A vehicle spatiotemporal travel pattern mining method and system based on ST-LDA

Abstract

The present invention relates to a method for mining spatiotemporal travel patterns of vehicles based on ST‑LDA, and belongs to the field of intelligent transportation technology. The method comprises the following steps: S1, data preparation: extracting a set of highway trips of a vehicle and converting it into a trip corpus; S2, constructing an ST‑LDA model: the ST‑LDA model comprises two parts: a polynomial distribution between travel patterns and spatiotemporal features, which increases the interpretability of the model from a semantic feature level; and a polynomial distribution between individual vehicles and travel patterns, which well reflects the diversified travel behaviors of individual vehicles; S3, model training: using the obtained trip words as input to train the ST‑LDA model, and using the folded Gibbs sampling method to solve the latent variables in the model to obtain travel pattern recognition results. The technical solution of the present invention reflects the diversity of individual travel behaviors, not only mines the travel patterns of vehicles from the two dimensions of time and space, but also increases the interpretability of the model, and has broad application prospects.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation technology, in particular to the field of intelligent transportation technology using big data, and relates to a vehicle spatiotemporal travel pattern mining method and system based on ST-LDA. Background Art

[0002] A pattern is defined as "a repetitive or regular way in which something occurs or is accomplished." Travel patterns can be defined as repetitive travel behaviors associated with various characteristics, including travel distance, travel time, and travel frequency. Understanding and studying people's travel patterns is a crucial prerequisite for developing appropriate transportation policies, optimizing transportation networks, and improving transportation services in transportation planning and management.

[0003] While much research has focused on mining individual travel patterns and analyzing individual travel behaviors and their patterns, various limitations remain. Some methods utilize spatiotemporal data features to fit distribution patterns, such as power-law, logarithmic, gamma, and exponential distributions. While these methods can effectively capture the distributional properties of data, they can only understand the distribution between single or two-dimensional features and fail to capture the diversity and interactions between multidimensional features. Some studies have employed clustering algorithms, which are unsupervised learning algorithms. For travel pattern mining, these algorithms do not require a predefined set of travel patterns, but instead identify different travel patterns based on data similarity. While clustering algorithms can mine multiple travel patterns, they often only provide a quantitative and categorical representation of the results and fail to provide a clear semantic interpretation of the characteristics of different travel patterns. Furthermore, because clustering algorithms classify results based on data similarity, they often assume that regions or individuals belong to only one category. However, individual travel needs are no longer monolithic; regions or individuals often reflect diverse travel patterns. For example, individuals may travel partly for commuting and partly for entertainment or other purposes. When performing cluster analysis, if there is large heterogeneity between different data, the clustering results may be inaccurate.

[0004] Topic modeling-based methods analyze vehicle travel patterns through topic modeling. Derived from generative probabilistic models, topic models were originally used to discover themes or central ideas in articles within large text corpora. This method considers a vehicle trip as a single word, all individual vehicle trip records as a document, and all vehicle trip records as a document collection. It then discovers potential topics, or potential travel patterns, from this document collection. Topic models assume that individual trips are described by travel patterns with a certain probability, avoiding the single-division problem associated with clustering algorithms. However, most of these studies simply combine multidimensional features into one-dimensional text data, ignoring the correlations between features.

[0005] Therefore, there is an urgent need for a method and system suitable for mining vehicle spatiotemporal travel patterns that can overcome the above shortcomings. Summary of the Invention

[0006] In light of this, the present invention aims to provide a method and system for mining spatiotemporal vehicle travel patterns based on ST-LDA. This method and system models travel patterns as a polynomial distribution of spatiotemporal features, and individual travel behaviors as a polynomial distribution of travel patterns. Specifically, this method represents travel behaviors as spatiotemporal feature vectors, effectively avoiding the problem of previous methods that lack attention to the spatiotemporal correlation of individual travel behaviors.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A vehicle spatiotemporal travel pattern mining method based on ST-LDA, the method comprising the following steps:

[0009] S1. Data preparation: Extract the vehicle's highway trip set and convert it into a trip corpus;

[0010] S2. Constructing the ST-LDA model: The ST-LDA model consists of two parts: a polynomial distribution between travel patterns and spatiotemporal features, which increases the interpretability of the model at the semantic feature level; and a polynomial distribution between individual vehicles and travel patterns, which well reflects the diverse travel behaviors of individual vehicles.

