Travel destination prediction method and system based on Logit model

Abstract

The invention discloses a travel destination prediction method and system based on a Logit model, and belongs to the technical field of travel destination prediction.The travel destination prediction method comprises the steps that a spatial-temporal characteristic matrix containing a POI type, a traffic state and a user portrait is constructed based on multi-modal data, a nested mixed Logit model is introduced, travel destination options are divided into a plurality of logic groups through a nested structure, and the POI type, the traffic state and the user portrait are predicted; relevance of options in a group is quantified through a containing value and an elastic coefficient, so that a real selection behavior is reflected more accurately, the fitting ability of the model to a complex decision path is further enhanced through an upper and lower layer probability separation design, and meanwhile individual preference differences are captured through random parameters, so that the algorithm is more accurate. Parameters of the mixed Logit model are dynamically updated through Bayesian maximum likelihood estimation, and a Bayesian framework corrects fixed effect parameters and random parameters in combination with new observation data in each iteration by introducing a parameter prior distribution and posteriori updating mechanism.

Description

Technical Field

[0001] The present invention belongs to the technical field of travel destination prediction, and specifically refers to a travel destination prediction method and system based on the Logit model. Background Art

[0002] With the rapid growth of urban traffic data and the wide application of intelligent transportation systems, travel destination prediction plays an important role in navigation services, traffic management, shared travel scheduling, etc.; However, there are still certain defects in existing travel destination prediction. When introducing new options in existing travel destination prediction, the IIA assumption will cause the relative probabilities of other options to be wrongly adjusted, which is contradictory to real behavior. It cannot simulate the complex decision-making path of individuals first selecting the regional level and then refining to specific locations. It only provides fixed effect parameters and cannot reflect the different perceptions of travel costs by different users. The fixed parameters cannot respond to emergencies, resulting in lagging prediction results. It lacks prior constraints and dynamic update mechanisms, and the parameter estimation is prone to distortion in data-scarce scenarios. It ignores cross-scale interactions, resulting in poor adaptability of the model to non-uniform spatial distributions. Therefore, a travel destination prediction method and system based on the Logit model are proposed. Summary of the Invention

[0003] The purpose of the present invention is to provide a travel destination prediction method and system based on the Logit model to solve the problems raised in the above background art.

[0004] To achieve the above purpose, the present invention provides the following technical solutions: A travel destination prediction method based on the Logit model, including the following steps: S1. Based on multi-modal data, construct a spatio-temporal feature matrix including POI types, traffic states, and user portraits; S2. According to the preprocessed spatio-temporal feature matrix in S1, screen key variables through Lasso regression and combine PCA dimensionality reduction to generate a low-dimensional feature vector; S3. Combine the low-dimensional feature vector in S2 to construct a mixed Logit model; S4. Based on the mixed Logit model in S3, dynamically update the parameters through Bayesian maximum likelihood estimation; S5. Based on the parameters of the updated mixed Logit model in S4, construct a multi-scale spatial modeling by combining gridization and attention mechanism; S6. According to the gridized spatial features and attention weights, combine real-time data, and generate the destination selection probability through the updated mixed Logit model, and simulate the impact of the planning scheme on travel distribution; S7. Obtain the actual travel data, compare it with the prediction result, and perform error feedback optimization.

[0005] Preferably, S1: Based on multi-modal data, extract features of POI types, traffic states, and user portraits. Concatenate the feature vectors of POI types, traffic states, and user portraits by grid cells to form a unified feature vector. Assign weights according to feature importance, and perform dimensionality reduction through PCA to retain the main information and reduce redundancy. Construct a spatio-temporal feature matrix, which includes: Time dimension: Slice by a fixed time granularity to form a time series; Space dimension: Use grid cells as the basic unit, and each cell corresponds to a spatial coordinate; Feature dimension: Integrate the feature vectors of POI types, traffic states, and user portraits to form high-dimensional features; Dynamically update the spatio-temporal feature matrix according to real-time traffic data and user behaviors.

