An individual daily travel mode prediction method and system based on a deep neural network

Abstract

The application discloses an individual whole-day travel mode prediction method based on a deep neural network, which comprises the following steps: obtaining a classification label system of individual whole-day travel modes, wherein the classification label system at least comprises single-mode travelers, non-bus multi-mode combination travelers and bus multi-mode combination travelers; and constructing a cascade deep neural network model, taking environmental features and travel plan features as inputs, and performing step-by-step prediction on the travel modes in the classification label system; the cascade deep neural network model comprises multiple sub-networks, and the inputs and outputs of a previous sub-network are simultaneously used as inputs of a next sub-network, so that step-by-step prediction of user labels is realized. The overall prediction accuracy of the application reaches 94.90%, and the problem of low capturing accuracy of multi-mode combination travelers in the prior art is effectively solved.

Description

Technical Field

[0001] This invention relates to the fields of traffic behavior analysis and deep learning technology, and in particular to a method and system for predicting individual all-day travel patterns based on deep neural networks. Background Technology

[0002] Multimodal travel is defined as "using multiple modes of transportation within a certain time frame." Early studies primarily categorized individuals' typical travel patterns by statistically analyzing their mode of transportation usage within the survey period. For example, Nobis, using data from a one-week travel survey in Germany, categorized travelers into seven typical patterns: car-only travel, bicycle-only travel, public transport-only travel, car + bicycle combination travel, car + public transport combination travel, bicycle + public transport combination travel, and car + bicycle + public transport combination travel. Yin et al., using data from a one-day activity-travel survey in Paris, identified seven typical travel patterns: car-only travel, public transport-only travel, slow-moving-only travel, car + public transport combination travel, car + slow-moving-combination travel, public transport + slow-moving-combination travel, and car + public transport + slow-moving-combination travel. This indicates that the characteristics of typical individual travel patterns are primarily influenced by local mobility culture, and establishing traveler profiles that conform to local characteristics is a crucial foundation for understanding multimodal travel.

[0003] However, the above method of statistical analysis based solely on travel mode has obvious shortcomings: since it only considers travel mode as an indicator to distinguish individual travel patterns, there are still significant differences in individual travel purposes, travel distances, and number of trips within the same mode. Using this general classification method makes it difficult to truly understand the heterogeneity of individual travel decisions.

[0004] To address the aforementioned issues, some researchers have proposed in recent years that travel patterns are a representation of multidimensional travel characteristics, and that typical travel patterns should be identified by comprehensively considering multiple indicators such as the number of trips, travel purpose, travel distance, and mode of travel. For example, Diana used cluster analysis based on travel intensity and mode of travel to identify four typical travel patterns in the Paris and Lyon metropolitan areas of France: single-mode car travelers, car-dominated multi-mode travelers, and multi-mode travelers with two different travel frequencies. Typical travel patterns in the San Francisco Bay Area of ​​the United States showed more car-dominated characteristics, including a single-mode car travel pattern with a low travel frequency and three car-dominated multi-mode travelers with different travel frequencies. Compared to relying solely on travel modes for statistical analysis, identifying typical combination patterns based on multidimensional characteristic indicators allows for a direct comparison of the heterogeneity of travel decisions from an overall travel perspective, which is more helpful in understanding the decision-making mechanisms of multi-mode travel behavior.

[0005] However, the existing technology still has the following technical problems: First, existing research does not adequately consider multi-modal travel patterns at the single-trip level. In fact, since almost every public transport trip involves other modes of transportation, single-trip behavior exhibits more common and complex multi-modal combinations. For example, individual single rail trip statistics show seven typical access and exit combinations: walking-walking, walking-bus, bicycle-walking, bicycle-bus, bus-walking, bus-bus, and car-walking. Krygsman et al., using Dutch survey data based on detailed trip phase segmentation, found that access and exit combinations differ depending on the primary mode of transport within a single trip. When the primary mode of transport is a car, the typical access mode is walking, and the typical exit mode is bus / metro / train. When the primary mode of transport is a bus, the typical access mode is walking / bicycle / car, and the typical exit mode is walking. When the primary mode of transport is a metro / tram, the typical access mode is walking / bicycle / car / bus, and the typical exit mode is walking / bus. When the primary mode of transport is a train, the typical access mode is walking / bicycle / car / bus / metro, and the typical exit mode is walking / bicycle / bus. However, these access and outbound processes are often overlooked in the recognition of all-day travel patterns. This results in the fact that the accuracy of capturing multi-modal travelers in existing studies is not high, and there is a lack of identification and analysis of typical multi-modal travel patterns of individuals from the perspective of multiple "door-to-door" trips throughout the day.

[0006] Second, existing clustering analysis methods have limitations in identifying travel patterns.

