A travel trajectory generation, deep learning network model construction and training method

Abstract

The application discloses a kind of travel trajectory generation, deep learning network model and training method, the method includes: constructing deep learning network, obtaining road network and according to the adjacency relation of each road section in the road network, construct adjacency matrix, the adjacency matrix is embedded in deep learning network;Through training data, the embedded deep learning network is trained;Get departure, destination and current location;The departure, destination and current location are input into the well-trained deep learning network to generate travel trajectory;The adjacency relation of each road section is embedded in deep learning network in the form of adjacency matrix, and the discontinuity problem of trajectory caused by non-adjacent road section is avoided.

Description

Technical Field

[0001] This invention relates to the field of big data and artificial intelligence technology, and more specifically to a method for generating travel trajectories, constructing and training deep learning network models. Background Technology

[0002] Trajectory generation research can support many downstream applications in the field of smart cities, such as urban infrastructure planning, trajectory data privacy protection, and epidemic spread analysis.

[0003] To address the trajectory generation problem, researchers have employed generative deep neural network models, primarily using neural network generative model paradigms such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs). Unlike methods based on hand-designed mechanisms, deep neural network methods abandon the construction of specific individual travel mechanisms. Instead, they directly utilize neural networks to learn the true latent distribution of trajectory data and generate synthetic trajectories from the same distribution. Thanks to the depth and self-learning capabilities of deep neural networks, generative model paradigms like GANs and VAEs can simultaneously capture different aspects of individual travel patterns. Furthermore, deep neural network models can learn high-dimensional and nonlinear patterns that are difficult to analyze and construct manually using traditional methods. Therefore, deep neural network-based models can generate more realistic trajectory data than traditional methods based on hand-designed mechanisms.

[0004] However, existing generative models based on deep neural networks still have the following three unresolved problems: (1) First, existing generative models ignore the continuity of generated trajectories. The trajectories generated by current methods are not continuous trajectories on the road network, which makes these synthetic trajectories unusable for downstream applications such as traffic simulation. (2) Second, existing generative models do not make full use of knowledge of crowd movement patterns, resulting in poor model training efficiency and performance. (3) Third, the random generation process of existing methods suffers from error accumulation. That is, in existing methods, trajectories are randomly generated according to the probability given by the generator. However, once the generator makes a wrong prediction, the process will continue to generate under the premise of error, thereby reducing the quality of generated trajectories.

[0005] Therefore, how to provide a method for generating travel trajectories, constructing deep learning network models, and training them is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method for generating travel trajectories, constructing and training deep learning network models, which embeds the adjacency relationships of each road segment into the deep learning network in the form of an adjacency matrix, thereby avoiding the problem of trajectory discontinuity caused by non-adjacent road segments.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for generating travel trajectories, characterized by the following steps:

[0009] Building deep learning networks

[0010] Obtain the road network and construct an adjacency matrix based on the adjacency relationship of each road segment in the road network, then embed the adjacency matrix into a deep learning network;

[0011] The embedded deep learning network is trained using training data;

[0012] Get the origin, destination, and current location;

[0013] The origin, destination, and current location are input into the trained deep learning network to generate a travel trajectory.

[0014] Furthermore, the origin, destination, and current location are input into the trained deep learning network to generate a travel trajectory, including:

[0015] The generated trajectory and candidate road segment set are determined based on the origin, destination, and current location;

[0016] Calculate the observation condition probability between each candidate road in the candidate road segment set and the generated trajectory, and use the adjacency matrix to set the observation condition probability of non-adjacent candidate roads to 0, and then generate the observation cost.

[0017] Calculate the expected cost based on the candidate route set and the destination;

[0018] Calculate the travel cost and generate a route trajectory based on the travel cost until the destination is reached to obtain the travel trajectory.