[0006] Preferably, S2: According to the preprocessed spatio-temporal feature matrix in S1, screen key variables through Lasso regression and combine with PCA dimensionality reduction to generate low-dimensional feature vectors; standardize each feature dimension of the spatio-temporal feature matrix; Define the Lasso regression objective function as: , In the formula, represents the standardized spatio-temporal feature matrix, y represents the target variable, represents the regression coefficient vector, represents the regularization parameter, represents the L1 norm; Through K-fold cross-validation, calculate the mean squared error corresponding to different λ, and select the one with the minimum mean squared error; According to the subset of key variables screened by the Lasso regression objective function, generate low-dimensional feature vectors through PCA dimensionality reduction.

[0007] Preferably, for S3, construct a mixed Logit model. The steps to realize the probability of individual selection options include: Obtain the low-dimensional feature vectors of S2 , and assume that the utility of individual i for option j is , represents , represents the fixed utility parameter, represents the random parameter, represents the independent error term; Preferably, group the options logically, with each group as a nest, and calculate the inclusive value of nest r. The implementation formula is: , In the formula, Represents the inclusive value of nest r, reflecting the overall utility of the options within the nest. Indicates that option j belongs to nest r. Represents the elasticity coefficient of nest r, regulating the correlation of the options within the nest. Represents the utility of option j for individual i.

[0008] Preferably, in S3, the probability of individual i choosing an option is divided into the upper - level nest selection probability and the lower - level nest selection probability. The upper - level nest is used for the individual to first select the nest to which it belongs, and the lower - level nest is used to further select a specific option within the selected nest. The upper - level nest selection probability is implemented as: , In the formula, Represents the probability that individual i selects nest r. Represents the fixed - effect parameter of the nest level. Represents the inclusive value of nest r. The lower - level nest selection probability is implemented as: , In the formula, Represents the probability that individual i selects option j within nest r. Represents the utility of option j for individual i. Represents the elasticity coefficient of nest r.

[0009] Preferably, in S3, the total probability of individual i choosing option j includes the following steps: multiplying the upper - level nest selection probability and the lower - level nest selection probability. The total probability that individual i chooses option j is: , In the formula, Represents the total probability that individual i selects option j.

[0010] Preferably, in S4, the steps of dynamically updating the parameters through Bayesian maximum likelihood estimation include: , In the formula, Represents the set of updated parameters after the (t + 1)-th iteration. Represents the set of model parameters, including fixed - effect parameters and random parameters. Represents the integral over the distribution of the random parameter of. Represents the product of the joint probabilities of J options. Represents the probability that individual i selects option j. Represents the random parameter of the distribution density function. Represents the integration variable, the integration operation over the random parameter. Represents the parameter The logarithm of the prior distribution, maximizing the posterior log-likelihood, combining the data likelihood and the parameter prior, for each individual i, integrating denotes the probability of the choice behavior of individual i under the parameter , and restricts the value range of the parameter through the prior. By prior, restrict the value range of the parameter.

[0011] Preferably, for the step S5 of constructing the multi-scale spatial modeling: divide the research area into multiple regular grid cells, each grid corresponding to an independent spatial unit, and extract multi-dimensional features for each grid cell, including static features, dynamic features, and behavior features; Define multi-scale grid levels, including coarse-grained grids, medium-grained grids, and fine-grained grids. Coarse-grained grids: capture large-scale spatial patterns; medium-grained grids: balance local and global features; fine-grained grids: focus on microscopic details; Extract neighborhood features for each grid cell, capture large-scale spatial dependencies through a graph convolutional network, splice local and global features into a multi-scale feature vector, and perform weighted fusion on grid features of different scales to generate a unified spatial feature representation; Perform weighted fusion on attention maps of different scales, generate channel weights through average pooling and max pooling, emphasize important features, calculate the spatial correlation between grids, generate a spatial attention map, and process multiple attention heads in parallel, with each head focusing on features of different scales; Perform weighted fusion according to the spatial attention maps of different scales to generate the final spatial attention weight matrix; Update the fixed-effect parameters of the mixed Logit model based on the weighted spatial features, and update the random-effect parameters through the distribution of grid features.