[0007] Existing research has shown that identifying single and multi-modal travel patterns based on multi-dimensional indicators yields more comprehensive and accurate results than simply classifying travel patterns based on individual travel modes. In these studies, both regression analysis and cluster analysis have demonstrated superior ability to handle multiple indicators. Regression analysis, as a supervised learning method, is primarily used to analyze the influencing factors of differences between travel patterns when travel pattern classifications are predetermined. These factors mainly include four categories: personal and family characteristics, travel characteristics, built environment characteristics (or spatial characteristics), and subjective attitude characteristics. Typical studies include those by Klinger, Buehler, and Mao, whose findings provide important clues for user profiling of different travel patterns and can be considered the foundation for travel pattern classification. Cluster analysis, as an unsupervised learning method, can be used to find homogeneous groups in multivariate data, and is therefore considered the most commonly used method for identifying travel patterns. Compared to regression analysis, cluster analysis generally only includes core indicators that significantly influence travel pattern selection and can expand the dimensions of travel, focusing on travel purpose, travel distance by mode, travel time, and frequency of travel by mode.

[0008] Based on the specific algorithms used in cluster analysis, existing research can be divided into latent class clustering and K-means clustering. While latent class clustering can use statistical criteria to determine the optimal number of clusters, its calculation method is complex, unsuitable for large datasets, and improper initial selection of indicators may lead to suboptimal results. K-means clustering is simple and efficient, and is considered a more effective method for processing massive amounts of data; however, this algorithm is sensitive to the location of the initial centroids and is prone to converging to local optima. More importantly, existing research typically uses a single clustering method to perform a one-time clustering of all travelers, failing to consider the significant differences in characteristics within different categories of travelers, making it difficult to achieve refined user segmentation.

[0009] Third, existing technologies lack effective methods for predicting the spatiotemporal heterogeneity of an individual's daily "door-to-door" travel patterns. Existing studies mainly identify and classify travel patterns through the statistical characteristics of mediating variables such as travel intensity, travel mode, and travel distance. These methods cannot accurately characterize the spatiotemporal heterogeneity of an individual's multimodal travel decision-making behavior in multimodal transportation networks. Furthermore, they have failed to construct predictive models of an individual's daily "door-to-door" travel patterns that can be quantified based on built environment characteristics and travel plan characteristics. Moreover, they lack technical solutions that can achieve progressively refined predictions.

[0010] In summary, the existing technologies urgently need to address the following technical problems: (1) How to incorporate the entire process of access, main travel, and pick-up / drop-off in a single trip into the individual's daily travel pattern recognition from the perspective of multiple "door-to-door" trips throughout the day, so as to improve the accuracy of capturing multi-mode travelers; (2) How to construct a hierarchical clustering analysis framework suitable for large datasets, overcome the problem that single clustering methods are difficult to handle the heterogeneity of multiple types of travelers, and achieve refined segmentation of travelers; (3) How to construct a deep learning model that can make high-precision predictions of the identified multiple subdivided travel modes based on multi-source input variables (built environment features, travel plan features). Summary of the Invention

[0011] This application provides a method for predicting individual all-day travel patterns based on deep neural networks, which can effectively solve the problems of low accuracy in capturing multi-modal travelers and lack of typical travel pattern recognition from multiple "door-to-door" perspectives throughout the day in existing technologies.

[0012] This application provides a method for predicting an individual's all-day travel patterns based on deep neural networks, including: A classification tag system for individual daily travel patterns is obtained, wherein the classification tag system includes at least the categories of single-mode travelers, non-public transport multi-mode travelers, and public transport multi-mode travelers; A cascaded deep neural network model is constructed, taking built environment features and travel plan features as inputs. The model is configured to predict user labels step by step through multiple subnetworks, wherein the input and output of the previous subnetwork are simultaneously used as the input of the next subnetwork. The following steps are performed using the cascaded deep neural network model: The system inputs the built environment characteristics and travel plan characteristics into the first-level network and outputs a judgment on whether the user is a public transport-type combined traveler. By inputting the output of the first-level network into the second-level network, along with the built environment characteristics and travel plan characteristics of the initial input, the system outputs a judgment on whether the user is a non-public transport combined traveler. Combining the judgment results of the first-level network and the second-level network, the system outputs predictions for user category labels for single-mode travelers, non-public transport combination travelers, and public transport combination travelers. Based on the different predicted labels, they are fed into their respective sub-networks for further processing: For single-mode travelers, the third-level network outputs a prediction of whether they prefer single slow-moving or single-car travel. Then, based on the result, the fourth-level network is selected to predict their travel elasticity for single slow-moving travel and their travel intensity and elasticity for single-car travel. Finally, the prediction results with detailed labels are given. For those who do not use public transportation, the system directly provides detailed labels for judgment through the third-level network. For travelers using a combination of public transport and rail travel, the third-level network predicts whether they prefer regular public transport or rail travel, and the fourth-level network predicts the frequency of their primary mode of travel. The fourth-level subnetwork for regular public transport preference directly outputs the predicted sub-label results, while the fourth-level subnetwork for rail preference outputs the results and adds feature variables as input to the subsequent fifth-level network to further predict the degree of motorization of connections. The fifth-level network combines the output results of the fourth and fifth levels to determine the sub-label of rail preference.

[0013] Preferably, the first-level network contains two hidden layers, which are composed of 8 Tanh activation functions and 4 Sigmoid activation functions, respectively. The Sigmoid activation function produces two outputs about the probability of whether the user is a public transport traveler, and the Argmax function explicitly gives the judgment. The hidden layer and output layer structures of the second-level network are the same as those of the first-level network; The network structure of the third-level network consists of two hidden layers and an output layer composed of Tanh and Sigmoid functions; the hidden layers of the orbital-biased fourth-level sub-network contain neurons with ReLU activation functions; the fifth-level network contains three hidden layers composed of ReLU and Tanh activation functions.

