An artificial intelligence-based method and system for predicting tourism indicators

Abstract

The present invention relates to the field of artificial intelligence technology. The present invention discloses a tourism index prediction method and system based on artificial intelligence. The method includes performing a correlation analysis on traffic flow data and tourism platform data through a spatio-temporal cross-attention mechanism to obtain a spatio-temporal correlation matrix, obtaining the historical number of tourists, inputting the spatio-temporal correlation matrix and the historical number of tourists into a trained first neural network model to obtain an initial tourist prediction value, performing a time series decomposition on the traffic flow data to obtain a first variation factor, extracting semantic emotion features from the tourism platform data to obtain a second variation factor, and then constructing a variation network based on a graph neural network according to the first variation factor and the second variation factor, and performing adaptive topology optimization on a preset second neural network model based on the variation network, effectively alleviating the problem that the platform popularity is decoupled from the actual number of tourists during peak periods or in complex traffic situations.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology. More specifically, the present invention relates to a method and system for predicting tourism indicators based on artificial intelligence. Background Art

[0002] In the prior art, the prediction of the number of scenic area visitors mainly relies on the modeling analysis of historical visitor flow data and the mining of platform behavior data such as search volume, comment quantity, and ticket purchase behavior on tourism platforms. Common methods include trend prediction models based on time series, regression models, and neural network models gradually introduced in recent years. These methods usually rely on the statistical laws of historical visitor data and tourism popularity, and achieve the prediction of the number of visitors in a future period through model fitting, providing auxiliary decision-making support for scenic area management and resource allocation.

[0003] However, in actual tourism scenarios, the change in the number of scenic area visitors is not only affected by tourism platform behavior data, but also significantly restricted by offline traffic factors. Especially during peak periods or in popular scenic areas, due to the concentrated influx of visitors, the traffic flow around the scenic area increases significantly, leading to road congestion and parking difficulties. This traffic pressure, in turn, reduces the travel willingness of visitors, resulting in a deviation between platform popularity and the actual number of visitors. The prior art fails to effectively feedback factors such as "parking pressure" or "traffic congestion" into the visitor behavior prediction model, resulting in deviations in prediction results under the complex and fluctuating traffic environment of the scenic area, and being unable to accurately reflect the dynamic changes in the actual number of visitors.

[0004] In view of this, the present invention proposes a method and system for predicting tourism indicators based on artificial intelligence to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for predicting tourism indicators based on artificial intelligence.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] In a first aspect, a method for predicting tourism indicators based on artificial intelligence is provided, including:

[0008] Performing a correlation analysis on traffic flow data and tourism platform data based on a spatio-temporal cross-attention mechanism to obtain a spatio-temporal correlation matrix;

[0009] Obtaining historical visitor numbers, and inputting the spatio-temporal correlation matrix and the historical visitor numbers into a trained first neural network model to obtain an initial visitor prediction value;

[0010] Perform time series decomposition on traffic flow data to obtain a first variation factor, and perform semantic emotion feature extraction on tourism platform data to obtain a second variation factor;

[0011] Construct a variation network based on the first variation factor and the second variation factor, and perform adaptive topology optimization on a preset second neural network model based on the variation network to obtain an index correction model;

[0012] Input the initial tourist prediction value into the trained index correction model to obtain the target tourist prediction value.

[0013] In some embodiments, the traffic flow data includes parking space occupancy rate, vehicle detention duration, and road congestion index. The method for performing correlation analysis on the traffic flow data and the tourism platform data through a spatio-temporal cross-attention mechanism to obtain a spatio-temporal correlation matrix includes:

[0014] S101: Align the time stamps of the traffic flow data and the tourism platform data, perform spatial feature encoding on the traffic flow data to generate a spatial embedding vector; perform temporal feature encoding on the tourism platform data to generate a temporal embedding vector;

[0015] S102: Construct a spatio-temporal cross-attention mechanism, and calculate bidirectional attention weight matrices respectively with the spatial embedding vector as Query, the temporal embedding vector as Key-Value, and the temporal embedding vector as Query, the spatial embedding vector as Key-Value, where Query is a query matrix, Key is a key matrix, and Value is a value matrix;

[0016] S103: Concatenate the bidirectional attention weight matrices through a preset fully connected layer to obtain a spatio-temporal correlation matrix;

[0017] S104: Obtain the stability value of the spatio-temporal correlation matrix, and calculate the Pearson correlation coefficient between the traffic flow data and the tourism platform data;

[0018] S105: Determine whether the stability value is less than a preset stability threshold. If so, go to S106; if not, output the spatio-temporal correlation matrix;

[0019] S106: Determine whether the Pearson correlation coefficient is less than a preset correlation coefficient threshold. If so, return to step S101 and adjust the encoding parameters; if not, output the spatio-temporal correlation matrix.