[0012] Preferably, according to the grid-based spatial features and attention weights, combined with real-time data, generate the destination choice probability through the updated mixed Logit model, and simulate the impact of the planning scheme on the travel distribution; obtain real-time traffic state data, environmental data, and user behavior data in real time, and update the grid-based spatial features; Dynamically adjust the feature weights according to real-time data, calculate the real-time correlation between grids, and combine the attention weights at the local, mesoscopic, and macroscopic scales to generate the final spatial attention weight matrix; Correct the time cost in the utility function according to the real-time traffic state. If a road closure / accident occurs, dynamically block the choice probability of the affected grids. Pre-define the planning scheme, map the impact of the planning scheme to the grid features, correct the model parameters according to the feedback data after the implementation of the plan, count the choice results of all individuals, generate the regional-level travel distribution, predict the travel demand at different time periods in combination with the time dimension, and simulate the impact of the planning scheme on the travel distribution; Based on the updated mixed Logit model, generate a new probability distribution of destination choices for each individual.

[0013] Preferably, a travel destination prediction system based on the Logit model includes: Multimodal data feature construction module: Based on multimodal data, construct a spatio-temporal feature matrix including POI types, traffic states, and user portraits, and perform preprocessing; Feature screening and dimensionality reduction module: According to the preprocessed spatio-temporal feature matrix, screen key variables through Lasso regression and combine with PCA dimensionality reduction to generate a low-dimensional feature vector; Mixed Logit model construction module: Based on the low-dimensional feature vector, construct a mixed Logit model; Bayesian parameter dynamic update module: Based on the mixed Logit model, dynamically update parameters through Bayesian maximum likelihood estimation; Multi-scale spatial modeling module: According to the updated mixed Logit model parameters, combine grid-based and attention mechanisms to construct multi-scale spatial modeling; Destination probability generation and simulation module: According to the grid-based spatial features and attention weights, combine real-time data, and generate destination choice probabilities through the updated mixed Logit model, and simulate the impact of the planning scheme on travel distribution; Error feedback optimization module: Obtain the actual travel data, compare it with the prediction results, and perform error feedback optimization.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. Through the synergistic effect of screening key variables by Lasso regression and PCA dimensionality reduction, the present invention realizes the efficient compression of high-dimensional features and the removal of redundancy. The L1 regularization feature of Lasso regression can automatically eliminate features with low correlation with the target variable, significantly reducing the model complexity, while retaining variables crucial for the prediction results. Subsequently, PCA dimensionality reduction further extracts the principal components through eigenvalue decomposition of the covariance matrix, improving the practicality and real-time performance of the model; 2. Among them, by introducing the nested mixed Logit model, the nested structure divides travel destination options into multiple logical groups, quantifies the correlation within the group options through the inclusion value and elasticity coefficient, thereby more accurately reflecting the real choice behavior. The upper and lower layer probability separation design further enhances the model's fitting ability for complex decision-making paths, and at the same time captures individual preference differences through random parameters; 3. The present invention dynamically updates the parameters of the mixed Logit model through Bayesian maximum likelihood estimation. The Bayesian framework, by introducing the prior distribution of parameters and the posterior update mechanism, corrects the fixed-effect parameters and random parameters in each iteration by combining new observed data, ensuring that the model parameters always fit the latest data distribution. The combination of the integration operation and the prior constraint effectively suppresses the risk of overfitting, enabling the model to remain robust even in small-sample scenarios. 4. The present invention breaks through the limitations of single-scale models in modeling complex spatial relationships through the combination of multi-granularity grid partitioning and the attention mechanism. Coarse-granularity grids capture macroscopic spatial patterns, medium-granularity grids (such as community-level) balance local features and global associations, and fine-granularity grids focus on microscopic details. The collaborative application of graph convolutional networks and multi-head attention mechanisms further extracts cross-scale spatial dependence relationships, dynamically adjusts the influence of each grid, and improves the adaptability to non-uniform spatial distributions. 5. The present invention realizes the accurate prediction of the dynamic adjustment of travel distribution and the impact simulation of the planning scheme through the deep combination of real-time data and multi-scale spatial modeling. The real-time traffic state dynamically masks the selection probability of affected grids by correcting the time-cost parameter in the utility function, and the impact of the planning scheme is quantified by mapping to grid features, enabling the model to preview the changes in travel distribution under different policies. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is the operation flow of the travel destination prediction method based on the Logit model of the present invention Figure 1 ; Figure 2 is the operation flow of the travel destination prediction method based on the Logit model of the present invention Figure 2 ; Figure 3 is the operation flow of the travel destination prediction method based on the Logit model of the present invention Figure 3 ; Figure 4 is the structural schematic diagram of the travel destination prediction system based on the Logit model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0017] Embodiment 1: Please refer to Figures 1 - 4 as shown. The present invention provides a technical solution, including the following steps: S1. Based on multi-modal data, construct a spatio-temporal feature matrix including POI types, traffic states, and user portraits. S2. According to the preprocessed spatio-temporal feature matrix in S1, screen key variables through Lasso regression and combine with PCA dimensionality reduction to generate low-dimensional feature vectors. S3. Combine the low-dimensional feature vectors in S2 to construct a mixed Logit model. S4. Based on the mixed Logit model in S3, dynamically update the parameters through Bayesian maximum likelihood estimation. S5. Based on the updated parameters of the mixed Logit model in S4, construct a multi-scale spatial model by combining gridification and attention mechanism. S6. According to the gridified spatial features and attention weights, combine with real-time data, and generate the destination selection probability through the updated mixed Logit model, and simulate the impact of the planning scheme on the travel distribution. S7. Obtain the actual travel data, compare it with the prediction results, and perform error feedback optimization.