[0014] Preferably, the built environment characteristics include 52 measurement indicators in three categories of the "7D" elements of the built environment; the travel plan characteristics include the number of trips per day, the number of commuting trips, the number of trips within 5 kilometers of the origin-destination distance, the proportion of trips with departure times during peak hours, the maximum distance of a single commuting and non-commuting trip, the maximum time spent on a single trip, and the maximum cost of a single trip.

[0015] Preferably, during the training process of the cascaded deep neural network model, the loss function used in each layer of the network is the cross-entropy loss function, the RAdam algorithm is used to optimize the model parameters, and the learning rate is 0.001; before training, the minority class is oversampled and the majority class is undersampled, and a relatively balanced dataset is resampled and recombined; the training set and the test set are randomly divided in a 4:1 ratio, and the training iteration step size is set to 500 steps.

[0016] Preferably, the individual all-day travel mode includes 6 single-mode travel modes, 2 non-public transport multi-mode combination travel modes, and 12 public transport multi-mode combination travel modes; The six single-mode travel patterns are: low-elasticity single slow-moving travel pattern, high-elasticity single slow-moving travel pattern, low-elasticity low-intensity single car travel pattern, low-elasticity high-intensity single car travel pattern, high-elasticity low-intensity single car travel pattern, and high-elasticity high-intensity single car travel pattern. The two non-public transport-related multi-modal travel modes are: a car-dominated + slow-moving travel mode and a car-dominated + slow-moving travel mode. The 12 types of multi-modal travel modes include: low-frequency public transport mode with low-motorization connections, medium-frequency public transport mode with low-motorization connections, high-frequency public transport mode with low-motorization connections, low-frequency rail transport mode with low-motorization connections, low-frequency rail transport mode with medium-motorization connections, low-frequency rail transport mode with high-motorization connections, medium-frequency rail transport mode with low-motorization connections, medium-frequency rail transport mode with medium-motorization connections, medium-frequency rail transport mode with high-motorization connections, high-frequency rail transport mode with low-motorization connections, high-frequency rail transport mode with medium-motorization connections, and high-frequency rail transport mode with high-motorization connections.

[0017] This invention also discloses an individual's all-day travel pattern prediction system based on deep neural networks, comprising: The classification tag system acquisition module is used to acquire the classification tag system of an individual's daily travel patterns. The classification tag system includes at least the categories of single-mode travelers, non-public transport multi-mode travelers, and public transport multi-mode travelers. A cascaded deep neural network model is used to predict travel patterns in the classification label system step by step, taking built environment features and travel plan features as inputs. The cascaded deep neural network model includes multiple subnetworks, and the step-by-step prediction of user labels is achieved by using the input and output of the previous subnetwork as the input of the next subnetwork. The cascaded deep neural network model specifically includes: The first-level network is used to input the built environment characteristics and travel plan characteristics, and output a judgment on whether the user is a public transport-type combined traveler. The second-level network takes the output of the first-level network, along with the built environment features and travel plan features of the initial input, as input, and outputs a judgment on whether the user is a non-public transport combined traveler. The output module combines the judgment results of the first-level network and the second-level network to output the predicted user category labels for single-mode travelers, non-public transport combination travelers, and public transport combination travelers. The third-level network is used to process data according to the different predicted labels: For single-mode travelers, the third-level network outputs a prediction of whether the traveler prefers single slow-moving mode or single car mode. Based on the result, the fourth-level network is selected to predict the travel elasticity of single slow-moving mode and the travel intensity and elasticity of single car mode, and finally, the prediction results with detailed labels are given. For those who do not use public transportation, the third-level network directly provides a detailed label for judgment; For travelers using public transport, the third-level network predicts whether they prefer regular public transport or rail travel, while the fourth-level network predicts the frequency of their primary mode of travel. The fourth-level subnetwork indicating preference for regular public transport directly outputs the predicted sub-label, while the output of the fourth-level subnetwork indicating preference for rail travel incorporates feature variables as input to the subsequent fifth-level network to further predict the degree of motorization of connections. The fifth-level network combines the outputs of the fourth and fifth levels to determine the sub-label of preference for rail travel.

[0018] Preferably, the first-level network contains two hidden layers, which are composed of 8 Tanh activation functions and 4 Sigmoid activation functions, respectively. The Sigmoid activation function produces two outputs about the probability of whether the user is a public transport traveler, and the Argmax function explicitly gives the judgment. The hidden layer and output layer structures of the second-level network are the same as those of the first-level network; The network structure of the third-level network consists of a double hidden layer and an output layer composed of Tanh and Sigmoid functions; The hidden layers of the orbital-type-oriented four-level subnetwork contain neurons with ReLU activation functions; The fifth-level network contains three hidden layers, which are composed of ReLU and Tanh activation functions.