[0020] In some embodiments, the method for performing spatial feature encoding on the traffic flow data to generate a spatial embedding vector includes:

[0021] Obtain the target area map corresponding to the traffic flow data, perform fixed-size division on the target area map to obtain W grids;

[0022] Bind each grid with the corresponding parking space occupancy rate, vehicle residence time, and road congestion index to obtain a traffic flow feature map with spatial location attributes;

[0023] Calculate the Euclidean distance between each grid, and use the Euclidean distance as the weight of each edge in the traffic flow feature map. Among them, the nodes in the traffic flow feature map are grids, and the edges are the spatial connection relationships between grids;

[0024] Encode the traffic flow feature map through a preset graph convolutional network, perform feature aggregation and update in the spatial dimension, and obtain a spatial embedding vector.

[0025] In some embodiments, the method for obtaining the stability value of the spatio-temporal correlation matrix includes:

[0026] Obtain two spatio-temporal correlation matrices generated by the current time period H and the previous time H - 1 respectively, that is, and ;

[0027] Calculate and The average change rate and variance change rate in the global feature distribution;

[0028] Judge whether the average change rate is less than the preset average change threshold, and whether the variance change rate is less than the preset variance change threshold;

[0029] If so, calculate the cosine similarity between the two spatio-temporal correlation matrices, and use the cosine similarity as the stability value;

[0030] If not, calculate the KL divergence between the two spatio-temporal correlation matrices, and use the KL divergence as the stability value.

[0031] In some embodiments, the travel platform data includes user search volume, user comment data, and ticket reservation volume. The method for adjusting the encoding parameters includes:

[0032] Traverse the traffic flow data of each grid. If the parking space occupancy rate or road congestion index of at least one grid exceeds G times the standard deviation of the historical same period threshold, it is determined that the traffic flow data is abnormal;

[0033] Traverse the travel platform data of each time period. If the user search volume or ticket reservation volume of at least one time period exceeds G times the standard deviation of the historical same period threshold, it is determined that the travel platform data is abnormal;

[0034] If only the traffic flow data is abnormal, adjust the spatial feature encoding parameters. If only the travel platform data is abnormal, adjust the time feature encoding parameters;

[0035] If both the abnormal traffic flow data and the abnormal tourism platform data exist, adjust the spatial feature encoding parameters and the temporal feature encoding parameters simultaneously;

[0036] If no abnormal traffic flow data and abnormal tourism platform data are detected, adjust the attention mechanism parameters.

[0037] In some embodiments, the method for performing time series decomposition on traffic flow data to obtain the first variation factor includes:

[0038] Judge whether there is missing traffic flow data in each grid within the target time window. If there is missing data, calculate the spatial similarity of adjacent grids, and select adjacent grids whose spatial similarity is greater than the preset similarity threshold;

[0039] Based on the average value of the traffic flow data of the adjacent grids, perform time filling and alignment on the missing traffic flow data, resample the filled grids according to the unified timestamp set, and construct a multi-dimensional time series dataset;

[0040] Perform additive decomposition processing on the multi-dimensional time series dataset through the STL algorithm to obtain the residual term in the multi-dimensional time series dataset;

[0041] Calculate the residual variance value corresponding to the residual term, and use the residual variance value as the first variation factor.

[0042] In some embodiments, the method for extracting semantic emotion features from tourism platform data to obtain the second variation factor includes:

[0043] Extract the user comment data in the tourism platform data, preprocess the user comment data, and construct a target comment corpus;

[0044] Based on the preset semantic encoding model, perform context semantic encoding on the target comment corpus to obtain the corresponding semantic text features;

[0045] Use the semantic text features as the second variation factor.

[0046] In some embodiments, the method for constructing a variation network based on a graph neural network according to the first variation factor and the second variation factor includes:

[0047] Take each grid in the target area as a node of the graph neural network, construct a topological initial graph structure, and the edge weight value between any two nodes in the topological initial graph structure is a function value of their spatial distance;

[0048] Embed the first variation factor as a node dynamic attribute into the topological initial graph structure;

[0049] Extract the corresponding emotional intention vector in the second variation factor, and use the emotional intention vector as the dynamic weight adjustment factor for the associated edges in the initial topological graph structure to obtain a heterogeneous graph structure;

[0050] Based on a preset multi-layer graph neural network, perform feature propagation and topological convolution operations on the heterogeneous graph structure to generate a node coupling strength matrix, and use the node coupling strength matrix as the variation network.

