Deep learning-based intelligent travel scenic area passenger flow prediction method

Abstract

The invention discloses a deep learning-based smart travel scenic area passenger flow prediction method. The method comprises the steps of obtaining a preliminary multi-scale distribution feature set; obtaining an enhanced intrinsic correlation feature vector; obtaining an initial dynamic grouping result set; obtaining a mode conversion tracking sequence; a re-clustering mechanism is applied to the recognized breaking points to fuse the internal association feature vectors, and an updated dynamic grouping system is obtained; the updated dynamic grouping system is matched with the real-time passenger flow distribution data, whether the dynamic change triggers a grouping adjustment threshold value or not is judged, and an adaptive grouping updating signal is obtained; and a final tourist group division structure is generated according to the adaptive grouping update signal, the fusion mode conversion tracking sequence and the inherent association feature vector, and a passenger flow management system data set is obtained. According to the method, the problems that a traditional method is difficult to process group behavior mutation, holiday behavior drift and group boundary instability are solved from the algorithm level.

Description

Technical Field

[0001] This invention relates to the field of smart cultural tourism technology, and in particular to a method for predicting visitor flow in smart cultural tourism scenic spots based on deep learning. Background Technology

[0002] With the continuous advancement of smart tourism and smart scenic area construction, more and more scenic areas are beginning to utilize big data, IoT, and AI technologies to dynamically manage and accurately predict tourist behavior and visitor flow. Current visitor flow prediction and visitor group segmentation technologies commonly employ historical visitor flow data statistics, sensor data monitoring, and simple time-series analysis models to analyze trends in visitor numbers. These methods primarily use fixed-granularity hourly, daily, and monthly data to statistically model visitor flow in and out of the scenic area, using rule-based algorithms or conventional machine learning models to predict and warn of visitor flow trends. However, existing technologies have significant shortcomings in the following aspects: Existing methods primarily rely on single-scale time-series analysis or static features, making it difficult to fully explore the multi-scale distribution characteristics of tourist flows. They also lack in-depth modeling of the nesting and correlation relationships between different time granularities (hours, days, months), resulting in limited understanding of dynamic changes in tourist flows. Research on the heterogeneity of tourist group behavior and intra-group pattern transitions is relatively weak, lacking a detailed characterization of the transition paths between different tourist commuting and roaming behavior patterns, making it impossible to dynamically and accurately identify and classify multiple types of tourist groups. Most existing clustering methods rely only on single-moment or static features, lacking dynamic grouping based on time-series features and tracking of behavioral pattern evolution, resulting in group segmentation results that fail to reflect real-time changes and complex behavioral migrations. Summary of the Invention

[0003] One objective of this invention is to propose a method for predicting visitor flow in smart cultural tourism scenic spots based on deep learning. This invention solves the problems of traditional methods being unable to handle sudden changes in group behavior, holiday behavior drift, and unstable group boundaries from an algorithmic perspective.

[0004] According to an embodiment of the present invention, a deep learning-based method for predicting visitor flow in smart cultural tourism scenic areas includes: By collecting hourly, daily, and monthly visitor entry and exit data and location sequences from sensors and motion track recorders at the scenic area entrance, a preliminary set of multi-scale distribution features was obtained. Based on the obtained multi-scale distribution feature set, the nesting characteristics between hourly, daily, and monthly sequences are decomposed using time series analysis. For each scale, the distribution similarity index is calculated. The nesting association strength is judged by comparing the similarity index with a preset threshold, and the enhanced intrinsic association feature vector is obtained. If the similarity index in the enhanced intrinsic association feature vector exceeds the preset threshold, the corresponding intrinsic association feature vector is input into the hierarchical clustering algorithm for preliminary tourist group segmentation. Clustering iteration is performed based on weekend behavior differences to obtain an initial dynamic grouping result set. Obtain a subset of pattern transitions from the initial dynamic grouping result set. By tracking the migration path from commuting to roaming behavior in the subset, use time series analysis to quantify the transition rate and obtain the pattern transition tracking sequence. If the mode transition tracking sequence shows that the transition rate is higher than the preset threshold, the group affiliation label in the initial dynamic grouping result set is adjusted, and the re-clustering mechanism is applied to the identified breakpoint to fuse the intrinsic correlation feature vectors to obtain an updated dynamic grouping system. By matching the updated dynamic grouping system with real-time passenger flow distribution data, it is determined whether dynamic changes trigger the grouping adjustment threshold, thus obtaining an adaptive grouping update signal. Based on the adaptive grouping update signal, the fusion mode transformation tracking sequence and intrinsic correlation feature vector are used to generate the final tourist group segmentation structure, thus obtaining the passenger flow management system dataset.