[0019] Preferably, the built environment characteristics include 52 measurement indicators in three categories of the "7D" elements of the built environment; the travel plan characteristics include the number of trips per day, the number of commuting trips, the number of trips within 5 kilometers of the origin-destination distance, the proportion of trips with departure times during peak hours, the maximum distance of a single commuting and non-commuting trip, the maximum time spent on a single trip, and the maximum cost of a single trip.

[0020] Preferably, the cascaded deep neural network model uses the cross-entropy loss function, optimizes the parameters using the RAdam algorithm, and has a learning rate of 0.001. Before training, the minority class is oversampled and the majority class is undersampled, and the training set and test set are divided in a 4:1 ratio, with an iteration step size of 500 steps.

[0021] Preferably, the individual all-day travel mode includes 6 single-mode travel modes, 2 non-public transport multi-mode combination travel modes, and 12 public transport multi-mode combination travel modes; The six single-mode travel patterns are: low-elasticity single slow-moving travel pattern, high-elasticity single slow-moving travel pattern, low-elasticity low-intensity single car travel pattern, low-elasticity high-intensity single car travel pattern, high-elasticity low-intensity single car travel pattern, and high-elasticity high-intensity single car travel pattern. The two non-public transport-related multi-modal travel modes are: a car-dominated + slow-moving travel mode and a car-dominated + slow-moving travel mode. The 12 types of multi-modal travel modes include: low-frequency public transport mode with low-motorization connections, medium-frequency public transport mode with low-motorization connections, high-frequency public transport mode with low-motorization connections, low-frequency rail transport mode with low-motorization connections, low-frequency rail transport mode with medium-motorization connections, low-frequency rail transport mode with high-motorization connections, medium-frequency rail transport mode with low-motorization connections, medium-frequency rail transport mode with medium-motorization connections, medium-frequency rail transport mode with high-motorization connections, high-frequency rail transport mode with low-motorization connections, high-frequency rail transport mode with medium-motorization connections, and high-frequency rail transport mode with high-motorization connections.

[0022] Compared with the prior art, the technical solution provided in this application has at least the following technical effects or advantages: 1. High prediction accuracy: This invention achieves high-precision prediction of an individual's daily travel patterns by constructing a four-layer clustering analysis architecture and a cascaded deep neural network model.

[0023] 2. High pattern recognition resolution: This invention is the first to incorporate the entire process of a single trip—the access phase, the main travel phase, and the pick-up / drop-off phase—into all-day travel pattern recognition, effectively solving the problem of low resolution in capturing multi-modal travelers in existing technologies. Through a multi-layer clustering analysis model, a total of 20 typical travel patterns were identified, including 6 single-modal travel patterns, 2 non-public transport multi-modal travel patterns, and 12 public transport multi-modal travel patterns, forming a complete classification system for individual all-day "door-to-door" travel patterns.

[0024] 3. Stable model training and strong generalization ability: The cascaded deep neural network model constructed in this invention adopts a staged processing strategy, using the input and output of the previous sub-network as the input of the next sub-network. During training, the loss value converges stably, and the accuracy is mostly above 90% after multiple iterations. The model also shows stable performance on the test set and has good generalization ability.

[0025] 4. The role of the technical framework in connecting the upper and lower levels: The present invention plays a connecting role in the technical framework. It can realize the efficient transmission of built environment regulation variables at the macro level of the community, and provide high-quality prediction results for the lower-level path optimization model, so as to achieve comprehensive optimization of individual daily travel paths. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the multi-layer clustering model analysis framework based on the daily travel chain of this invention; Figure 2This is a schematic diagram of the cascaded neural network structure constructed in this invention; Figure 3 This is a schematic diagram of the training iteration process of the cascaded neural network of the present invention, where (a) represents the change in training loss value and (b) represents the change in prediction accuracy during the training process. Detailed Implementation

[0027] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0028] Example: This example provides a method for predicting individual daily travel patterns based on deep neural networks, including the construction of a multi-layer clustering analysis model, the clustering and subdivision results of typical daily travel patterns of individuals, and the construction and prediction of a deep neural network model. The specific process is as follows: 1. Construction of multi-level clustering analysis model This embodiment employs cluster analysis to identify typical patterns of single and multi-modal travel combinations. Since the individual's daily travel trajectory features, mined from spatiotemporal big data, contain detailed multi-stage features of a single "door-to-door" trip, they can effectively capture the various modes of transportation (car, rail, public transport, and bicycles) that may be used at different stages of the entire travel process. At the statistical level of travel modes alone, there are 8042 different combination types. To ensure analytical efficiency and improve accuracy, this embodiment uses a multi-dimensional K-means clustering algorithm to identify typical combination patterns.