[0051] In some embodiments, the method for adaptively optimizing the topology of a preset second neural network model based on the variation network to obtain an index correction model includes:

[0052] Align the hidden layer structure in the preset second neural network model with the node structure in the variation network to establish a mapping relationship between the neural network structure and the grid;

[0053] Based on the node coupling strength matrix between the nodes in the variation network, adjust the connection edge weights between the hidden layer neurons in the second neural network model;

[0054] If there is a node connection relationship with a coupling strength lower than the preset pruning threshold, perform a structure pruning operation in the neural network to remove the corresponding connection;

[0055] If there is a regional cluster with a coupling strength significantly higher than the average value, perform local topology deepening processing to obtain an index correction model.

[0056] In a second aspect, a tourism index prediction system based on artificial intelligence is provided, which is used to implement the above-mentioned tourism index prediction method based on artificial intelligence, including:

[0057] The association analysis module: used to perform association analysis on the traffic flow data and tourism platform data based on the spatio-temporal cross-attention mechanism to obtain a spatio-temporal association matrix;

[0058] The index prediction module: used to obtain the historical tourist volume, and input the spatio-temporal association matrix and the historical tourist volume into the trained first neural network model to obtain an initial tourist prediction value;

[0059] The first processing module: used to perform time series decomposition on the traffic flow data to obtain a first variation factor, and perform semantic emotion feature extraction on the tourism platform data to obtain a second variation factor;

[0060] The second processing module: used to construct a variation network based on the first variation factor and the second variation factor, and perform adaptive topology optimization on a preset second neural network model based on the variation network to obtain an index correction model;

[0061] Index correction module: used to input the initial tourist prediction value into the trained index correction model to obtain the target tourist prediction value.

[0062] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0063] By introducing a spatio-temporal cross-attention mechanism, the present invention establishes the correlation between traffic flow data and tourism platform data, can reveal the interference degree of traffic status on platform popularity. Further, based on the change network constructed by the first change factor and the second change factor, it can capture the abnormal fluctuations of vehicle flow and the changes in user sentiment in a scenic area within a specific time period, thereby adaptively optimizing the topology of the preset second neural network model, improving the model's adaptability to local fluctuations, and finally dynamically correcting the initial tourist prediction value through the index correction model, effectively alleviating the problem that the platform popularity is decoupled from the actual number of tourists during peak periods or in complex traffic situations. Description of the Drawings

[0064] Figure 1 It is a schematic flowchart of a tourism index prediction method based on artificial intelligence in the present invention;

[0065] Figure 2 It is a schematic flowchart of a method for adjusting coding parameters in the present invention;

[0066] Figure 3 It is a schematic structural diagram of a tourism index prediction system based on artificial intelligence in the present invention;

[0067] Figure 4 It is a comparison chart for evaluating the stability value of the spatio-temporal correlation matrix in the present invention;

[0068] Figure 5 It is a trend chart for characterizing the influence of the first and second change factors on the tourist prediction error in the present invention;

[0069] Figure 6 It is a comparison chart of the mean square error between the initial tourist prediction value and the target tourist prediction value in the present invention. Detailed Embodiments

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the described embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0071] Unless otherwise defined, the technical or scientific terms used in the present invention shall have the usual meanings understood by persons with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar terms used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. The words "include" or "comprise" and similar terms mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, but do not exclude other elements or objects. The words "connect" or "connected" and similar terms are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.

[0072] Example 1

[0073] See also Figure 1 As shown, this embodiment discloses a tourism index prediction method based on artificial intelligence, including:

[0074] S10: Based on the spatiotemporal cross-attention mechanism, the correlation analysis of traffic flow data and tourism platform data is performed to obtain the spatiotemporal correlation matrix;

[0075] In this embodiment, the traffic flow data includes parking space occupancy rate, vehicle detention time and road congestion index. The parking space occupancy rate refers to the average stay time of vehicles from entering the parking lot to leaving the parking lot, the percentage of the number of parking spaces occupied in the target area in real time to the total number of parking spaces, the vehicle detention time refers to the road congestion index, which represents a quantitative indicator of traffic status calculated based on real-time traffic flow, average vehicle speed and road capacity, usually expressed in levels of 0-10 (the higher the value, the more severe the congestion). The above-mentioned roads refer to roads near scenic spots. Tourism platform data includes user search volume, user comment data and ticket reservation volume. User search volume refers to the number of keyword searches actively initiated by users on tourism platforms (such as Ctrip, Qunar, Fliggy, etc.) for specific scenic spots or related services (such as tickets, hotels, transportation) during the target time period.