[0019] Furthermore, calculating the expected cost based on the candidate road segment set and destination includes:

[0020] Obtain the relative positions and distances of each road segment from the road network, and calculate the spherical distance between two road segments;

[0021] After inputting the candidate road segment set and the destination into the trained graph attention network to output the spherical distance, the expected conditional probability is predicted through a multilayer perceptron network, and the expected cost is calculated.

[0022] A method for constructing a deep learning network model of travel trajectories, comprising an input layer, an LSTM network layer, a constrained linear layer, and an output layer;

[0023] The input layer is used to receive training data;

[0024] The LSTM network layer is used to extract deep mobility situation features;

[0025] A constrained linear layer is used to make decisions based on the characteristics of the deep mobile station, and an adjacency matrix is ​​introduced as a constraint on the decision.

[0026] The output layer is used to generate travel trajectories based on the decision results.

[0027] Furthermore, it also includes an expected cost calculation module, used to calculate the expected cost based on the candidate road segment set and the destination;

[0028] The training data includes real trajectories, a portion of historical movement trajectories extracted from the real trajectories, and a set of candidate road segments corresponding to the historical movement trajectories;

[0029] The decision-making process specifically involves: calculating the observation conditional probability between each neighboring candidate road in the candidate road segment set and the historical movement trajectory, and then using the adjacency matrix to set the observation conditional probability of non-adjacent candidate roads to 0, thereby generating the observation cost.

[0030] The output layer is also used to calculate the travel cost based on the expected cost and the observed cost, and to generate a route trajectory based on the travel cost until the destination is reached to obtain the travel trajectory.

[0031] Furthermore, an attention network layer is introduced between the LSTM network layer and the constrained linear layer;

[0032] The correlation weights between the candidate road segment set and the historical movement trajectory are calculated using an attention mechanism;

[0033] The constrained linear layer incorporates the correlation weights when calculating conditional probabilities.

[0034] Furthermore, the expected cost calculation module includes a graph attention network layer, a calculation module, and a multilayer perceptron;

[0035] The graph attention network layer is used to learn road features and generate a road representation matrix;

[0036] The calculation module is used to index the road representation matrix according to the candidate road segment set and the destination to obtain the road representation vector of each road segment in the candidate road segment set and the road representation vector of the destination; and to calculate the spherical distance according to the road representation vector of each road segment in the candidate road segment set and the road representation vector of the destination.

[0037] Based on the road representation matrix, the relative position and relative distance between the candidate road segment set and the destination are determined, and the spherical distance is calculated.

[0038] The multilayer perceptron is used to calculate the expected cost based on the road representation vectors of each road segment in the candidate road segment set, the road representation vector of the destination, and the spherical distance.

[0039] A method for training a deep learning network model, employing the deep learning network model described in claims 4-7, includes the following steps:

[0040] A generator is constructed based on the deep learning model;

[0041] Acquire training data, which includes real trajectories composed of multiple road segment trajectories; wherein the road segment trajectories include spatial and temporal information of the trajectories;

[0042] The origin and destination of the multiple road segments are input into the generator to generate a composite travel trajectory.

[0043] The spatial and temporal information corresponding to the road segments in the actual trajectory and the travel trajectory are processed to obtain the trajectory spatial sequence and time sequence.

[0044] The trajectory spatial sequence and the time sequence are respectively input into the trained discriminator to obtain a reward signal; wherein, the reward signal includes a sequential reward and a mobility yaw reward;

[0045] The generator gradient is calculated based on the reward signal.

[0046] Furthermore, the discriminator includes a preprocessing module, a sequence order extraction module, a sequence reward module, and a mobility yaw reward module;

[0047] The preprocessing module is a generator for historical training rounds, used to supplement the current travel trajectory into a complete travel trajectory;

[0048] The sequence order extraction module receives the conversion mode of the extracted time series of the complete travel trajectory and predicts the true probability of the trajectory.

[0049] The sequential reward module calculates the sequential reward based on the true probability;

[0050] The mobility deviation reward module calculates the mobility deviation reward based on the actual trajectory label and the complete travel trajectory.