[0011] S3. Model training: The obtained itinerary words are used as input to train the ST-LDA model, and the latent variables in the model are solved using the folded Gibbs sampling method to obtain the travel mode recognition results;

[0012] S4. Use the model to mine vehicle spatiotemporal travel patterns.

[0013] Furthermore, in step S1, the highway trip set of the vehicle is extracted and converted into a trip corpus, specifically including:

[0014] S11: Extracting travel records for a specified time period on the expressway, integrating the travel records of individual vehicles according to their vehicle identification IDs, and arranging them in ascending order according to their entry time to obtain a set of individual vehicle travel records;

[0015] S12: Convert each trip of each vehicle into trip words. After the conversion is completed, all trip words of a vehicle constitute a trip document, and the trip documents of all vehicles constitute a trip corpus.

[0016] Furthermore, in step S12, converting the itinerary into itinerary words specifically includes:

[0017] (1): Get the travel time interval label e of the travel word;

[0018] (2): Get the trip end label r of the trip word;

[0019] (3): Get the departure time tag t of the trip word;

[0020] (4): Get the day of the week label q of the itinerary word.

[0021] Furthermore, in step S2, the ST-LDA model assumes that the itinerary corpus is generated by the following steps:

[0022] S21: For each travel mode of category z (z=1,2,..,Z), generate a multinomial distribution of the corresponding trip word labels from the Dirichlet distribution;

[0023] Repeat step S21 Z times to obtain the spatiotemporal feature distribution of all travel mode categories.

[0024] S22: Sampling to get Car i Itinerary documents.

[0025] Further, the step S22 samples and obtains Car i The itinerary documents specifically include:

[0026] S221: Generate Car from Dirichlet Distribution i Multinomial distribution of travel modes;

[0027] S222: According to Car i The multinomial distribution of travel modes is used to sample a travel mode category z, and the corresponding travel word label is sampled from the spatiotemporal distribution of travel under the travel mode of category z.

[0028] Furthermore, in step S3, the model solving step is divided into the following steps:

[0029] S31: Randomly initialize the travel mode category to which the travel words in the travel corpus belong;

[0030] S32: For Car in the itinerary corpus i Itinerary words:

[0031] S321: removing the spatiotemporal feature category label corresponding to the word from the travel mode of the current trip word;

[0032] S322: Calculate the posterior probability of the trip word belonging to the travel mode of category z when other trip words and travel mode categories are given;

[0033] S323: Perform multinomial sampling according to the calculated posterior distribution to obtain a new travel mode category;

[0034] S324: updating the travel mode category of the current trip word to the newly extracted travel mode category;

[0035] Repeat step S32 until the rate of change of the parameters in the “travel mode-spatial-temporal characteristics” distribution and the “vehicle-travel mode” distribution tends to be stable, and the model can be considered to have reached a convergence state;

[0036] S33: Based on the samples, estimate the “vehicle-travel mode” distribution and the “travel mode-spatiotemporal characteristics” distribution.

[0037] The beneficial effects of the present invention are:

[0038] The proposed ST-LDA-based method for mining vehicle travel patterns is superior to traditional methods. Compared with existing technologies, this method starts from the travel behavior of individual vehicles and models their travel behavior as a polynomial distribution of travel patterns, which reflects the diversity of individual travel behaviors. Travel patterns are also modeled as a polynomial distribution of spatiotemporal features. This semantic-level feature not only mines vehicle travel patterns from both temporal and spatial dimensions but also increases the model's interpretability. The proposed ST-LDA model can discover potential travel patterns within a travel corpus and has broad application prospects.