[0076] The method of analyzing the correlation between traffic flow data and tourism platform data through the spatiotemporal cross attention mechanism and obtaining the spatiotemporal correlation matrix includes:

[0077] S101: aligning the timestamps of the vehicle flow data and the tourism platform data, encoding the spatial features of the vehicle flow data, and generating a spatial embedding vector; encoding the temporal features of the tourism platform data, and generating a temporal embedding vector;

[0078] S102: Construct a spatio-temporal cross-attention mechanism. Calculate the bidirectional attention weight matrix by using the spatial embedding vector as the Query, the temporal embedding vector as the Key-Value, and the temporal embedding vector as the Query, the spatial embedding vector as the Key-Value respectively, where Query is the query matrix, Key is the key matrix, and Value is the value matrix;

[0079] S103: Concatenate the bidirectional attention weight matrix through a preset fully connected layer to obtain a spatio-temporal correlation matrix;

[0080] S104: Obtain the stability value of the spatio-temporal correlation matrix and calculate the Pearson correlation coefficient between the traffic flow data and the tourism platform data;

[0081] S105: Determine whether the stability value is less than a preset stability threshold. If so, go to S106; if not, output the spatio-temporal correlation matrix;

[0082] S106: Determine whether the Pearson correlation coefficient is less than a preset correlation coefficient threshold. If so, return to step S101 and adjust the encoding parameters; if not, output the spatio-temporal correlation matrix.

[0083] It should be noted that the time feature encoding of the tourism platform data can be achieved by organizing the tourism platform data (including user search volume, user comment data, ticket reservation volume, etc.) into a time series according to the time dimension (such as hour / day), constructing an input tensor, using the encoder part of the standard Transformer model to model the time series, and extracting the correlation between different time periods through the multi-head self-attention mechanism to obtain the time feature encoding.

[0084] In the process of timestamp alignment, if the time granularity of the traffic flow data is at the minute level while the tourism platform data is at the hour level, the alignment can be achieved through the following method: taking the mean or maximum value of the minute-level traffic flow data in an hourly window to generate an aggregated result with the same timestamp as the tourism platform data.

[0085] The methods for generating spatial embedding vectors by performing spatial feature encoding on the traffic flow data include:

[0086] Obtain the target area map corresponding to the traffic flow data, divide the target area map into a fixed size to obtain W grids;

[0087] Bind each grid with the corresponding parking space occupancy rate, vehicle residence time, and road congestion index to obtain a traffic flow feature map with spatial position attributes;

[0088] Calculate the Euclidean distance between each grid, and use the Euclidean distance as the weight of each edge in the traffic flow feature map, where the nodes in the traffic flow feature map are grids and the edges are the spatial connection relationships between grids;

[0089] Encode the traffic flow feature map through a preset graph convolutional network, perform feature aggregation and update in the spatial dimension, and obtain a spatial embedding vector.

[0090] In this embodiment, the target area map refers to a geographical information layer with a clear spatial boundary constructed based on a preset geographical boundary (for example, extending 3 kilometers outward from the scenic area) with the scenic area and its surrounding areas as the research scope. This layer is used to carry out spatial rasterization processing and serves as the basic area graphic structure for spatial analysis, supporting subsequent raster division and spatial mapping processes. The graph convolutional network encodes the traffic flow feature map by specifically performing multi-layer feature aggregation operations on this graph structure, enabling the representation vector of each node in the current layer to be iteratively updated into a high-dimensional spatial embedding vector by combining the features and structural information of its neighboring nodes.

[0091] In this embodiment, by performing spatial embedding and temporal embedding on traffic flow data and tourism platform data respectively, and using a spatio-temporal cross-attention mechanism to construct a bidirectional association structure, the spatio-temporal association matrix generated on this basis not only reveals the dynamic conduction path of "the impact of traffic status on tourist popularity", but also depicts the feedback mechanism of "the reverse effect of tourist popularity on regional traffic flow fluctuations". In addition, the spatial embedding vector extracted from the traffic flow raster map by means of the graph convolutional network can significantly enhance the model's perception ability of the spatial heterogeneity of the scenic area, enabling the prediction model to perform differential modeling between different geographical areas. The spatio-temporal association matrix and spatial embedding vector obtained in the above manner enable the model to have the ability of cross-modal data collaborative perception and local area traffic change sensitivity, which cannot be achieved by traditional single-channel historical data prediction methods. This significantly improves the accuracy and timeliness of tourism index prediction in peak periods or sudden fluctuation scenarios.