[0051] The beneficial effects of this invention are:

[0052] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a travel trajectory generation, a deep learning network model and a training method, which embeds the adjacency relationship of each road segment into the deep learning network in the form of an adjacency matrix, thereby avoiding the trajectory discontinuity problem caused by non-adjacent road segments. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0054] Figure 1 The attached figure is a schematic diagram of a travel trajectory generation, deep learning network model, and training method provided by the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see the appendix Figure 1 .

[0057] Example 1

[0058] This invention discloses a method for generating travel trajectories, comprising the following steps:

[0059] S1: Obtain the road network and construct an adjacency matrix based on the adjacency relationships of each road segment in the road network. Embed the adjacency matrix into a pre-constructed deep learning network. The road network is prior knowledge of a certain region and can confirm the adjacency relationships between roads.

[0060] S2: Train the embedded deep learning network using training data;

[0061] S3: Get the origin, destination, and current location;

[0062] S4: Input the origin, destination and current location into the trained deep learning network to generate the travel trajectory.

[0063] This invention uses the assumptions of the A* algorithm to model human movement patterns, thereby predicting individual movement strategies. The A* algorithm is a heuristic search algorithm widely used in road networks. In the A* assumption, individual movement behavior is determined by two factors: the observed travel cost from the starting road to the current candidate road, and the expected travel cost from the candidate road to the destination. Combining these two costs, the A* algorithm evaluates which candidate road is the best candidate road for the next search step, and then heuristically generates the optimal continuous trajectory. Specifically, the A* algorithm for candidate road segments... The evaluation formula is: ,in Indicates the observed passage costs, This indicates the expected toll cost.

[0064] In one embodiment, S4 specifically includes:

[0065] S41: Determine the generated trajectory and candidate road segment set based on the origin, destination and current location; if the current location is not the origin, determine the generated trajectory and execute S42; if the current location is the origin, execute S43 directly.

[0066] Here, the generated movement trajectory represents the trajectory that an individual has already traversed, and the candidate road segment trajectory set consists of the next possible road segment based on the current location and the destination direction, composed of multiple candidate roads; the trajectory can be defined as a sequence ordered by time. Among them, trajectory points Defined as a tuple . Indicates the road segment ID. This represents the time information corresponding to the trajectory points. A continuous trajectory is a trajectory that satisfies the following condition: each trajectory point pair... In a road network, all road segments are adjacent. Therefore, the trajectory generation problem can be described as: given a real-world dataset of movement trajectories, generate one or a set of continuous movement trajectories. .

[0067] S42: Calculate the observation conditional probability between each candidate road in the candidate road segment set and the generated trajectory, and use the adjacency matrix to set the observation conditional probability of non-adjacent candidate roads to 0, and then generate the observation cost.

[0068] In another embodiment, a deep learning network can be constructed to calculate the observation cost, and the specific steps include:

[0069] Build training data, obtain complete trajectory data, extract multiple trajectory points, extract a portion of trajectory data as the generated trajectory, and use the next segment of the trajectory as the label of the candidate road segment trajectory set;

[0070] Training process:

[0071] For the already generated trajectory We do this by each point Embedded into dense vectors To extract spatiotemporal information, as shown below:

[0072] For encoding time information: we map the timestamps of weekdays to a set of... Time slots (1 to 1440) are used to encode the timestamps, and weekend timestamps are mapped to another set of time slots (1441 to 2880). .

[0073] After embedding, we transform the current partial trajectory into a series of dense vectors. Then, we use an LSTM network to analyze the dense vectors and model the state of each trajectory point to obtain a spatiotemporal trajectory model. ;

[0074] Using dot product attention mechanism to enhance latent movement states As shown in the formula below:

[0075] Where the function () Calculate the current movement state hi and the historical movement state using the dot product method. The correlation weights between them.