[0039] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0041] Figure 1 Schematic diagram of the ST-LDA model structure;

[0042] Figure 2 Schematic diagram of the perplexity change of the ST-LDA model under different numbers of travel mode categories;

[0043] Figure 3 Schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0044] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0045] The present invention provides a method and system for mining vehicle spatiotemporal travel patterns based on ST-LDA. Figure 3 This is a flow chart of the method of the present invention. As shown in the figure, the method includes the following steps: S1, data preparation: extracting the vehicle's highway trip set and converting it into a trip corpus; S2, constructing the ST-LDA model: the ST-LDA model includes two parts: a polynomial distribution between travel patterns and spatiotemporal features, which increases the interpretability of the model from the semantic feature level; and a polynomial distribution between individual vehicles and travel patterns, which well reflects the diverse travel behaviors of individual vehicles; S3, model training: using the obtained trip words as input to train the ST-LDA model, and using the folded Gibbs sampling method to solve the latent variables in the model to obtain travel pattern recognition results; S4, using the model to mine vehicle spatiotemporal travel patterns.

[0046] In this example, the data used is Chongqing Expressway exit traffic data. The time range of the data is July 5, 2021, to July 25, 2021, and the spatial scope of the data is the Chongqing metropolitan area. The subjects of the study are ordinary Class A passenger vehicles with at least three trips. A total of 10,559,698 trip records were obtained for 1,306,003 vehicles. Figure 1 Schematic diagram of the ST-LDA model structure.

[0047] Step 1: Data preparation: Extract the vehicle's highway trip set and convert it into a trip word set.

[0048] Step 11: A highway trip record can be expressed as Tr = (cid, tid o , t o , tid e , t d ), cid indicates the vehicle to which the record belongs, tid o and t o They correspond to the vehicle's entry point and entry time, tid d and t d The vehicle's individual Car can be obtained based on cid. i All highway trip records within a given time period are sorted in ascending order according to the entry time, and the Car i Itinerary Collection in Car i The j-th highway trip record, N i Represents Car i The number of trip records in a given time period.

[0049] According to the CID field, extract all highway travel records of the vehicle in a given period and sort them by entry time TID o The fields are sorted in ascending order to obtain the vehicle individual travel record set (as shown in Table 1). i Take the itinerary record of as an example.

[0050] Table 1 Car i Example of itinerary record (partial)

[0051] Serial number <![CDATA[tid o ]]> <![CDATA[t o ]]> <![CDATA[tid d ]]> <![CDATA[t d ]]> 1 50013610 2021-07-05 08：05：53 50015620 2021-07-05 08：18：53 2 50013620 2021-07-05 18：25：13 50013610 2021-07-05 18：38：26 3 50013610 2021-07-06 08：02：35 50015620 2021-07-06 08：15：51

[0052] Step 12: For Car i Itinerary Set Trs i The itinerary in the , convert it into itinerary words. Specifically including:

[0053] (1) Get the travel time interval label e of the travel word. i Itinerary Set Trs i Itinerary The next adjacent trip is Travel time interval In units of hours, convert it into one of the categories of {"[0, 3)","[3, 7)","[7, 11)","[11, 15)","[15, 24)","[24, 48)","[48, 72)","[72, ∞)"} to obtain the travel time interval label e of the travel word.

[0054] (2) Get the trip end tag r of the trip word. i Itinerary Set Trs i Itinerary The label r of the end point of the trip is The terminal toll station ID, tid d .

[0055] (3) Get the departure time tag t of the trip word. i Itinerary Set Trs i Itinerary The end label t of its trip word is Departure time t o hours.

[0056] (4) Get the day of the week label q of the trip word. i Itinerary Set Trs i Itinerary The end label q of its trip word is Days of the week.

[0057] For Car i , the first trip occurs at 08:05 on Monday, and the interval between it and the next trip is 10 hours and 6 minutes. The trip word corresponding to the first trip is Car i All of a person's itineraries constitute their itinerary document for The corresponding itinerary word, is the number of trip words, The trip word set of all vehicles in the research dataset constitutes the trip corpus Trc = {Trd1, Trd2, ..., Trd M}, Trd i Represents Car i The itinerary document is , where M is the number of vehicles.