[0092] The methods for calculating the bidirectional attention weight matrix include:

[0093] ;

[0094] In the formula, represents the attention weight matrix of the spatial embedding vector to the temporal embedding vector, represents the attention weight matrix of the temporal embedding vector to the spatial embedding vector, represents the query matrix generated for the spatial embedding vector, represents the query matrix generated for the temporal embedding vector, represents the transposed matrix of the key matrix generated for the temporal embedding vector, represents the transposed matrix of the key matrix generated for the spatial embedding vector, represents the value matrix generated for the temporal embedding vector, represents the value matrix generated for the spatial embedding vector, is the scaling factor, represents the normalized exponential function.

[0095] The method for obtaining the stability value of the spatio-temporal correlation matrix includes:

[0096] Obtain two spatio-temporal correlation matrices generated by the current time period H and the previous time H - 1 respectively, namely and ;

[0097] Calculate and the average change rate and variance change rate in the global feature distribution;

[0098] Judge whether the average change rate is less than the preset average change threshold and whether the variance change rate is less than the preset variance change threshold;

[0099] If so, calculate the cosine similarity between the two spatio-temporal correlation matrices, and use the cosine similarity as the stability value;

[0100] If not, calculate the KL divergence between the two spatio-temporal correlation matrices, and use the KL divergence as the stability value.

[0101] As Figure 4 shown, Figure 4 shows the stability values of the cosine similarity and KL divergence in three adjacent time periods. Among them, the cosine similarity curve always remains above 0.88, indicating that the overall distribution of the spatio-temporal correlation matrix is relatively stable. The KL divergence curve rises to 0.32 in the H - 2~H - 1 period and then falls back to 0.22 from H - 1~H, indicating that there has been an obvious distribution drift in this period but it has been corrected. It can be seen that in the stability determination process, when the matrix mean and variance change rates are small, it can be directly determined as stable based on the cosine similarity. Once a distribution drift is detected, the KL divergence is used to timely identify anomalies and adjust the coding parameters to ensure that the output spatio-temporal correlation matrix remains reliable and provides a robust input for subsequent tourist prediction.

[0102] In this embodiment, the average change rate is obtained by calculating the change amplitude of the difference between the mean of the spatio-temporal correlation matrix generated in the current time period and the mean of the matrix in the previous time period relative to the mean of the previous time period. The variance change rate indicates whether there is a significant change in the overall characteristic distribution fluctuation degree of the two-period matrices. The calculation methods of the above average value and variance are both common statistical quantity extraction methods in the prior art, which are applicable to the global feature analysis of numerical matrices. When the average change rate or the variance change rate is large, it indicates that the numerical distribution of the current spatio-temporal correlation matrix has a large deviation. At this time, the KL divergence is used to measure the distribution difference between the two matrices. The KL divergence is a common method for measuring the distance between two probability distributions. By normalizing the matrix (such as Softmax), it can be regarded as a probability distribution, and then the KL divergence is calculated to reflect its information deviation degree.

[0103] In this embodiment, by organically integrating the cosine similarity and the KL divergence, an adaptive judgment of the stability of the spatio-temporal correlation matrix from multiple angles is realized. This method maintains high efficiency when the matrix feature distribution is stable, and automatically switches to a more robust distribution metric index when the data fluctuates violently, effectively avoiding misjudgment problems caused by early model convergence, abnormal fluctuations or attention imbalance, and improving the self-adjusting ability and generalization reliability of the prediction model.

[0104] Please refer to Figure 2 As shown, the method for adjusting the coding parameters includes:

[0105] Traverse the traffic flow data of each grid. If the parking space occupancy rate or road congestion index of at least one grid exceeds G times the standard deviation of the historical same period threshold, it is determined that the traffic flow data is abnormal;

[0106] Traverse the tourism platform data of each time period. If the user search volume or ticket reservation volume of at least one time period exceeds G times the standard deviation of the historical same period threshold, it is determined that the tourism platform data is abnormal;

[0107] If only the traffic flow data is abnormal, adjust the spatial feature coding parameters. If only the tourism platform data is abnormal, adjust the time feature coding parameters;

[0108] If both the traffic flow data anomaly and the tourism platform data anomaly exist simultaneously, adjust the spatial feature coding parameters and the time feature coding parameters simultaneously;

[0109] If no traffic flow data anomaly and tourism platform data anomaly are detected, adjust the attention mechanism parameters.