[0029] The overall cluster analysis process is as follows: Figure 1 As shown. This embodiment constructs a multi-layer clustering analysis model based on this, specifically including one classification label layer and three clustering analysis layers. The first layer is the classification label layer, which divides travelers in the dataset into three categories based on the travel modes used in an individual's daily "door-to-door" travel chain: single-mode travelers (40.2%), non-public transport multi-mode travelers (37.3%), and public transport multi-mode travelers (22.5%). The second, third, and fourth layers are clustering analysis layers, which use the K-means clustering method to segment users for each category of travelers. The clustering indicators are selected as follows: For single-mode travelers, the total number of trips per day, the proportion of commuting trips, and the proportion of short-distance trips per day are used as clustering indicators. For non-public transport users who use a combination of modes of transportation, the proportion of slow-moving trips in short-distance trips, the proportion of car trips in commuting trips, and the proportion of car trips in total daily trips will be used as clustering indicators. For travelers using multiple modes of transportation, the clustering indicators are as follows: during the primary mode of transportation phase, the percentage of public transportation trips in the total daily trips, the percentage of public transportation trips in the commuting trips, and the percentage of public transportation trip distance in the total daily trip distance; during the connection and pick-up phase, the clustering indicators are as follows: the percentage of motorized shuttle trips in the total number of shuttle trips and the percentage of motorized shuttle trip distance in the total shuttle trip distance.

[0030] 2. Clustering and subdivision results of typical daily travel patterns of individuals The multi-level clustering analysis model identified 20 typical travel patterns, including 6 single-mode travel patterns, 2 non-public transport multi-mode travel patterns, and 12 public transport multi-mode travel patterns. Details are as follows: (1) Clustering subdivision results of single-mode travelers: The clustering subdivision results and mean values ​​of indicators for single-mode travelers are shown in Table 1. According to the clustering results, there are six subdivisions of single-mode travelers: Low-elasticity single slow travel mode (LAM), High-elasticity single slow travel mode (HAM), Low-elasticity low-intensity single car travel mode (LLC), Low-elasticity high-intensity single car travel mode (LHC), High-elasticity low-intensity single car travel mode (HLC), and High-elasticity high-intensity single car travel mode (HHC).

[0031] The clustering index means show that the low-elasticity single slow travel mode (LAM) and the high-elasticity single slow travel mode (HAM) have similar average daily trip counts and short-distance trip percentages, but the average commuting trip percentages differ significantly. In the LAM mode, commuting trips account for 65.34%, while in the HAM mode it is only 15.11%. The four car-only travel groups show significant differences mainly in the percentages of commuting trips and short-distance trips. In summary, the low-elasticity low-intensity single car travel mode (LLC) is dominated by short-distance commuting, while the low-elasticity high-intensity single car travel mode (LHC) is dominated by long-distance commuting. The high-elasticity low-intensity single car travel mode (HLC) is dominated by short-distance leisure or shopping trips, while the high-elasticity high-intensity single car travel mode (HHC) is dominated by long-distance leisure or shopping trips.

[0032] Table 1. Clustering subdivisions and clustering index levels of daytime single-mode travelers

[0033] (2) Clustering subdivision results of non-public transport combined travelers: The clustering subdivision results and mean values ​​of non-public transport combined travelers are shown in Table 2. They can be further divided into car + slow-moving travel mode (HC+AM) which is highly dominated by private cars and car + slow-moving travel mode (MC+AM) which is moderately dominated by private cars. In the car + slow-moving travel mode (HC+AM), the proportion of private cars used throughout the day is 87.18%, and the proportion of private cars in commuting trips is 86.24%. In the car + slow-moving travel mode (MC+AM), which is moderately dominated by private cars, slow-moving travel mode is used more frequently for short distances and commuting trips.

[0034] Table 2. Clustering subdivisions and clustering index levels of all-day non-public transport combined travelers.

[0035] (3) Clustering and segmentation results of public transport-based multimodal travelers: Twelve further subdivision patterns were identified among public transport-based multimodal travelers, including three multimodal travel patterns primarily based on public transport and nine multimodal travel patterns primarily based on rail transport. Specifically: The tiered multimodal transport modes are: Low-frequency public transport mode (LBLM) with low-motorization connections, Medium-frequency public transport mode (MBLM) with low-motorization connections, High-frequency public transport mode (HBLM) with low-motorization connections, Low-frequency rail transport mode (LRLM) with low-motorization connections, Low-frequency rail transport mode (LRMM) with medium-motorization connections, Low-frequency rail transport mode (LRHM) with high-motorization connections, Medium-frequency rail transport mode (MRLM) with low-motorization connections, Medium-frequency rail transport mode (MRMM) with medium-motorization connections, Medium-frequency rail transport mode (MRHM) with high-motorization connections, High-frequency rail transport mode (HRLM) with low-motorization connections, High-frequency rail transport mode (HRMM) with medium-motorization connections, and High-frequency rail transport mode (HRHM) with high-motorization connections. In terms of overall proportion, the three multimodal transport modes primarily based on public transport account for 65% of public transport users, while the nine multimodal transport modes primarily based on rail account for 35%.

[0036] 3. Deep Neural Network Model Construction and Prediction (1) Input variable analysis: The input variables of this layer model include 52 measurement indicators of three categories of the “7D” elements of the built environment, as well as the number of trips per day, the number of commuting trips, the number of trips within 5 kilometers of the OD distance, the proportion of trips with departure time at peak times, the maximum distance of a single commuting and non-commuting trip, the maximum time consumption of a single trip, and the maximum cost of a single trip.