[0018] Preferably, in S1, based on multi-modal data, extract features of POI types, traffic states, and user portraits, splice the feature vectors of POI types, traffic states, and user portraits according to grid cells to form a unified feature vector, assign weights according to feature importance, and perform dimensionality reduction through PCA to retain the main information and reduce redundancy. In this embodiment, a spatio-temporal feature matrix is constructed, and the spatio-temporal feature matrix includes: Time dimension: Slice according to a fixed time granularity to form a time series. Space dimension: Take grid cells as the basic unit, and each unit corresponds to a spatial coordinate. Feature dimension: Integrate the feature vectors of POI types, traffic states, and user portraits to form high-dimensional features. Dynamically update the spatio-temporal feature matrix according to real-time traffic data and user behaviors.

[0019] Preferably, in S2, according to the preprocessed spatio-temporal feature matrix in S1, screen key variables through Lasso regression and combine with PCA dimensionality reduction to generate low-dimensional feature vectors; standardize each feature dimension of the spatio-temporal feature matrix. Specifically, define the Lasso regression objective function as: , In the formula, represents the standardized spatio-temporal feature matrix, y represents the target variable, represents the regression coefficient vector, represents the regularization parameter, represents the L1 norm; Through K-fold cross-validation, calculate the mean squared error corresponding to different λ, and select the one with the minimum mean squared error. According to the subset of key variables screened by the Lasso regression objective function, generate low-dimensional feature vectors through PCA dimensionality reduction.

[0020] In this embodiment, for S3, to construct a mixed Logit model, the steps to realize the probability of an individual's choice option include: obtaining the low-dimensional feature vectors of S2 ; Specifically, let the utility of individual i for option j be , denotes , denotes the fixed utility parameter, denotes the random parameter, denotes the independent error term. Group the options logically, with each group as a nest, and calculate the inclusive value of nest r. The implementation formula is: , In the formula, denotes the inclusive value of nest r, reflecting the overall utility of the options within the nest, ratio indicates that option j belongs to nest r, denotes the elasticity coefficient of nest r, adjusting the correlation of the options within the nest, denotes the utility of individual i for option j.