[0005] Optionally, the step of using time series analysis to decompose the nesting characteristics between hourly, daily, and monthly sequences includes: The time-series decomposition method is used to process the multi-scale distribution feature set to obtain hourly, daily, and monthly sequences. Distribution similarity was calculated for hourly, daily, and monthly series respectively, resulting in three sets of distribution similarity values; The three sets of distribution similarity values ​​are compared with the preset similarity threshold. If the distribution similarity is greater than the similarity threshold, it is determined that there is a strong nested association between the corresponding scales. If the distribution similarity is less than or equal to the similarity threshold, the nested association between the corresponding scales is determined to be weak, and the set of nested association strengths for each scale is obtained. The original multi-scale distribution feature set is weighted and adjusted based on the nested correlation strength set of each scale. Scales with high correlation strength are given high weights and scales with low correlation strength are given low weights, resulting in the adjusted distribution feature set. The k-means clustering method is used to process the adjusted distribution feature set to obtain the cluster center set. The cluster center set and the adjusted distribution feature set are then concatenated as vectors to obtain the enhanced intrinsic correlation feature vector.

[0006] Optionally, if the similarity index in the enhanced intrinsic association feature vector exceeds a preset threshold, the corresponding intrinsic association feature vector is input into a hierarchical clustering algorithm for preliminary tourist group segmentation, including: By detecting the similarity index of the enhanced intrinsic correlation feature vectors, if the detection result exceeds the preset threshold, it is included in the dataset to be processed, and an initial feature set that meets the conditions is obtained. Based on the initial feature set that meets the conditions, a hierarchical clustering algorithm is used to initially divide the tourist groups. The feature vectors of each group are classified and labeled to obtain the preliminary group categories. For the initially classified groups, weekend behavior data from the pattern transition signals are obtained, the significant characteristics of behavioral differences are analyzed, and the key points of behavioral differences are determined. If the key points of behavioral differences meet the preset significance conditions, then perform clustering iteration on the relevant group categories and update the intermediate results of dynamic grouping; Based on the intermediate results of dynamic grouping, the core behavioral patterns of each group are extracted, and a second comparison is performed in conjunction with the pattern transition signal to determine the stability of the grouping. Based on the assessment of group stability, unstable group categories are adjusted, and outliers are removed using data cleaning tools to obtain an initial dynamic grouping result set.

[0007] Optionally, the migration path from commuting to roaming by tracking behavioral characteristics in a subset includes: By filtering out the user set that includes the transition from commuting mode to roaming mode from the initial dynamic grouping result set using preset rules, a mode conversion subset is obtained; For the pattern conversion subset, all user behavior records are arranged in chronological order to obtain a behavior sequence arranged in chronological order. Extract consecutive commuting mode segments and roaming mode segments from the time-ordered behavior sequence to determine the switching point position each time from commuting mode to roaming mode; Based on the determined switching point location, calculate the time interval between two adjacent switching points to obtain the time span data for each mode transition; The time span data of all transformations are processed using time series analysis methods to calculate the rate value at the time of each transformation, thus obtaining the transformation rate sequence; The conversion rate sequence is rearranged according to the original user time order to obtain the mode conversion tracking sequence for each user.

[0008] Optionally, the step of applying a re-clustering mechanism to fuse intrinsic correlation feature vectors for identifying breakpoints to obtain an updated dynamic grouping system includes: Acquire the mode transition tracking sequence and determine whether the transition rate is higher than a preset threshold; If the conversion rate is higher than the preset threshold, the group affiliation label of the initial dynamic grouping result set is extracted; The tracking sequence is scanned segment by segment based on the group affiliation label to determine the location of the breakpoint in the continuity of the tracking sequence; For each breakpoint, the intrinsic correlation feature vectors on both sides are collected. The k-means clustering algorithm is used to perform re-clustering on the intrinsic correlation feature vectors collected before and after the breakpoint, resulting in a new set of group affiliation labels after re-clustering. The corresponding labels in the initial dynamic grouping result set are replaced with the new set of group affiliation labels to form a dynamic grouping system that includes the updated group affiliation labels.

[0009] Optionally, the step of matching the updated dynamic grouping system with real-time passenger flow distribution data to determine whether dynamic changes trigger a grouping adjustment threshold includes: Passenger flow distribution data is acquired through a real-time data acquisition system. The acquired data is then preliminarily cleaned and formatted to obtain a standardized passenger flow distribution dataset. Using a standardized passenger flow distribution dataset and a pre-established dynamic grouping system, a data matching operation is performed to determine the correspondence between the current passenger flow distribution and the grouping system; Based on the data matching results, analyze the dynamic changes in passenger flow distribution. If the dynamic changes exceed the preset adjustment threshold, generate a preliminary grouping adjustment demand signal. For the initial grouping adjustment demand signal, historical grouping adjustment records and real-time environmental data are obtained. Logical comparison is used to determine whether the adaptive update conditions are met, and an adaptive grouping update signal is obtained.