[0076] Finally, after capturing the enhanced mobility state, we construct a topologically constrained linear layer to predict the observed conditional probabilities. In the topologically constrained linear layer, we introduce a topological adjacency weight. This weighting coefficient is used when candidate roads... With current roads The value is 1 if the paths are adjacent, and 0 otherwise. By using topological adjacency weights, we can identify non-adjacent candidate roads. conditional probability Set to zero to ensure the continuity of the generated trajectory. The observed conditional probability is expressed as...

[0077]

[0078] Wherein, parameter vector It is the first of the linear layer weight matrix. Okay, the weight matrix of the linear layer is the trainable parameters of the model. The function of the linear layer is expressed as y = w*x + b, where w is the weight matrix, which is updated through gradients during training. Ultimately, the observed cost... It can be calculated as the negative logarithm of the conditional probability, i.e. .

[0079] S43: Calculate the expected cost based on the candidate road segment set and the destination;

[0080] Specifically, the process involves: obtaining the relative positions and distances of each road segment from the road network, calculating the spherical distance between two road segments; inputting the candidate road segment trajectory set and the destination into the trained graph attention network to output the spherical distance, predicting the expected conditional probability through a multilayer perceptron network, and calculating the expected cost.

[0081] In another embodiment, a deep learning network model can be used for computation, and the specific steps include:

[0082] A graph attention network (GAT) is constructed to learn a structured road representation that includes relative positional information. The update of the graph attention network can be represented as... .in, It is the road representation matrix of the z-th iteration, the z-th iteration. OK Corresponding road segment The representation vector, It is a road adjacency matrix.

[0083] During the network initialization phase, we construct the road context embedding vector. and set Specifically, given a road segment ,

[0084] We consider six road attributes: segment length, segment width, maximum speed limit, number of lanes, road type, and latitude and longitude. Next, we normalize the continuous attributes and perform one-hot encoding on the discrete attributes. We then concatenate the one-hot encoded discrete attributes with the normalized continuous attributes to form a context embedding vector.

[0085] After obtaining road representations, we use a multilayer perceptron network to predict the road characteristics from candidate roads. Arrive at the destination The expected conditional probability is given by the following formula:

[0086]

[0087] in and It is a road and The road representation vector, It is the spherical distance between them. Ultimately, the expected cost. Calculated as conditional probability The negative logarithm, i.e. .

[0088] S44: Calculate the travel cost and generate a route trajectory based on the travel cost until the destination is reached to obtain the travel trajectory.

[0089] In another embodiment, the destination includes intermediate destinations across multiple stages.

[0090] The specific steps include:

[0091] In the first phase, our goal is to generate regional trajectories.

[0092] The road network is divided into multiple structural regions, and the mapping relationship between each structural region and the road segments within it is obtained.

[0093] The trajectory start and end points are mapped to the regional level according to the mapping relationship to obtain regional start and regional end points; initially, we sample start and end point tuples from the historical OD matrix. and start timestamp This can be calculated from a given real-world trajectory dataset. Then, we map road-level start and end points to region-level start and end points based on the road-to-region mapping. Among them we will Mapped to If and only if .

[0094] The regional starting point and regional ending point are input into the trained deep learning network to generate regional trajectories. Each structural region in the regional trajectory has a starting point and an ending point. The ending point is only the end point of the regional trajectory, so it can be regarded as an intermediate destination directly from the starting point to the destination. The deep learning network is trained using regional trajectories.

[0095] In the second stage, we generate continuous trajectories guided by regional trajectories. In practice, we construct each region... and region Boundary road segment set The access frequency of each boundary road segment is pre-calculated. Then, the regional trajectory is mapped to the road level according to the algorithm process shown in the attached figure. This results in discontinuous road segment trajectories. Then, we use a road-level generator to complete the process. Each pair The trajectory between these points is the final continuous trajectory. This way, we only need to generate it within a small, structured region, instead of generating it within a vast road network.