[0037] (2) Model Structure Construction: This embodiment constructs a cascaded deep neural network model. By nesting multiple sub-networks and using the input and output of the previous sub-network as the input of the next sub-network, step-by-step prediction of user labels is achieved. The specific model structure is as follows: Figure 2 As shown. Wherein: The first-level network (NN1) takes the feature variables as input and contains two hidden layers (composed of 8 Tanh activation functions and 4 Sigmoid activation functions, respectively). After passing through the Sigmoid activation function, it obtains two outputs regarding the probability of whether the user is a public transport traveler. The Argmax function then explicitly gives the judgment.

[0038] The second-level network (NN2) takes the output of the first-level network and the initial input feature variables as input. Its hidden and output layer structures are identical to the first-level network. It outputs the probability that a user is a non-public transport combination traveler, which is made explicit through the Argmax function. Combining the judgment results from both levels, it outputs predictions for user category labels: single-mode traveler, non-public transport combination traveler, and public transport combination traveler.

[0039] Based on the predicted labels, each level's network feeds into its respective sub-network. The input variables for the third-level networks are the same 52 feature variables as the first-level networks, and the network structure consists of two hidden layers and an output layer composed of Tanh and Sigmoid functions. For single-mode travelers, the third-level network (NN3) outputs a prediction of whether they prefer single-mode travel (slow-moving vehicles) or single-mode travel (cars). Based on this result, a fourth-level network is selected to predict their travel elasticity (NN4) for single-mode travel and their travel intensity and elasticity (NN5) for single-mode travel, ultimately providing the sub-label prediction. For non-public transport combined travelers, the third-level network (NN6) directly provides the sub-label judgment. For public transport combined travelers, the third-level network (NN7) predicts whether they prefer regular public transport or rail travel, and the fourth-level network predicts the frequency of their primary mode of travel. The fourth-level subnetwork (NN8) for conventional public transport orientation directly outputs the predicted sub-label results. The fourth-level subnetwork (NN9) for rail transit orientation contains ReLU activation function neurons in its hidden layers. Its output is added to feature variables as input to the subsequent fifth-level network (NN10) to further predict the degree of rail transit mobility. The fifth-level network contains three hidden layers (composed of ReLU and Tanh activation functions). The sub-labels for rail transit orientation are determined by combining the output results of the fourth and fifth levels.

[0040] The structural design of the cascaded deep neural network model in this embodiment strictly follows the decision-making hierarchy of an individual's daily travel behavior. Specifically: Individual travel mode choices exhibit a clear hierarchy: individuals first roughly decide whether they rely on public transportation (bus or rail) for their entire day's travel, a decision that corresponds to the output of the first-level network in this model (whether they are a combined public transportation traveler); secondly, for non-public transportation travelers (those who do not use public transportation throughout the day), a further distinction is made between those who use a combination of car and motorized travel throughout the day and those who use only car or motorized travel throughout the day, corresponding to the output of the second-level network (whether they are a non-combined public transportation traveler). The segmentation dimensions differ across different traveler categories: For single-mode travelers, segmentation primarily relies on travel elasticity (commuting percentage) and intensity (trip frequency). Therefore, the third-level network and its subsequent sub-networks (NN3, NN4, NN5) predict travel mode preference and travel elasticity and intensity, respectively. For non-public transport multi-mode travelers, segmentation mainly depends on the ratio of car to road travel; therefore, the third-level network (NN6) directly outputs segmentation labels. For public transport multi-mode travelers, segmentation involves a deeper decision chain: first, the primary mode of transport (bus or rail), then usage frequency (low, medium, high), and finally, for rail travelers, the degree of connecting mobility (low, medium, high). In this model, the third-level network (NN7) predicts primary mode preference, the fourth-level network predicts usage frequency, and the fifth-level network (NN10) predicts connecting mobility, achieving a complete modeling of the decision chain for public transport travelers.

[0041] Therefore, it can be seen that the cascaded neural network in this embodiment is not a simple application of a general network structure, but a dedicated architecture specifically designed for the technical problem of predicting individual all-day travel patterns, based on the decision-making level of travel behavior.

[0042] (3) Model Training Results: The training and calibration of the cascaded neural network followed a hierarchical approach. After completion, the model state dictionary was saved as a .pth file. The cross-entropy loss function was used for each level of the network, and the RAdam algorithm was used to optimize the model parameters with a learning rate of 0.001. Before training, oversampling was performed on the minority class, and undersampling was performed on the majority class, resulting in a relatively balanced dataset. The training and test sets were randomly divided in a 4:1 ratio, and the training iteration step size was set to 500 steps. The changes in loss values ​​and accuracy of each sub-network during training and testing are shown below. Figure 3 As shown, both the loss value and accuracy converge stably, and the accuracy is mostly above 90% after multiple iterations.

[0043] (4) Model prediction results for individual travel patterns: The model results show high prediction accuracy for 20 single or multi-modal travel patterns, with an overall prediction accuracy of 94.90%. Among them, the prediction accuracy for single-modal travelers is 98.7%, the prediction accuracy for non-public transport combined travelers is 95.51%, and the prediction accuracy for public transport combined travelers is 93.73%. The prediction results for the sub-patterns of travel patterns are shown in Table 3. The sub-patterns with the highest prediction accuracy are the high-elasticity single slow-moving travel pattern (HAM), the low-elasticity low-intensity single car travel pattern (LLC), and the high-elasticity high-intensity single car travel pattern (HHC), with model accuracy exceeding 99%.