[0010] Optionally, the step of generating the final tourist group segmentation structure by fusing the mode transition tracking sequence and intrinsic correlation feature vector based on the adaptive grouping update signal includes: An adaptive grouping mechanism is used to dynamically cluster the adaptive grouping update signal to generate a preliminary tourist group classification set. Based on the preliminary tourist group classification set, information related to pattern transformation is obtained, and behavioral pattern analysis is performed over time using tracking sequence data to determine the dynamic changing trend of tourist groups. By extracting intrinsic correlation information from dynamic trends and combining it with feature vectors, a multi-dimensional profile of tourist groups is constructed, resulting in refined tourist group segmentation. Based on the refined tourist group segmentation results, a pre-established decision tree model is used to structure the segmentation results and generate the initial framework of the passenger flow management system. Based on the initial framework, combined with passenger flow data and evolvability requirements, the management system is dynamically adjusted. If the response speed of the management system is lower than the preset threshold, the response capability is improved by adding data processing nodes, thus obtaining the passenger flow management system dataset.

[0011] The beneficial effects of this invention are: In the process of passenger flow data processing, this invention introduces a nested correlation analysis mechanism of hourly, daily, and monthly multi-scale distribution features. By quantitatively calculating the distribution similarity between sequences at different time scales and combining it with thresholds to determine the nested correlation strength, it realizes dynamic weight adjustment of multi-scale features. It can identify the degree of intrinsic correlation between different time scales and focus on strengthening highly correlated scales in the feature space, making the generated intrinsic correlation feature vectors more compact in structure and more discriminative. This effectively alleviates the problems of scale conflict, information redundancy, and feature dilution in traditional multi-scale modeling.

[0012] In the process of segmenting tourist groups, this invention not only utilizes hierarchical clustering for initial grouping, but also further tracks and analyzes the migration paths of tourist behavior from commuting to roaming patterns. By quantifying the transformation rate over time, it identifies breakpoints in group behavior and triggers a re-clustering mechanism when a high transformation rate is detected, dynamically correcting the original group affiliation. Compared to static clustering or grouping methods based solely on single feature updates in existing technologies, this invention explicitly introduces behavioral pattern changes into the group segmentation logic, so that the clustering results are no longer limited to the similarity at a certain time segment, but reflect the continuous process of tourist behavior evolution over time. From an algorithmic perspective, it solves the problems of traditional methods struggling to handle sudden changes in group behavior, holiday behavior drift, and unstable group boundaries.

[0013] This invention introduces an adaptive grouping update signal generation mechanism based on real-time passenger flow distribution matching, which is based on dynamic grouping results. By comprehensively analyzing real-time data changes, historical grouping adjustment records, and environmental factors, it determines whether to trigger an update of the grouping structure. During the update process, it integrates the mode conversion tracking sequence and the intrinsic correlation feature vector to form the final tourist group division structure. By using a signal-driven approach to constrain grouping adjustments, the group structure update has clear triggering conditions and continuous evolution logic, avoiding over-adjustment caused by short-term fluctuations. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a smart cultural tourism scenic area visitor flow prediction method based on deep learning proposed in this invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0016] refer to Figure 1 As shown in Example 1: A method for predicting visitor flow in smart cultural tourism scenic spots based on deep learning, the method includes: By collecting hourly, daily, and monthly visitor entry and exit data and location sequences from sensors and motion track recorders at the scenic area entrance, a preliminary set of multi-scale distribution features was obtained. Based on the obtained multi-scale distribution feature set, the nesting characteristics between hourly, daily, and monthly sequences are decomposed using time series analysis. For each scale, the distribution similarity index is calculated. The nesting association strength is judged by comparing the similarity index with a preset threshold, and the enhanced intrinsic association feature vector is obtained. If the similarity index in the enhanced intrinsic association feature vector exceeds the preset threshold, the corresponding intrinsic association feature vector is input into the hierarchical clustering algorithm for preliminary tourist group segmentation. Clustering iteration is performed based on weekend behavior differences to obtain an initial dynamic grouping result set. Obtain a subset of pattern transitions from the initial dynamic grouping result set. By tracking the migration path from commuting to roaming behavior in the subset, use time series analysis to quantify the transition rate and obtain the pattern transition tracking sequence. If the mode transition tracking sequence shows that the transition rate is higher than the preset threshold, the group affiliation label in the initial dynamic grouping result set is adjusted, and the re-clustering mechanism is applied to the identified breakpoint to fuse the intrinsic correlation feature vectors to obtain an updated dynamic grouping system. By matching the updated dynamic grouping system with real-time passenger flow distribution data, it is determined whether dynamic changes trigger the grouping adjustment threshold, thus obtaining an adaptive grouping update signal. Based on the adaptive grouping update signal, the fusion mode transformation tracking sequence and intrinsic correlation feature vector are used to generate the final tourist group segmentation structure, thus obtaining the passenger flow management system dataset.