[0096] Example 2

[0097] This invention discloses a method for constructing a deep learning network model for travel trajectories. The model adopts the trajectory generation method described in the embodiments, including an input layer, an LSTM network layer, a constrained linear layer, and an output layer.

[0098] The input layer is used to receive training data or actual travel data, including the origin, destination and current location.

[0099] LSTM network layers are used to extract deep motion situation features.

[0100] Decisions are made based on the characteristics of deep mobile stations, and an adjacency matrix is ​​introduced as a constraint on the decision.

[0101] A constrained linear layer is used to make decisions based on the characteristics of the deep mobile station, and an adjacency matrix is ​​introduced as a constraint on the decision.

[0102] The output layer is used to generate travel trajectories based on the decision results.

[0103] In one embodiment, the system further includes an expected cost calculation module for calculating the expected cost based on the candidate road segment set and the destination.

[0104] The training data includes real trajectories, partial historical movement trajectories extracted from real trajectories, and a set of candidate road segments corresponding to the historical movement trajectories;

[0105] The decision-making process is as follows: calculate the observation conditional probability between each neighboring candidate road in the candidate road segment set and the historical movement trajectory, and use the adjacency matrix to set the observation conditional probability of non-adjacent candidate roads to 0, and then generate the observation cost.

[0106] The output layer is also used to calculate travel costs based on expected and observed costs, and to generate route trajectories based on travel costs until the destination is reached, thus obtaining the travel trajectory.

[0107] In one embodiment, an attention network layer is introduced between the LSTM network layer and the constrained linear layer;

[0108] The correlation weights between the candidate road segment trajectory set and the historical movement trajectory are calculated using an attention mechanism;

[0109] When calculating conditional probabilities, the constrained linear layer introduces correlation weights.

[0110] In one embodiment, the expected cost calculation module includes a graph attention network layer, a calculation module, and a multilayer perceptron. The graph attention network layer is used to learn road features and generate a road representation matrix. The calculation module is used to index the road representation matrix according to the candidate road segment set and the destination to obtain the road representation vector of each road segment in the candidate road segment set and the road representation vector of the destination. It is also used to calculate a spherical distance based on the road representation vectors of each road segment in the candidate road segment set and the road representation vector of the destination, wherein the spherical distance is calculated based on the latitude and longitude information of each road contained in the road network information. The module is used to confirm the relative position and relative distance between the candidate road segment set and the destination according to the road representation matrix and calculate the spherical distance. The multilayer perceptron is used to calculate the expected cost based on the road representation vectors of each road segment in the candidate road segment set, the road representation vector of the destination, and the spherical distance.

[0111] Example 3

[0112] This invention provides a method for training a deep learning network model, which can be used to train the deep learning network model in Embodiment 2. The steps include:

[0113] Build a generator based on a deep learning model;

[0114] Acquire training data, which includes real trajectories composed of multiple road segment trajectories; whereby the road segment trajectories include spatial and temporal information of the trajectories;

[0115] Input the origin and destination of multiple road segments into the generator to generate a composite travel trajectory;

[0116] The spatial and temporal information corresponding to road segments in the actual trajectory and the travel trajectory are processed to obtain the trajectory spatial sequence and time sequence;

[0117] The trajectory spatial sequence and time series are respectively input into the trained discriminator to obtain the reward signal; the reward signal includes sequential reward and mobility yaw reward.

[0118] Calculate the generator gradient based on the reward signal.

[0119] Generally, Generative Adversarial Networks (GANs) enhance the performance of the generator by training a mini-maximum game between the generator and the discriminator, as shown below.

[0120]

[0121] in This represents a sample from the real data distribution. This represents the samples generated by the generator. yes - Parameterized discriminator. In the above minimax game, the discriminator... The goal is to distinguish input samples, generator It then performs self-optimization based on the discriminator's output signal. We call this signal the reward signal. .