[0021] In this embodiment, the probability of individual i choosing an option is divided into the upper-level nest selection probability and the lower-level nest selection probability. The upper-level nest is used for the individual to first select the nest to which they belong, and the lower-level nest is used to further select a specific option within the selected nest. Specifically, the upper-level nest selection probability is realized as: , In the formula, denotes the probability that individual i selects nest r, denotes the fixed effect parameter of the nest level, denotes the inclusive value of nest r; Specifically, the lower-level nest selection probability is realized as: , In the formula, denotes the probability that individual i selects option j within nest r, denotes the utility of individual i for option j, denotes the elasticity coefficient of nest r.

[0022] Preferably, for the above S3, the total probability that individual i selects option j, the implementation steps include: multiplying the upper-layer nest selection probability and the lower-layer nest selection probability, and the total probability that individual i selects option j is: , In the formula, represents the total probability that individual i selects option j.

[0023] Preferably, for the above S4, the steps of dynamically updating parameters through Bayesian maximum likelihood estimation include: , In the formula, represents the updated parameter set after the (t + 1)-th iteration, represents the model parameter set, including fixed effect parameters and random parameters, represents the integral of the distribution of the random parameter , represents the product of the joint probabilities of J options, represents the probability that individual i selects option j, represents the random parameter of the distribution density function, represents the integral variable, the integral operation of the random parameter, represents the parameter of the logarithm of the prior distribution, maximizing the posterior log-likelihood, combining the data likelihood and the parameter prior, for each individual i, integrating represents the probability of the choice behavior of individual i under the parameter , constrains the value range of the parameter through the prior.

[0024] Preferably, for the above S5, the implementation steps of constructing a multi-scale spatial model: divide the research area into multiple regular grid cells, each grid corresponds to an independent spatial unit, and extract multi-dimensional features for each grid cell, including static features, dynamic features, and behavioral features; In this embodiment, a multi-scale grid hierarchy is defined, including a coarse-grained grid, a medium-grained grid, and a fine-grained grid. Coarse-grained grid: capture large-scale spatial patterns; Medium-grained grid: balance local and global features; Fine-grained grid: focus on microscopic details; In this embodiment, extract neighborhood features for each grid cell, capture large-scale spatial dependencies through a graph convolutional network, splice local and global features into a multi-scale feature vector, and perform weighted fusion on grid features of different scales to generate a unified spatial feature representation; In this embodiment, attention maps of different scales are weighted and fused, channel weights are generated through average pooling and max pooling, the spatial correlation between computational grids of important features is emphasized to generate a spatial attention map, multiple attention heads are processed in parallel, and each head focuses on features of different scales; Weighted fusion is performed according to the spatial attention maps of different scales to generate a final spatial attention weight matrix; Based on the weighted spatial features, the fixed effect parameters of the mixed Logit model are updated, and the random effect parameters are generated and updated through the distribution of grid features.

[0025] Preferably, according to the gridded spatial features and attention weights, combined with real-time data, the destination selection probability is generated through the updated mixed Logit model, and the impact of the planning scheme on the travel distribution is simulated; real-time traffic state data, environmental data and user behavior data are obtained to update the gridded spatial features; In this embodiment, the feature weights are dynamically adjusted according to real-time data, the real-time correlation between grids is calculated, and combined with the attention weights at the local, mesoscopic and macroscopic scales, a final spatial attention weight matrix is generated; The time cost in the utility function is corrected according to the real-time traffic state. If a road is closed / there is an accident, the selection probability of the affected grids is dynamically blocked. A predefined planning scheme is used to map the impact of the planning scheme to the grid features. The model parameters are corrected according to the feedback data after the implementation of the plan. The selection results of all individuals are counted to generate a regional-level travel distribution, and the travel demands in different time periods are predicted in combination with the time dimension, and the impact of the planning scheme on the travel distribution is simulated; Based on the updated mixed Logit model, a new destination selection probability distribution is generated for each individual.