[0017] In this embodiment, the decomposition of the nesting characteristics between hourly, daily, and monthly sequences using time series analysis methods includes: The time-series decomposition method is used to process the multi-scale distribution feature set to obtain hourly, daily, and monthly sequences. Distribution similarity was calculated for hourly, daily, and monthly series respectively, resulting in three sets of distribution similarity values; The three sets of distribution similarity values ​​are compared with the preset similarity threshold. If the distribution similarity is greater than the similarity threshold, it is determined that there is a strong nested association between the corresponding scales. If the distribution similarity is less than or equal to the similarity threshold, the nested association between the corresponding scales is determined to be weak, and the set of nested association strengths for each scale is obtained. The original multi-scale distribution feature set is weighted and adjusted based on the nested correlation strength set of each scale. Scales with high correlation strength are given high weights and scales with low correlation strength are given low weights, resulting in the adjusted distribution feature set. The k-means clustering method is used to process the adjusted distribution feature set to obtain the cluster center set. The cluster center set and the adjusted distribution feature set are then concatenated as vectors to obtain the enhanced intrinsic correlation feature vector.

[0018] In this embodiment, if the similarity index in the enhanced intrinsic association feature vector exceeds a preset threshold, the corresponding intrinsic association feature vector is input into a hierarchical clustering algorithm for preliminary tourist group segmentation, including: By detecting the similarity index of the enhanced intrinsic correlation feature vectors, if the detection result exceeds the preset threshold, it is included in the dataset to be processed, and an initial feature set that meets the conditions is obtained. Based on the initial feature set that meets the conditions, a hierarchical clustering algorithm is used to initially divide the tourist groups. The feature vectors of each group are classified and labeled to obtain the preliminary group categories. In Example 1, based on the initial feature set that meets the conditions, a hierarchical clustering algorithm is used to perform similarity calculation on the input feature vector set, gradually merging feature vectors that are close to each other to form a cluster tree. According to the preset number of categories or clustering threshold, the cluster tree is cut into several sub-groups, and all feature vectors in each sub-group are assigned the same category label. Finally, the preliminary category division results and corresponding feature vector classification labels of each tourist group are obtained, and the preliminary group categories are determined.

[0019] If the key points of behavioral differences meet the preset significance conditions, then perform clustering iteration on the relevant group categories and update the intermediate results of dynamic grouping; Based on the intermediate results of dynamic grouping, the core behavioral patterns of each group are extracted, and a second comparison is performed in conjunction with the pattern transition signal to determine the stability of the grouping. In Example 1, based on the intermediate results of dynamic grouping, the behavioral records of each group member are aggregated. A frequent sequence pattern mining algorithm is used to statistically analyze the dominant behavioral sequences of each member under different time periods, regions, and activity categories, obtaining the core behavioral patterns characterizing the group's behavioral features. The core behavioral patterns of each group are then compared twice with the behavioral sequences identified in the pattern transition signal using the Jaccard similarity coefficient. If the similarity coefficient is lower than a preset threshold, the group is considered to have significant pattern fluctuations and is deemed unstable, requiring adjustment. If the similarity coefficient is higher than the threshold, the group's stability is deemed satisfactory.

[0020] Based on the assessment of group stability, unstable group categories are adjusted, and outliers are removed using data cleaning tools to obtain an initial dynamic grouping result set.

[0021] In this embodiment, the migration path from commuting to roaming by tracking behavioral characteristics in a subset includes: By filtering out the user set that includes the transition from commuting mode to roaming mode from the initial dynamic grouping result set using preset rules, a mode conversion subset is obtained; For the pattern conversion subset, all user behavior records are arranged in chronological order to obtain a behavior sequence arranged in chronological order. Extract consecutive commuting mode segments and roaming mode segments from the time-ordered behavior sequence to determine the switching point position each time from commuting mode to roaming mode; Based on the determined switching point location, calculate the time interval between two adjacent switching points to obtain the time span data for each mode transition; The time span data of all transformations are processed using time series analysis methods to calculate the rate value at the time of each transformation, thus obtaining the transformation rate sequence; The conversion rate sequence is rearranged according to the original user time order to obtain the mode conversion tracking sequence for each user.

[0022] In this embodiment, the step of applying a re-clustering mechanism to fuse intrinsic correlation feature vectors for identifying breakpoints to obtain an updated dynamic grouping system includes: Acquire the mode transition tracking sequence and determine whether the transition rate is higher than a preset threshold; If the conversion rate is higher than the preset threshold, the group affiliation label of the initial dynamic grouping result set is extracted; The tracking sequence is scanned segment by segment based on the group affiliation label to determine the location of the breakpoint in the continuity of the tracking sequence; For each breakpoint, the intrinsic correlation feature vectors on both sides are collected. The k-means clustering algorithm is used to perform re-clustering on the intrinsic correlation feature vectors collected before and after the breakpoint, resulting in a new set of group affiliation labels after re-clustering. In Example 1, for each break point location, a preset number of intrinsic correlation feature vectors are collected before and after the break point to form a feature vector set before the break point and a feature vector set after the break point. After normalizing the feature vectors in the set and removing outliers, the feature vector sets on both sides are merged into a local feature vector set. The k-means clustering algorithm is used for re-clustering. A new group affiliation label set is generated based on the clustering results, and this label set is used to replace the corresponding members in the initial dynamic grouping result set, thereby realizing the adaptive update of the group affiliation labels before and after the break point.