[0110] In this embodiment, the method for adjusting the spatial feature encoding parameters can be to increase or decrease the number of layers of the graph convolutional network, or to replace the type of activation function. The method for adjusting the temporal feature encoding parameters can be to increase or decrease the number of attention heads of the Transformer. The method for adjusting the attention mechanism parameters can be to modify the scaling factor. For example, in the calculation of the bidirectional attention weight matrix, item.

[0111] This embodiment introduces a dynamic adjustment mechanism for encoding parameters driven by anomaly detection results. By statistically discriminating the traffic flow data and tourism platform data in the spatial and temporal dimensions respectively, it can judge in real time whether "spatial feature anomaly" or "temporal feature anomaly" occurs, and accordingly adjust the number of layers of the graph convolutional network, the number of Transformer attention heads or the attention mechanism parameters. Compared with the processing method of unified encoding structure and fixed network depth in the prior art, in this embodiment, when the traffic flow data is abnormal, the spatial aggregation ability is enhanced, when the tourism platform data is abnormal, the cross-time segment modeling depth is enhanced, and joint adjustment is performed in the spatio-temporal double anomaly scenario, thus significantly improving the dynamic response ability and prediction accuracy stability of the model in complex spatio-temporal coupling situations such as the emergence of hot spots and holiday impacts, and effectively solving the problem of lack of adaptability of traditional models under the fixed structure.

[0112] S20: Obtain the historical number of tourists, input the spatio-temporal correlation matrix and the historical number of tourists into the trained first neural network model, and obtain the initial tourist prediction value;

[0113] In this embodiment, the training method of the first neural network model includes:

[0114] The input layer in the first neural network model receives the historical number of tourists and the historical spatio-temporal correlation matrix, and the output layer in the first neural network model outputs the historical initial tourist prediction value;

[0115] When training the first neural network model, select the mean square error loss function as the loss function, and minimize the loss function by the gradient descent method;

[0116] Update the weight parameters of the first neural network model, and through iterative training, obtain the trained first neural network model.

[0117] In this embodiment, the first neural network model can be an LSTM model, and both the historical number of tourists and the spatiotemporal association matrix are accompanied by corresponding time labels. The spatiotemporal association matrix, as the interactive mapping result between traffic flow data and tourism platform data, not only reflects the coupling strength between the two types of data in terms of time dimension and spatial distribution, but also reveals the nonlinear relationship between traffic status and user travel intention. Specifically, the matrix is ​​generated through a bidirectional cross-attention mechanism, in which each weight value represents the response degree of traffic characteristics in a specific spatial grid (such as parking space occupancy rate, vehicle detention time, road congestion index) to platform behavior characteristics in a certain time period (such as user search volume, user comment data, ticket reservation volume).

[0118] In the initial tourist prediction process, the matrix is ​​input into the first neural network model as a spatiotemporal intervention variable of the coupling degree between platform popularity and traffic status, and together with the historical number of tourists, it constitutes a joint input tensor. During the training process, the neural network can automatically learn through repeated optimization: the positive correlation between "strong traffic-heat coupling" and "actual visitor growth", and the negative feedback phenomenon that the actual number of tourists decreases instead of increases when "the popularity is high but the traffic is extremely congested", thereby forming an accurate modeling mechanism for whether the platform popularity can be effectively converted into tourist traffic.

[0119] Taking a holiday peak period as an example, assuming that the user search volume and ticket reservation volume of a scenic spot increase rapidly within a certain period of time, according to the traditional model, the predicted value of tourists in this area should increase simultaneously. However, through traffic flow monitoring, it was found that the road congestion index of multiple grids around the scenic spot remained high and the parking space occupancy rate was close to saturation, causing a large number of potential tourists to face long-term detention and parking difficulties, and their actual willingness to travel decreased.

[0120] In this embodiment, the bidirectional attention mechanism responds to the asynchronous pattern between the platform popularity and the traffic status, and the weight of the corresponding position in the spatiotemporal association matrix is ​​automatically weakened, indicating that "the popularity has not formed an effective conversion". After integrating the matrix, the first neural network will revise the prediction result downward, thereby avoiding excessive overestimation based on the platform popularity. In contrast, if the traffic flow data and the tourism platform data in the same period show a high degree of synchronization, the weight in the matrix will be enhanced, and the model will increase the prediction value, forming a prediction mechanism that is more sensitive to the actual trend of changes in tourists and has more timely feedback.