[0122] In one embodiment, the discriminator includes a preprocessing module, a sequence order extraction module, a sequence reward module, and a mobility yaw reward module;

[0123] The preprocessing module is a generator for historical training rounds, used to process the current travel trajectory and supplement it into a complete travel trajectory;

[0124] Specifically, because the generator's trajectory is generated sequentially step by step, a discriminator is needed to provide a reward signal for each step in the trajectory. However, for intermediate steps... Current trajectory It is not yet fully generated. This means that based on the current sub-trajectory... The reward only considers the sequence transformation pattern of the current sub-trajectory, ignoring future results. Therefore, we apply Monte Carlo search to evaluate intermediate steps. To do this, we first maintain an older version of the generator. This is the version of the current generator from the previous training epoch. Then, we use... The current sub-trajectory is completed by repeating the Monte Carlo search N times.

[0125] The sequence order extraction module receives the transformation patterns of the time series of the complete travel trajectory and predicts the true probability of the trajectory. The sequence order extraction module includes an embedding layer, an LSTM network layer and a linear layer. The completed trajectory is input to the embedding layer, the LSTM network layer extracts the hidden sequence transformation patterns in the trajectory to evaluate the similarity, and the linear layer predicts the probability that the input trajectory is true.

[0126] The sequential reward module calculates sequential rewards based on true probabilities; the specific steps include using Monte Carlo search to evaluate the average continuous reward of intermediate steps in the trajectory, as shown in the following formula.

[0127]

[0128] The mobility deviation reward module calculates mobility deviation rewards based on the actual trajectory label and the complete travel trajectory.

[0129] Specifically, input trajectory Mobility yaw distance is defined as the distance to the true trajectory set. The minimum distance is given by the following formula.

[0130]

[0131] Specifically, we utilize the widely used trajectory distance metric DTW to measure mobility yaw distance in the real trajectory set. Includes input trajectory The true trajectories share the same original OD. Furthermore, we also use Monte Carlo search for the trajectories. Each step in the calculation calculates the mobility yaw distance. The mobility yaw reward is defined as the change in yaw distance. ,Right now Furthermore, we applied normalization to measure the dimensionless mobility yaw distance variation.

[0132] Yaw distance The change in the preference distance indicates how much the current step's yaw distance has decreased compared to the previous step. Therefore, the greater the decrease in preference distance, the more correct the decision in this step is. Correspondingly, the reward for the current step should be greater.

[0133] In this embodiment, at the start of training, we pre-train using the trajectory next position prediction task. functions and Functions are used because they learn individual movement patterns from different perspectives. Then, we pre-train a discriminator based on the generated and real trajectories. After all pre-training is complete, we compute gradients for the generator using the REINFORCE reinforcement learning algorithm, as shown below.

[0134]

[0135] in, Represents generator The parameters, It is the generated trajectory, reward It is a sequential reward from the discriminator Mobility yaw reward The sum. Based on policy gradient. ,parameter In the following manner: Update, in which It is the learning rate.

[0136] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0137] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating travel trajectories, characterized in that, Includes the following steps: Obtain the road network and construct an adjacency matrix based on the adjacency relationship of each road segment in the road network, then embed the adjacency matrix into a pre-constructed deep learning network; The embedded deep learning network is trained using training data; Get the origin, destination, and current location; The origin, destination, and current location are input into the trained deep learning network to generate a travel trajectory, specifically including: The generated trajectory and candidate road segment set are determined based on the origin, destination, and current location; Calculate the observation condition probability between each candidate road in the candidate road segment set and the generated trajectory, and use the adjacency matrix to set the observation condition probability of non-adjacent candidate roads to 0, and then generate the observation cost. Calculate the expected cost based on the candidate route set and the destination; specifically including: A graph attention network layer is used to learn road features and generate a road representation matrix; Based on the candidate road segment set and the destination, the road representation vectors of each road segment in the candidate road segment set and the road representation vector of the destination are obtained by indexing the road representation matrix. Calculate the spherical distance based on the road representation vector of each road segment in the candidate road segment set and the road representation vector of the destination; The expected cost is calculated based on the road representation vector of each road segment in the candidate road segment set, the road representation vector of the destination, and the spherical distance. Calculate the travel cost and generate a route trajectory based on the travel cost until the destination is reached to obtain the travel trajectory.