[0023] The corresponding labels in the initial dynamic grouping result set are replaced with the new set of group affiliation labels to form a dynamic grouping system that includes the updated group affiliation labels.

[0024] In Example 1, for each breakpoint location, the intrinsic correlation feature vectors on both sides are collected. Using the user ID as an index, the k-means clustering algorithm is used to perform re-clustering on the intrinsic correlation feature vectors collected before and after the breakpoint, obtaining a new group affiliation label for each sample in the re-clustering results, forming a new set of group affiliation labels. The new set of group affiliation labels replaces the corresponding labels in the initial dynamic grouping result set one by one according to the user ID or index, completing the update of the dynamic grouping system, and finally forming a dynamic grouping system including the updated group affiliation labels.

[0025] In this embodiment, the step of matching the updated dynamic grouping system with real-time passenger flow distribution data to determine whether dynamic changes trigger a grouping adjustment threshold includes: Passenger flow distribution data is acquired through a real-time data acquisition system. The acquired data is then preliminarily cleaned and formatted to obtain a standardized passenger flow distribution dataset. Using a standardized passenger flow distribution dataset and a pre-established dynamic grouping system, a data matching operation is performed to determine the correspondence between the current passenger flow distribution and the grouping system; In Example 1, a standardized passenger flow distribution dataset is used, combined with a pre-established dynamic grouping system, to perform a data matching operation. The data matching operation includes: calculating a similarity index between the real-time collected passenger flow distribution data and the historical distribution feature vectors of each group in the dynamic grouping system. The similarity index is cosine similarity, Euclidean distance, DTW distance, etc. The group with the highest similarity index (or the smallest distance) is selected as the group to which the current passenger flow distribution belongs, thus determining the correspondence between the current passenger flow distribution and the grouping system.

[0026] Based on the data matching results, analyze the dynamic changes in passenger flow distribution. If the dynamic changes exceed the preset adjustment threshold, generate a preliminary grouping adjustment demand signal. For the initial grouping adjustment demand signal, historical grouping adjustment records and real-time environmental data are obtained. Logical comparison is used to determine whether the adaptive update conditions are met, and an adaptive grouping update signal is obtained.

[0027] In this embodiment, the step of generating the final tourist group segmentation structure by fusing the mode transition tracking sequence and intrinsic correlation feature vector based on the adaptive grouping update signal includes: An adaptive grouping mechanism is used to dynamically cluster the adaptive grouping update signal to generate a preliminary tourist group classification set. In Example 1, when the adaptive grouping mechanism is adopted, tourist feature data with significant recent behavioral changes are selected from the current dynamic grouping system based on the adaptive grouping update signal. Combined with the pattern transformation tracking sequence and intrinsic correlation feature vector, a dynamic clustering input dataset is formed. The input dataset is then fed into the incremental k-means clustering algorithm based on the time window to dynamically cluster the tourist group. During the clustering process, the number and boundaries of groups are automatically adjusted, and a preliminary tourist group classification set is output.

[0028] Based on the preliminary tourist group classification set, information related to pattern transformation is obtained, and behavioral pattern analysis is performed over time using tracking sequence data to determine the dynamic changing trend of tourist groups. By extracting intrinsic correlation information from dynamic trends and combining it with feature vectors, a multi-dimensional profile of tourist groups is constructed, resulting in refined tourist group segmentation. Based on the refined tourist group segmentation results, a pre-established decision tree model is used to structure the segmentation results and generate the initial framework of the passenger flow management system. In Example 1, a pre-established decision tree model is used to structure the partitioning results, specifically including: The core features of each group in the refined tourist group segmentation results (including behavioral patterns, activity time periods, spatial distribution, and group size) are encoded into feature vectors to form the model input dataset; The model input dataset is fed into a pre-trained decision tree model. Based on the model node judgment rules, the tourist groups are categorized and management measures are suggested, forming a set of structured management suggestions. Based on the set of structured management suggestions, an initial framework for a passenger flow management system is generated. The initial framework includes group management nodes, early warning threshold parameters, and dynamic adjustment strategies.

[0029] Based on the initial framework, combined with passenger flow data and evolvability requirements, the management system is dynamically adjusted. If the response speed of the management system is lower than the preset threshold, the response capability is improved by adding data processing nodes, thus obtaining the passenger flow management system dataset.