[0121] S30: performing time series decomposition on the traffic flow data to obtain a first variation factor, and extracting semantic sentiment features from the travel platform comment data to obtain a second variation factor;

[0122] Methods for performing time series decomposition on vehicle flow data to obtain the first variation factor include:

[0123] Determine whether there is missing traffic flow data in each grid within the target time window. If there is missing data, calculate the spatial similarity of adjacent grids, and select adjacent grids whose spatial similarity is greater than the preset similarity threshold;

[0124] Based on the average value of the traffic flow data of the adjacent grids, perform time filling and alignment on the missing traffic flow data, resample the filled grids according to the unified timestamp set, and construct a multi-dimensional time series dataset;

[0125] Perform additive decomposition processing on the multi-dimensional time series dataset through the STL algorithm to obtain the residual term in the multi-dimensional time series dataset;

[0126] Calculate the residual variance value corresponding to the residual term, and use the residual variance value as the first change factor.

[0127] In this embodiment, STL is the abbreviation of Seasonal-Trend decomposition using Loess, which is often referred to in Chinese as the seasonal-trend decomposition method based on Loess regression. It is a classic and robust additive time series decomposition technique. As mentioned above, the residual term represents the unexplained abnormal fluctuations and mutation characteristics that still exist after removing the long-term trend and periodic changes. This term is used to capture the non-periodic and non-stable components in the traffic flow state and reflect short-term disturbances or occasional traffic behaviors.

[0128] It should be added that the method for calculating the residual variance value in this embodiment can be the unbiased sample variance calculation method in statistical analysis, which is used to quantify the fluctuation intensity of the residual term of each grid within the target time window. This embodiment will not elaborate on this too much.

[0129] The method for extracting the semantic emotion characteristics of the tourism platform review data to obtain the second change factor includes:

[0130] Extract the user review data in the tourism platform data, preprocess the user review data, and construct a target review corpus;

[0131] Based on the preset semantic coding model, perform context semantic coding on the target review corpus to obtain the corresponding semantic text features;

[0132] Use the semantic text features as the second change factor.

[0133] In this embodiment, the above preprocessing may perform operations such as removing HTML tags, punctuation, and stop word cleaning, word segmentation, lemmatization, or stemming on the user comment data to improve the parsing ability of the semantic encoding model for natural language context, and at the same time reduce the interference of invalid or redundant words on feature extraction. The semantic encoding model may adopt a deep semantic representation method based on a pre-trained language model, such as the BERT model, ERNIE, or ELECTRA model. By performing bidirectional context encoding on the comment text, a high-dimensional semantic embedding vector representing the user's emotion, subjective intention, and topic association is extracted to construct the second variation factor.

[0134] In this embodiment, aiming at the data missing problem in the traffic flow data, adjacent grids are selected by spatial similarity for time filling, and the STL algorithm is combined to separate periodic and mutation information, enhancing the sensitivity of the model to abnormal traffic flow behaviors. By using context semantic encoding models such as BERT to extract features from user comments, it is no longer limited to surface emotion judgment, but can identify the complex intention expressions of users in different geographical, traffic, and service scenarios, and the generated semantic features are more context-dependent and generalization-capable.

[0135] S40: Construct a variation network based on a graph neural network according to the first variation factor and the second variation factor, and perform adaptive topology optimization on a preset second neural network model based on the variation network to obtain an index correction model;

[0136] The method for constructing a variation network based on a graph neural network according to the first variation factor and the second variation factor includes:

[0137] Taking each grid in the target area as a node of the graph neural network, construct an initial topological graph structure, and the edge weight value between any two nodes in the initial topological graph structure is a function value of their spatial distance;

[0138] Embed the first variation factor as a node dynamic attribute into the initial topological graph structure;

[0139] Extract the corresponding emotion intention vector in the second variation factor, and use the emotion intention vector as the dynamic weight adjustment factor of the associated edge in the initial topological graph structure to obtain a heterogeneous graph structure;

[0140] Perform feature propagation and topological convolution operations on the heterogeneous graph structure based on a preset multi-layer graph neural network to generate a node coupling strength matrix, and use the node coupling strength matrix as the variation network.

