle number, starting and ending node locations, and task time period; The dispatch instruction is sent to the corresponding vehicle driver terminal.

Images