[0030] Example 2: In a large-scale smart scenic area operation system, the system collected the following sample data (partial examples) through entrance sensors and Wi-Fi probes over a continuous 72-hour period: From 08:00 to 09:00, 1,432 people entered the park, 87 people left the park, and the average stay time was 8.7 minutes. From 09:00 to 10:00, 2,851 people entered the park, 245 people left the park, and the average stay time was 21.3 minutes. From 10:00 to 11:00, 4,326 people entered the park, 1,278 people left the park, and the average stay was 37.5 minutes. Daily cumulative total: 23,541 people entered the park, and 22,283 people left the park. Monthly cumulative: There were 14 peak days (daily passenger flow > 20,000) and 8 low days (less than 8,000).

[0031] The system synchronously collects visitor movement trajectory sequences. Each sequence includes: entry time, time spent in each area, movement path length, and number of stops. In Example 2, the trajectory sequence of one visitor is as follows: [Entry 08:42] → [12-minute stop at the main square] → [18-minute stop at Hall A] → [27-minute stop at the food area] → [Exit 10:12]; The data constitutes a multi-scale distribution feature set. The system employs a time-series decomposition algorithm to divide the original data into hourly, daily, and monthly sequences. In Example 2, the hourly sequence for a certain day is as follows: {08:00, 1432}, {09:00, 2851}, {10:00, 4326}, … {20:00, 687}; Daily sequence: {Day1, 23874}, {Day2, 18553}, {Day3, 22147}, …; Calculate the distribution similarity for each sequence group. Using the Pearson correlation coefficient and Dynamic Time Warping (DTW) algorithm, we obtain: Hourly to daily sequence similarity: 0.88; Daily to monthly sequence similarity: 0.69; Hourly to monthly sequence similarity: 0.66.

[0032] The system's preset distribution similarity threshold is 0.70. Therefore, there is a strong nested relationship between hourly and daily data, and a weak relationship between daily and monthly data.

[0033] Based on the distribution similarity results, the system performs weighted adjustments on the multi-scale feature set. The weights are set as follows: Hourly weight: 0.45; Daily weight: 0.40; Monthly weight: 0.15.

[0034] After weighting, the adjusted distribution feature set is more sensitive to hourly fluctuations and can reflect peak / abnormal periods more quickly.

[0035] The adjusted distribution feature set is input into the k-means algorithm, with the number of cluster centers set to 4. The cluster center vector is then calculated. C1 (Commuter type): {Average stay time: 16.4 minutes, peak entry time 08:00-10:00, average trajectory length: 510 meters}; C2 (Family Leisure Type): {Average stay time: 64.2 minutes, peak entry time 09:30-12:30, average trajectory length: 1450 meters}; C3 (roaming type): {Average dwell time: 135.5 minutes, present throughout the day, average trajectory length: 2850 meters}; C4 (Group Study Tour): {Average stay time: 51.1 minutes, concentrated between 11:00 and 14:00; Average track length: 920 meters}; The enhanced intrinsic correlation feature vectors are constructed using vector concatenation, specifically: [Stay time, peak entry time, average trajectory length, ...] splice together [cluster center 1, cluster center 2, ...]; Subsequent testing of the similarity index revealed that the similarity between C3 and C2 was 0.81, exceeding the threshold, requiring further subdivision.

[0036] The system enhances the hierarchical clustering of input features to obtain preliminary tourist group segmentation results. Combined with weekend data, the system detected a high variance in the internal distribution of group C3 (wandering type) and significant differences in weekend behavior, automatically triggering clustering iterations. After the clustering iterations, C3 was further subdivided into C3a (deep exploration type) and C3b (random strolling type).

[0037] After cleaning outliers, the preliminary results of dynamic grouping are as follows: Commuter type: 3682 people, Family leisure type: 8125 people, In-depth tour type: 3342 people, Random stroll type: 1701 people, Group study tour type: 681 people; The system has set rules for transitioning between commuting and roaming. 673 visitors were detected switching modes, with the switching points mostly concentrated between 09:30-10:30 and 13:00-15:00.

[0038] Taking a user as an example, their time-series behavior records are as follows: 09:05-09:23, quick tour of Area A, 18-minute stop, classified as commuter type; From 09:23 to 10:50, I walked continuously through areas B, C, and D for 87 minutes, covering a total distance of 2100 meters, which is classified as a leisurely stroll.

[0039] The system detected the switching point at 09:23, with the interval between the commuter segment and the roaming segment changing from 18 minutes to 87 minutes, and the conversion rate being 69 minutes per cycle.

[0040] Overall statistics show that the average conversion rate on weekends is 71 minutes per conversion, higher than the 45 minutes per conversion on weekdays, indicating that behavioral pattern conversions are more likely to occur on weekends.

[0041] The system detected a breakpoint in the group affiliation in the initial dynamic grouping results: a group of tourists (241 people) were classified as commuter type before 10:40, and a distribution break occurred between 10:40 and 12:00, with the latter part of the trajectory overlapping with the wandering type group.

[0042] The system re-clustered the feature vectors before and after 10:40 and found that 92% of the tourists in this group were classified as in-depth tourers after the re-clustering.