[0044] Table 3. Prediction Results of Individual Travel Patterns

[0045] Based on the above method, this embodiment also provides an individual all-day travel pattern prediction system based on deep neural networks, the system comprising: The multi-level clustering analysis model construction module is used to identify typical patterns of single and multi-modal travel combinations based on an individual's daily travel chain data. This module specifically performs the construction and cluster subdivision of the aforementioned multi-level clustering analysis model.

[0046] A cascaded deep neural network model is used to predict typical travel patterns identified by the multi-layer clustering analysis model step by step, taking built environment features and travel plan features as inputs. This model specifically executes the aforementioned cascaded neural network construction and training.

[0047] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting an individual's all-day travel patterns based on deep neural networks, characterized in that, include: A classification tag system for individual daily travel patterns is obtained, wherein the classification tag system includes at least the categories of single-mode travelers, non-public transport multi-mode travelers, and public transport multi-mode travelers; A cascaded deep neural network model is constructed, taking built environment features and travel plan features as inputs. The model is configured to predict user labels step by step through multiple subnetworks, wherein the input and output of the previous subnetwork are simultaneously used as the input of the next subnetwork. The following steps are performed using the cascaded deep neural network model: The system inputs the built environment characteristics and travel plan characteristics into the first-level network and outputs a judgment on whether the user is a public transport-type combined traveler. By inputting the output of the first-level network into the second-level network, along with the built environment characteristics and travel plan characteristics of the initial input, the system outputs a judgment on whether the user is a non-public transport combined traveler. Combining the judgment results of the first-level network and the second-level network, the system outputs predictions for user category labels for single-mode travelers, non-public transport combination travelers, and public transport combination travelers. Based on the different predicted labels, they are fed into their respective sub-networks for further processing: For single-mode travelers, the third-level network outputs a prediction of whether they prefer single slow-moving or single-car travel. Then, based on the result, the fourth-level network is selected to predict their travel elasticity for single slow-moving travel and their travel intensity and elasticity for single-car travel. Finally, the prediction results with detailed labels are given. For those who do not use public transportation, the system directly provides detailed labels for judgment through the third-level network. For travelers using a combination of public transport and rail travel, the third-level network predicts whether they prefer regular public transport or rail travel, and the fourth-level network predicts the frequency of their primary mode of travel. The fourth-level subnetwork for regular public transport preference directly outputs the predicted sub-label results, while the fourth-level subnetwork for rail preference outputs the results and adds feature variables as input to the subsequent fifth-level network to further predict the degree of motorization of connections. The fifth-level network combines the output results of the fourth and fifth levels to determine the sub-label of rail preference.

2. The method according to claim 1, characterized in that, The first-level network contains two hidden layers, consisting of 8 Tanh activation functions and 4 Sigmoid activation functions respectively. The Sigmoid activation function produces two outputs regarding the probability of whether the user is a public transport user, and the Argmax function explicitly gives the judgment. The hidden layer and output layer structures of the second-level network are the same as those of the first-level network; The network structure of the third-level network consists of two hidden layers and an output layer composed of Tanh and Sigmoid functions; the hidden layers of the orbital-biased fourth-level sub-network contain neurons with ReLU activation functions; the fifth-level network contains three hidden layers composed of ReLU and Tanh activation functions.

3. The method according to claim 1, characterized in that, The built environment characteristics include 52 measurement indicators in three categories of the "7D" elements of the built environment; the travel plan characteristics include the number of trips per day, the number of commuting trips, the number of trips within 5 kilometers of the origin-destination distance, the proportion of trips with departure times during peak hours, the maximum distance of a single commuting and non-commuting trip, the maximum time spent on a single trip, and the maximum cost of a single trip.

4. The method according to claim 1, characterized in that, During the training process of the cascaded deep neural network model, the loss function used in each layer of the network is the cross-entropy loss function, the RAdam algorithm is used to optimize the model parameters, and the learning rate is 0.

001. Before training, the minority class is oversampled and the majority class is undersampled, and a relatively balanced dataset is resampled and recombined. The training set and the test set are randomly divided in a 4:1 ratio, and the training iteration step size is set to 500 steps.

5. The method according to claim 1, characterized in that, The individual all-day travel patterns include 6 single-mode travel patterns, 2 non-public transport multi-mode travel patterns, and 12 public transport multi-mode travel patterns; The six single-mode travel patterns are: low-elasticity single slow-moving travel pattern, high-elasticity single slow-moving travel pattern, low-elasticity low-intensity single car travel pattern, low-elasticity high-intensity single car travel pattern, high-elasticity low-intensity single car travel pattern, and high-elasticity high-intensity single car travel pattern. The two non-public transport-related multi-modal travel modes are: a car-dominated + slow-moving travel mode and a car-dominated + slow-moving travel mode. The 12 types of multi-modal travel modes include: low-frequency public transport mode with low-motorization connections, medium-frequency public transport mode with low-motorization connections, high-frequency public transport mode with low-motorization connections, low-frequency rail transport mode with low-motorization connections, low-frequency rail transport mode with medium-motorization connections, low-frequency rail transport mode with high-motorization connections, medium-frequency rail transport mode with low-motorization connections, medium-frequency rail transport mode with medium-motorization connections, medium-frequency rail transport mode with high-motorization connections, high-frequency rail transport mode with low-motorization connections, high-frequency rail transport mode with medium-motorization connections, and high-frequency rail transport mode with high-motorization connections.