[0026] Preferably, a travel destination prediction system based on the Logit model includes: Multimodal data feature construction module: Based on multimodal data, a spatio-temporal feature matrix including POI types, traffic states and user portraits is constructed and preprocessed; Feature screening and dimensionality reduction module: According to the preprocessed spatio-temporal feature matrix, key variables are screened through Lasso regression and combined with PCA dimensionality reduction to generate a low-dimensional feature vector; Mixed Logit model construction module: Based on the low-dimensional feature vector, a mixed Logit model is constructed; Bayesian parameter dynamic update module: Based on the mixed Logit model, the parameters are dynamically updated through Bayesian maximum likelihood estimation; Multi-scale spatial modeling module: According to the updated parameters of the mixed Logit model, multi-scale spatial modeling is constructed by combining gridization and the attention mechanism; Destination probability generation simulation module: According to the grid-based spatial features and attention weights, combined with real-time data, generate destination selection probabilities through the updated mixed Logit model, and simulate the impact of the planning scheme on travel distribution; Error feedback optimization module: Obtain the actual travel data, compare it with the prediction results, and perform error feedback optimization.

[0027] Working principle: By collecting POI type data, real-time traffic status, and user portraits, discretize the geographical area into unified units through grid-based spatial division. Each grid unit is assigned a multi-dimensional feature vector, including static and dynamic attributes. Through spatio-temporal alignment technology, splice the time dimension slices with the spatial grid features to form a high-dimensional spatio-temporal feature matrix, and dynamically assign weights based on feature importance to initially filter redundant information; For the preprocessed spatio-temporal feature matrix, use Lasso regression to screen key variables, retain the feature subset that has a significant impact on travel choices, and then compress the high-dimensional features into low-dimensional vectors through PCA dimensionality reduction technology to eliminate multicollinearity. Using the dimensionality-reduced features as input, construct a mixed Logit model: Simulate the user decision-making process for travel options through a nested selection structure - first select the major category, and then select specific options within the category, and introduce random parameters to capture individual heterogeneity. The parameters of the mixed Logit model are dynamically updated through Bayesian maximum likelihood estimation: Combine historical travel data as the prior distribution, and newly collected real-time data as the likelihood function, and iteratively optimize the posterior distribution of the parameters; Divide the research area into three levels of grids: coarse-grained, medium-grained, and fine-grained. Capture cross-scale spatial dependence relationships through graph convolutional networks, introduce an attention mechanism, assign weights to features at different scales, generate a spatial attention map to highlight key areas, and finally fuse multi-scale features to update the utility function of the mixed Logit model, making the prediction results have both global trends and local details. Real-time access to the real-time data stream, dynamically adjust the grid feature weights, based on the updated model, generate a destination selection probability distribution for each user, and evaluate the impact of the planning scheme on travel distribution through Monte Carlo simulation. Compare the actual travel data with the prediction results, and optimize by backpropagating the error to optimize the feature weights and model parameters.

[0028] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

[0029] The above describes the present invention and its implementation manners. Such description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. In general, if those of ordinary skill in the art are inspired by it and, without departing from the gist of the present invention, creatively design structural manners and embodiments similar to the technical solution, they shall fall within the protection scope of the present invention.

Claims

1. A travel destination prediction method based on the Logit model, characterized in that, It includes the following steps: S1. Based on multimodal data, construct a spatio-temporal feature matrix including POI types, traffic status, and user portraits; S2. According to the preprocessed spatio-temporal feature matrix in S1, screen key variables through Lasso regression and combine PCA dimensionality reduction to generate low-dimensional feature vectors; S3. Combine the low-dimensional feature vectors in S2 to construct a mixed Logit model; S4. Based on the mixed Logit model in S3, dynamically update parameters through Bayesian maximum likelihood estimation; S5. Based on the parameters of the updated mixed Logit model in S4, combine gridification and attention mechanism to construct a multi-scale spatial model; S6. According to the gridified spatial features and attention weights, combine real-time data, and generate the destination selection probability through the updated mixed Logit model, and simulate the impact of the planning scheme on the travel distribution; S7. Obtain the actual travel data, compare it with the prediction results, and perform error feedback optimization.