[0043] Once the breakpoint is triggered, the dynamic grouping system labels are updated to reduce misclassification of attribution.

[0044] The system collected real-time data between 12:00 and 13:00, showing a sudden increase in visitor flow in the southeast area from 1121 to 2485, exceeding the system threshold (+80%). This automatically triggered a group adjustment signal. The system retrieved historical group adjustment records, determined it to be a special peak period during a holiday, and performed group iterations, dynamically adjusting security and tour guide resource deployment. The system integrated mode conversion tracking and the latest group tags to generate multi-dimensional visitor profiles and dynamic group distribution. In Example 2: Commuter type: 2,874 people, average stay 20 minutes; Family leisure type: 7,890 people, average stay 71 minutes; In-depth tour type: 4,250 people, average stay 153 minutes; Random stroll type: 1,448 people, average stay 113 minutes; Group study tour type: 721 people, average stay 48 minutes.

[0045] The system simultaneously outputs a forecast of 23,000 total passengers during the peak period (10:00-13:00) the following day, with the proportion of C3 group increasing by 12%, suggesting that guidance and crowd control measures in relevant areas should be strengthened.

[0046] To verify the difference between the present invention and traditional methods, the system selected 100,000 data points within the same period for comparison: index Traditional methods Method of the present invention Time-sharing forecast of MAE (person) 611 186 Group label misclassification rate 11.5% 2.6% Mode switching detection rate 54.3% 89.7% Real-time group update delay 260 seconds 37 seconds Behavioral breakage missed detection rate 17.4% 2.9% The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting visitor flow in smart cultural tourism scenic areas based on deep learning, characterized in that, include: By collecting hourly, daily, and monthly visitor entry and exit data and location sequences from sensors and motion track recorders at the scenic area entrance, a preliminary set of multi-scale distribution features was obtained. Based on the obtained multi-scale distribution feature set, the nesting characteristics between hourly, daily, and monthly sequences are decomposed using time series analysis. For each scale, the distribution similarity index is calculated. The nesting association strength is judged by comparing the similarity index with a preset threshold, and the enhanced intrinsic association feature vector is obtained. If the similarity index in the enhanced intrinsic association feature vector exceeds the preset threshold, the corresponding intrinsic association feature vector is input into the hierarchical clustering algorithm for preliminary tourist group segmentation. Clustering iteration is performed based on weekend behavior differences to obtain an initial dynamic grouping result set. Obtain a subset of pattern transitions from the initial dynamic grouping result set. By tracking the migration path from commuting to roaming behavior in the subset, use time series analysis to quantify the transition rate and obtain the pattern transition tracking sequence. If the mode transition tracking sequence shows that the transition rate is higher than the preset threshold, the group affiliation label in the initial dynamic grouping result set is adjusted, and the re-clustering mechanism is applied to the identified breakpoint to fuse the intrinsic correlation feature vectors to obtain an updated dynamic grouping system. By matching the updated dynamic grouping system with real-time passenger flow distribution data, it is determined whether dynamic changes trigger the grouping adjustment threshold, thus obtaining an adaptive grouping update signal. Based on the adaptive grouping update signal, the fusion mode transformation tracking sequence and intrinsic correlation feature vector are used to generate the final tourist group segmentation structure, thus obtaining the passenger flow management system dataset.

2. The method for predicting visitor flow in smart cultural tourism scenic areas based on deep learning according to claim 1, characterized in that, The method of decomposing the nested characteristics between hourly, daily, and monthly sequences using time series analysis includes: The time-series decomposition method is used to process the multi-scale distribution feature set to obtain hourly, daily, and monthly sequences. Distribution similarity was calculated for hourly, daily, and monthly series respectively, resulting in three sets of distribution similarity values; The three sets of distribution similarity values ​​are compared with the preset similarity threshold. If the distribution similarity is greater than the similarity threshold, it is determined that there is a strong nested association between the corresponding scales. If the distribution similarity is less than or equal to the similarity threshold, the nested association between the corresponding scales is determined to be weak, and the set of nested association strengths for each scale is obtained. The original multi-scale distribution feature set is weighted and adjusted based on the nested correlation strength set of each scale. Scales with high correlation strength are given high weights and scales with low correlation strength are given low weights, resulting in the adjusted distribution feature set. The k-means clustering method is used to process the adjusted distribution feature set to obtain the cluster center set. The cluster center set and the adjusted distribution feature set are then concatenated as vectors to obtain the enhanced intrinsic correlation feature vector.