6. A system for predicting individual all-day travel patterns based on deep neural networks, characterized in that, include: The classification tag system acquisition module is used to acquire the classification tag system of an individual's daily travel patterns. The classification tag system includes at least the categories of single-mode travelers, non-public transport multi-mode travelers, and public transport multi-mode travelers. A cascaded deep neural network model is used to predict travel patterns in the classification label system step by step, taking built environment features and travel plan features as inputs. The cascaded deep neural network model includes multiple subnetworks, and the step-by-step prediction of user labels is achieved by using the input and output of the previous subnetwork as the input of the next subnetwork. The cascaded deep neural network model specifically includes: The first-level network is used to input the built environment characteristics and travel plan characteristics, and output a judgment on whether the user is a public transport-type combined traveler. The second-level network takes the output of the first-level network, along with the built environment features and travel plan features of the initial input, as input, and outputs a judgment on whether the user is a non-public transport combined traveler. The output module combines the judgment results of the first-level network and the second-level network to output the predicted user category labels for single-mode travelers, non-public transport combination travelers, and public transport combination travelers. The third-level network is used to process data according to the different predicted labels: For single-mode travelers, the third-level network outputs a prediction of whether the traveler prefers single slow-moving mode or single car mode. Based on the result, the fourth-level network is selected to predict the travel elasticity of single slow-moving mode and the travel intensity and elasticity of single car mode, and finally, the prediction results with detailed labels are given. For those who do not use public transportation, the third-level network directly provides a detailed label for judgment; For travelers using public transport, the third-level network predicts whether they prefer regular public transport or rail travel, while the fourth-level network predicts the frequency of their primary mode of travel. The fourth-level subnetwork indicating preference for regular public transport directly outputs the predicted sub-label, while the output of the fourth-level subnetwork indicating preference for rail travel incorporates feature variables as input to the subsequent fifth-level network to further predict the degree of motorization of connections. The fifth-level network combines the outputs of the fourth and fifth levels to determine the sub-label of preference for rail travel.

7. The system according to claim 6, characterized in that, The first-level network contains two hidden layers, consisting of 8 Tanh activation functions and 4 Sigmoid activation functions respectively. The Sigmoid activation function produces two outputs regarding the probability of whether the user is a public transport user, and the Argmax function explicitly gives the judgment. The hidden layer and output layer structures of the second-level network are the same as those of the first-level network; The network structure of the third-level network consists of a double hidden layer and an output layer composed of Tanh and Sigmoid functions; The hidden layers of the orbital-type-oriented four-level subnetwork contain neurons with ReLU activation functions; The fifth-level network contains three hidden layers, which are composed of ReLU and Tanh activation functions.

8. The system according to claim 6, characterized in that, The built environment characteristics include 52 measurement indicators in three categories of the "7D" elements of the built environment; the travel plan characteristics include the number of trips per day, the number of commuting trips, the number of trips within 5 kilometers of the origin-destination distance, the proportion of trips with departure times during peak hours, the maximum distance of a single commuting and non-commuting trip, the maximum time spent on a single trip, and the maximum cost of a single trip.

9. The system according to claim 6, characterized in that, The cascaded deep neural network model uses the cross-entropy loss function and the RAdam algorithm to optimize the parameters, with a learning rate of 0.

001. Before training, the minority class is oversampled and the majority class is undersampled. The training set and the test set are divided in a 4:1 ratio, and the iteration step size is 500 steps.

10. The system according to claim 6, characterized in that, The individual all-day travel patterns include 6 single-mode travel patterns, 2 non-public transport multi-mode travel patterns, and 12 public transport multi-mode travel patterns; The six single-mode travel patterns are: low-elasticity single slow-moving travel pattern, high-elasticity single slow-moving travel pattern, low-elasticity low-intensity single car travel pattern, low-elasticity high-intensity single car travel pattern, high-elasticity low-intensity single car travel pattern, and high-elasticity high-intensity single car travel pattern. The two non-public transport-related multi-modal travel modes are: a car-dominated + slow-moving travel mode and a car-dominated + slow-moving travel mode. The 12 types of multi-modal travel modes include: low-frequency public transport mode with low-motorization connections, medium-frequency public transport mode with low-motorization connections, high-frequency public transport mode with low-motorization connections, low-frequency rail transport mode with low-motorization connections, low-frequency rail transport mode with medium-motorization connections, low-frequency rail transport mode with high-motorization connections, medium-frequency rail transport mode with low-motorization connections, medium-frequency rail transport mode with medium-motorization connections, medium-frequency rail transport mode with high-motorization connections, high-frequency rail transport mode with low-motorization connections, high-frequency rail transport mode with medium-motorization connections, and high-frequency rail transport mode with high-motorization connections.

Images