[0058] Since the trip interval of an individual vehicle's last trip is unknown, the last trip of an individual vehicle is not considered when constructing trip words. The trip corpus obtained in step 1 finally contains 9,253,695 trip words, involving 1,306,003 vehicles.

[0059] In step 2, the ST-LDA model assumes that the itinerary corpus Trc is generated by the following steps:

[0060] Step 21: For a travel mode of category z (z=1, 2, .., Z), generate a multinomial distribution of the corresponding trip word labels from the Dirichlet distribution. Specifically, it includes:

[0061] (1) Using β as a hyperparameter, generate the “travel mode-trip time distribution” of category z from the corresponding Dirichlet distribution.

[0062] (2) Using γ as a hyperparameter, generate the “travel mode-trip time distribution” of category z from the corresponding Dirichlet distribution.

[0063] (3) Using τ as a hyperparameter, generate the “travel mode-trip time distribution” π of category z from the corresponding Dirichlet distribution z .

[0064] (4) Using κ as a hyperparameter, generate the “travel mode-trip time distribution” η for category z from the corresponding Dirichlet distribution z .

[0065] Repeat step 21 z times to obtain the corresponding spatiotemporal characteristic distribution of all travel mode categories.

[0066] Step 22: Sampling to get Cari Trd i . Specifically including:

[0067] (1) Using α as a hyperparameter, generate Car from the corresponding Dirichlet distribution i The multinomial distribution of travel modes θ i .

[0068] (2) According to Car i The multinomial distribution of travel modes θ i , from which a travel mode category z is sampled. The corresponding travel word label is sampled from the spatiotemporal distribution of the travel mode of category z.

[0069] From the multinomial distribution of travel time intervals of category z A travel time interval category e is obtained by sampling.

[0070] From the multinomial distribution of trip endpoints with category z A travel time interval category r is obtained by sampling.

[0071] The departure time of a trip of category z is multinomially distributed π z A travel time interval category t is obtained by sampling.

[0072] From the multinomial distribution η of the days of the week for trips of category z z A travel time interval category q is obtained by sampling.

[0073] In step 3, the model solving steps can be divided into the following steps:

[0074] Step 31: Randomly initialize the travel mode category to which the travel words in the travel corpus Trc belong.

[0075] Step 32: For Car in the itinerary corpus i The jth trip word

[0076] Step 321: Remove the spatiotemporal feature category label corresponding to the word from the travel mode of the current trip word.

[0077] Step 322: Calculate the trip word given other trip words and travel mode categories The posterior probability of a travel mode belonging to category z is:

[0078]

[0079] Among them, the superscript Indicates that it is not considered in the calculation The travel mode category or characteristic value of hiz Represents Car i The number of trip words under travel mode z, v ze represents the number of travel words belonging to travel mode z in the travel words with a time interval of e, w zr represents the number of trip words with destination r that belong to travel mode z, c zt represents the number of travel words with departure time t that belong to travel mode z, x zq represents the number of trip words belonging to travel mode z among the trip words with day q of the week.

[0080] Step 323: Perform multinomial sampling according to the calculated posterior distribution to obtain a new travel mode category.

[0081] Step 324: Update the travel mode category of the current trip word to the newly extracted travel mode category.

[0082] Repeat step 32 until the rate of change of the parameters in the “travel mode-spatial-temporal characteristics” distribution and the “vehicle-travel mode” distribution tends to be stable, and the model can be considered to have reached convergence.

[0083] Step 33: Based on the samples, estimate the “vehicle-travel mode” distribution and the “travel mode-spatiotemporal characteristics” distribution.