3. The method for predicting visitor flow in smart cultural tourism scenic areas based on deep learning according to claim 1, characterized in that, If the similarity index in the enhanced intrinsic association feature vector exceeds a preset threshold, the corresponding intrinsic association feature vector is input into a hierarchical clustering algorithm for preliminary tourist group segmentation, including: By detecting the similarity index of the enhanced intrinsic correlation feature vectors, if the detection result exceeds the preset threshold, it is included in the dataset to be processed, and an initial feature set that meets the conditions is obtained. Based on the initial feature set that meets the conditions, a hierarchical clustering algorithm is used to initially divide the tourist groups. The feature vectors of each group are classified and labeled to obtain the preliminary group categories. For the initially classified groups, weekend behavior data from the pattern transition signals are obtained, the significant characteristics of behavioral differences are analyzed, and the key points of behavioral differences are determined. If the key points of behavioral differences meet the preset significance conditions, then perform clustering iteration on the relevant group categories and update the intermediate results of dynamic grouping; Based on the intermediate results of dynamic grouping, the core behavioral patterns of each group are extracted, and a second comparison is performed in conjunction with the pattern transition signal to determine the stability of the grouping. Based on the assessment of group stability, unstable group categories are adjusted, and outliers are removed using data cleaning tools to obtain an initial dynamic grouping result set.

4. The method for predicting visitor flow in smart cultural tourism scenic areas based on deep learning according to claim 1, characterized in that, The migration path from commuting to roaming by tracking behavioral characteristics in a subset includes: By filtering out the user set that includes the transition from commuting mode to roaming mode from the initial dynamic grouping result set using preset rules, a mode conversion subset is obtained; For the pattern conversion subset, all user behavior records are arranged in chronological order to obtain a behavior sequence arranged in chronological order. Extract consecutive commuting mode segments and roaming mode segments from the time-ordered behavior sequence to determine the switching point position each time from commuting mode to roaming mode; Based on the determined switching point location, calculate the time interval between two adjacent switching points to obtain the time span data for each mode transition; The time span data of all transformations are processed using time series analysis methods to calculate the rate value at the time of each transformation, thus obtaining the transformation rate sequence; The conversion rate sequence is rearranged according to the original user time order to obtain the mode conversion tracking sequence for each user.

5. The method for predicting visitor flow in smart cultural tourism scenic areas based on deep learning according to claim 1, characterized in that, The method of applying a re-clustering mechanism to identify breakpoints and fusing intrinsically correlated feature vectors to obtain an updated dynamic grouping system includes: Acquire the mode transition tracking sequence and determine whether the transition rate is higher than a preset threshold; If the conversion rate is higher than the preset threshold, the group affiliation label of the initial dynamic grouping result set is extracted; The tracking sequence is scanned segment by segment based on the group affiliation label to determine the location of the breakpoint in the continuity of the tracking sequence; For each breakpoint, the intrinsic correlation feature vectors on both sides are collected. The k-means clustering algorithm is used to perform re-clustering on the intrinsic correlation feature vectors collected before and after the breakpoint, resulting in a new set of group affiliation labels after re-clustering. The corresponding labels in the initial dynamic grouping result set are replaced with the new set of group affiliation labels to form a dynamic grouping system that includes the updated group affiliation labels.

6. The method for predicting visitor flow in smart cultural tourism scenic areas based on deep learning according to claim 1, characterized in that, The step of matching the updated dynamic grouping system with real-time passenger flow distribution data to determine whether dynamic changes trigger a grouping adjustment threshold includes: Passenger flow distribution data is acquired through a real-time data acquisition system. The acquired data is then preliminarily cleaned and formatted to obtain a standardized passenger flow distribution dataset. Using a standardized passenger flow distribution dataset and a pre-established dynamic grouping system, a data matching operation is performed to determine the correspondence between the current passenger flow distribution and the grouping system; Based on the data matching results, analyze the dynamic changes in passenger flow distribution. If the dynamic changes exceed the preset adjustment threshold, generate a preliminary grouping adjustment demand signal. For the initial grouping adjustment demand signal, historical grouping adjustment records and real-time environmental data are obtained. Logical comparison is used to determine whether the adaptive update conditions are met, and an adaptive grouping update signal is obtained.

7. The method for predicting visitor flow in smart cultural tourism scenic areas based on deep learning according to claim 1, characterized in that, The step of generating the final tourist group segmentation structure by updating the adaptive grouping signal, fusing the mode transformation tracking sequence and the intrinsic correlation feature vector includes: An adaptive grouping mechanism is used to dynamically cluster the adaptive grouping update signal to generate a preliminary tourist group classification set. Based on the preliminary tourist group classification set, information related to pattern transformation is obtained, and behavioral pattern analysis is performed over time using tracking sequence data to determine the dynamic changing trend of tourist groups. By extracting intrinsic correlation information from dynamic trends and combining it with feature vectors, a multi-dimensional profile of tourist groups is constructed, resulting in refined tourist group segmentation. Based on the refined tourist group segmentation results, a pre-established decision tree model is used to structure the segmentation results and generate the initial framework of the passenger flow management system. Based on the initial framework, combined with passenger flow data and evolvability requirements, the management system is dynamically adjusted. If the response speed of the management system is lower than the preset threshold, the response capability is improved by adding data processing nodes, thus obtaining the passenger flow management system dataset.

Images