2. The travel destination prediction method based on the Logit model according to claim 1, characterized in that: In step S3, a mixed Logit model is constructed. The steps for realizing the probability of an individual's choice option include: obtaining the low-dimensional feature vector of S2 , let the utility of individual i for option j be , denotes , denotes the fixed utility parameter, denotes the random parameter, denotes the independent error term. The options are grouped logically, with each group being a nest. Calculate the inclusive value of nest r. The implementation formula is: ， In the formula, represents the inclusion value of nest r, indicating that option j belongs to nest r, represents the elasticity coefficient of nest r, which adjusts the correlation of options within the nest, represents the utility of individual i for option j.

3. The travel destination prediction method based on the Logit model according to claim 2, wherein: In S3, the probability that individual i selects an option is divided into the upper-level nest selection probability and the lower-level nest selection probability. The upper-level nest is used for the individual to first select the nest to which it belongs, and the lower-level nest further selects specific options within the selected nest. The upper-level nest selection probability is implemented as: ， In the formula, represents the probability that individual i selects nest r, represents the fixed effect parameter of the nest level, represents the inclusive value of nest r; The lower-level nest selection probability is implemented as: ， In the formula, represents the probability that individual i selects option j within nest r, represents the utility of individual i for option j, represents the elasticity coefficient of nest r.

4. The travel destination prediction method based on the Logit model according to claim 3, wherein: In S3, the steps for implementing the total probability that individual i selects option j include: multiplying the upper-level nest selection probability and the lower-level nest selection probability. The total probability that individual i selects option j is: ， In the formula, represents the total probability that individual i selects option j.

5. The travel destination prediction method based on the Logit model according to claim 1, characterized in that: In S4, the steps for dynamically updating parameters through Bayesian maximum likelihood estimation include: ， In the formula, represents the set of parameters updated after the (t + 1)-th iteration, represents the set of model parameters, represents the integration over the distribution of the random parameter represents the product of the joint probabilities of J options, represents the probability that individual i selects option j, represents the random parameter of the probability density function of the distribution, represents the integration variable, the integration operation over the random parameter, represents the parameter the logarithm of the prior distribution of.​ 6. The travel destination prediction method based on the Logit model according to claim 1, characterized in that: In S5, the implementation steps for constructing a multi-scale spatial model: Divide the area into multiple regular grid cells, each grid corresponding to an independent spatial unit, and extract multi-dimensional features for each grid cell, including static features, dynamic features, and behavioral features; Define multi-scale grid levels, including coarse-grained grids, medium-grained grids, and fine-grained grids, perform weighted fusion on grid features at different scales to generate a unified spatial feature representation, perform weighted fusion on attention maps at different scales to generate a final spatial attention weight matrix, update the fixed effect parameters of the mixed Logit model based on the weighted spatial features, and generate random effect parameters through the distribution of grid features for updating.

7. A travel destination prediction system based on the Logit model implemented by the method according to claim 1, including: Multimodal data feature construction module: Based on multimodal data, construct a spatio-temporal feature matrix including POI types, traffic status, and user portraits, and perform preprocessing; Feature screening and dimensionality reduction module: According to the preprocessed spatio-temporal feature matrix, screen key variables through Lasso regression and combine PCA dimensionality reduction to generate low-dimensional feature vectors; Mixed Logit model construction module: Based on low-dimensional feature vectors, construct a mixed Logit model; Bayesian parameter dynamic update module: Based on the mixed Logit model, dynamically update parameters through Bayesian maximum likelihood estimation; Multi-scale spatial modeling module: According to the parameters of the updated mixed Logit model, combine gridification and attention mechanism to construct a multi-scale spatial model; Destination probability generation and simulation module: Based on the grid-based spatial features and attention weights, combined with real-time data, generate destination selection probabilities through the updated mixed Logit model, and simulate the impact of the planning scheme on the trip distribution; Error feedback optimization module: Obtain the actual travel data, compare it with the prediction results, and perform error feedback optimization.

Images