2. A method for constructing a deep learning network model for travel trajectory generation, characterized in that, It includes an input layer, an LSTM network layer, a constrained linear layer, and an output layer; The input layer is used to receive training data; The LSTM network layer is used to extract deep mobility situation features; A constrained linear layer is used to make decisions based on the characteristics of the deep mobile station, and an adjacency matrix is ​​introduced as a constraint on the decision. The output layer is used to generate travel trajectories based on the decision results; It also includes an expected cost calculation module, used to calculate the expected cost based on the candidate road segment set and the destination; the expected cost calculation module includes a graph attention network layer, a calculation module and a multilayer perceptron. The graph attention network layer is used to learn road features and generate a road representation matrix; The calculation module is used to index the road representation matrix according to the candidate road segment set and the destination to obtain the road representation vector of each road segment in the candidate road segment set and the road representation vector of the destination; Used to calculate spherical distance based on the road representation vector of each road segment in the candidate road segment set and the road representation vector of the destination; Based on the road representation matrix, the relative position and relative distance between the candidate road segment set and the destination are determined, and the spherical distance is calculated. The multilayer perceptron is used to calculate the expected cost based on the road representation vector of each road segment in the candidate road segment set, the road representation vector of the destination, and the spherical distance; The training data includes real trajectories, a portion of historical movement trajectories extracted from the real trajectories, and a set of candidate road segments corresponding to the historical movement trajectories; The decision-making process specifically involves: calculating the observation conditional probability between each neighboring candidate road in the candidate road segment set and the historical movement trajectory, and then using the adjacency matrix to set the observation conditional probability of non-adjacent candidate roads to 0, thereby generating the observation cost. The output layer is also used to calculate the travel cost based on the expected cost and the observed cost, and to generate a route trajectory based on the travel cost until the destination is reached to obtain the travel trajectory.

3. The method for constructing a deep learning network model for travel trajectory generation according to claim 2, characterized in that, An attention network layer is introduced between the LSTM network layer and the constrained linear layer; The correlation weights between the candidate road segment set and the historical movement trajectory are calculated using an attention mechanism; The constrained linear layer incorporates the correlation weights when calculating conditional probabilities.

4. A training method for a deep learning network model for travel trajectory generation, characterized in that, The deep learning network model for generating travel trajectories as described in claim 2 or 3 is employed, and the steps include: A generator is constructed based on the deep learning model; Acquire training data, which includes real trajectories composed of multiple road segment trajectories; wherein the road segment trajectories include spatial and temporal information of the trajectories; The origin and destination of the multiple road segments are input into the generator to generate a composite travel trajectory. The spatial and temporal information corresponding to the road segments in the actual trajectory and the travel trajectory are processed to obtain the trajectory spatial sequence and time sequence. The trajectory spatial sequence and the time sequence are respectively input into the trained discriminator to obtain a reward signal; wherein, the reward signal includes a sequential reward and a mobility yaw reward; The generator gradient is calculated based on the reward signal.

5. The training method for a deep learning network model for travel trajectory generation according to claim 4, characterized in that, The discriminator includes a preprocessing module, a sequence order extraction module, a sequence reward module, and a mobility yaw reward module; The preprocessing module is a generator for historical training rounds, used to supplement the current travel trajectory into a complete travel trajectory; The sequence order extraction module receives the conversion mode of the extracted time series of the complete travel trajectory and predicts the true probability of the trajectory. The sequential reward module calculates the sequential reward based on the true probability; The mobility deviation reward module calculates the mobility deviation reward based on the actual trajectory label and the complete travel trajectory.

Images