[0084] More specifically, step 33 can be divided into the following steps:

[0085] Step 331: Car i The probability of exhibiting a travel mode of category z

[0086] Step 332: The probability of the time interval being e under the travel mode of category z The probability of the end point being r The probability of departure time t The probability that the day of the week is q

[0087] Randomly sample 5,000 trip words from the trip corpus as the training set, and the rest as the training set. When training the model in step 3, set the parameters α = 50 / Z, β = 0.01, γ = 0.01, τ = 0.01, and κ = 0.01 based on experience. Sampling is performed by folding Gibbs, and the perplexity is used as an evaluation indicator. The lower the perplexity, the better the model fits the data and has better interpretability. The perplexity calculation formula is Where Trc test is the test set, M is the number of vehicles in the test set, For Car i The number of words in the itinerary, is the likelihood of observing the trip,

[0088] like Figure 2 , shows the change of perplexity under different numbers of travel mode categories. When the number of travel mode categories Z = 14, the perplexity reaches an optimal value.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A vehicle spatiotemporal travel pattern mining method based on ST-LDA, characterized by: The method comprises the following steps: S1. Data preparation: Extract the vehicle's highway trip set and convert it into a trip corpus; S2. Constructing the ST-LDA model: The ST-LDA model consists of two parts: a polynomial distribution between travel patterns and spatiotemporal features, which increases the interpretability of the model at the semantic feature level; and a polynomial distribution between individual vehicles and travel patterns, which reflects the diverse travel behaviors of individual vehicles. S3. Model training: The obtained itinerary words are used as input to train the ST-LDA model, and the latent variables in the model are solved using the folded Gibbs sampling method to obtain the travel mode recognition results; S4. Use the model to mine vehicle spatiotemporal travel patterns; In step S1, the highway trip set of the vehicle is extracted and converted into a trip corpus, which specifically includes: S11: Extracting travel records for a specified time period on the expressway, integrating the travel records of individual vehicles according to their vehicle identification IDs, and arranging them in ascending order according to their entry time to obtain a set of individual vehicle travel records; S12: Convert each trip of each vehicle into trip words. After the conversion is completed, all trip words of a vehicle constitute a trip document, and the trip documents of all vehicles constitute a trip corpus. In step S12, converting the itinerary into itinerary words specifically includes: (1): Get the travel time interval label e of the travel word; (2): Get the trip end label r of the trip word; (3): Get the departure time tag t of the trip word; (4): Get the day of the week label q of the itinerary word; In step S2, the ST-LDA model assumes that the itinerary corpus is generated by the following steps: S21: For each travel mode of category z, z = 1, 2, .., Z, generate a multinomial distribution of the corresponding trip word labels from the Dirichlet distribution; Repeat step S21 Z times to obtain the spatiotemporal feature distribution of all travel mode categories. S22: Sampling to get Car i Travel documents; The step S22 samples and obtains Car i The itinerary documents specifically include: S221: Generate Car from Dirichlet Distribution i Multinomial distribution of travel modes; S222: According to Car i The multinomial distribution of travel modes is used to sample a travel mode category z, and the corresponding travel word label is sampled from the spatiotemporal distribution of travel under the travel mode of category z.

2. The method for mining spatiotemporal travel patterns of vehicles based on ST-LDA according to claim 1, characterized in that: In step S3, the model solving steps are divided into the following steps: S31: Randomly initialize the travel mode category to which the travel words in the travel corpus belong; S32: For Car in the itinerary corpus i Itinerary words: S321: removing the spatiotemporal feature category label corresponding to the word from the travel mode of the current trip word; S322: Calculate the posterior probability of the trip word belonging to the travel mode of category z when other trip words and travel mode categories are given; S323: Perform multinomial sampling according to the calculated posterior distribution to obtain a new travel mode category; S324: updating the travel mode category of the current trip word to the newly extracted travel mode category; Repeat step S32 until the rate of change of the parameters in the "travel mode-spatial characteristics" distribution and the "vehicle-travel mode" distribution tends to be stable, and the model is considered to have reached a convergence state; S33: Estimate the distribution of "vehicle-travel mode" and "travel mode-spatiotemporal characteristics" based on the sample.

3. A vehicle spatiotemporal travel pattern mining system based on ST-LDA, characterized by: The system uses the method described in any one of claims 1 to 2 to mine vehicle spatiotemporal travel patterns.

Images