[0141] Among the above, the spatial distance can be the Euclidean distance, Manhattan distance, or Haversine distance based on longitude and latitude coordinates between the center points of any two grids within the target area. The emotional intention vector represents the subjective tendency, emotional polarity, and semantic focus of the user's comment statement in a specific context. It not only reflects the user's emotional response to the scenic area service or environment (such as satisfaction, complaint, expectation, etc.), but also incorporates potential intentions in the comment, such as scenario-sensitive emotional content like "waiting too long", "traffic congestion", "worth visiting again", etc. It is an encoding representation of natural language semantics in a multi-dimensional emotional space.

[0142] As Figure 5 shown, Figure 5 The tourist prediction error curves corresponding to the increase of the first variation factor and the second variation factor intensity from 0.1 to 1.0 are shown. It can be seen that the two curves generally decline, and the error reduction is more obvious when the intensity of the second variation factor ≥ 0.6. The residual variance captures the short-term fluctuations of traffic flow, and the emotional intention vector reflects the changes in comment emotions; the two are jointly input into the index correction model, which can adaptively weaken the prediction deviation in scenarios of traffic congestion or sudden emotional changes. The curve trend verifies the "dual variation factor + adaptive topology optimization" mechanism, proving that the model has the ability to quickly converge errors and improve prediction credibility during peak periods.

[0143] In this embodiment, the method for extracting the emotional intention vector can be to perform sentiment word annotation and part-of-speech recognition on the comment text in combination with an emotion dictionary (such as NRC or SentiWordNet), count the occurrence frequencies of various emotion words (such as positive, negative, anxious, expectant, etc.) within the target time window, and construct a multi-dimensional emotion frequency vector based on their semantic categories, emotional polarities, and co-occurring word distributions in the context.

[0144] The method for adaptively optimizing the topology of the preset second neural network model based on the variation network to obtain the index correction model includes:

[0145] Align the hidden layer structure in the preset second neural network model with the node structure in the variation network to establish a mapping relationship between the neural network structure and the grid;

[0146] Based on the node coupling strength matrix between the nodes in the variation network, adjust the connection edge weights between the hidden layer neurons in the second neural network model. Among them, the neurons corresponding to the nodes with strong coupling strength have enhanced connection weights, while those with weak coupling strength have reduced connection weights;

[0147] If there is a node connection relationship with a coupling strength lower than the preset pruning threshold, perform a structure pruning operation in the neural network to remove the corresponding connection;

[0148] If there are clusters of regions where the coupling strength is significantly higher than the average value, local topology deepening processing is performed to obtain an index correction model.

[0149] In this embodiment, the principle of the optimization strategy is as follows: A variation network constructed by fusing two variation factors extracted from traffic flow and tourism platform data (i.e., the first variation factor representing traffic fluctuation intensity and the second variation factor representing changes in user intentions) is used as a guiding signal to dynamically reconstruct the internal connection structure of the neural network, so that the topological structure of the prediction model is closer to the actual spatio-temporal change state of the current region. Specifically:

[0150] Each grid node represents a local spatial region, and the coupling strength of this node reflects the degree of tight association with other regions in terms of spatio-temporal dynamic changes;

[0151] Mapping this coupling strength to the connection edge weights between neurons in the neural network is essentially adjusting the "information flow path" of the network according to the spatial change trend, enhancing the modeling depth of high-frequency interaction regions;

[0152] For edges with extremely low coupling strength, it is considered that their influence on the target output is limited, and they can be removed through pruning operations to reduce redundant calculations;

[0153] For local regions with extremely high coupling strength, add connection or skip-layer structures to enable the model to have stronger representation capabilities in these "hot regions".

[0154] Pruning threshold setting: According to the distribution of historical coupling strength data, set the pruning threshold to the 10th percentile of all coupling strength values. If the coupling strength of a certain connection is lower than this threshold, then remove this connection.

[0155] Local topology deepening processing: For clusters of regions where the coupling strength is higher than 2 standard deviations of the mean, insert residual blocks in the neural network hidden layer at the corresponding positions. The specific structure is as follows:

[0156] ;

[0157] In the formula, is the output feature of the residual block, is the hidden layer feature input to the residual block, is the weight matrix of the first fully connected layer, is the weight matrix of the second fully connected layer, is the activation function.

[0158] In this embodiment, compared with the neural network with a fixed structure, this method can flexibly adjust the structure according to the current spatio-temporal state, improving the model's response ability to scenarios of drastic traffic fluctuations and sudden changes in user emotions. Through connection enhancement and topology deepening driven by coupling strength, it focuses on modeling hot grids and high-intensity interaction regions, avoiding the problem of "prediction failure" for local abnormal regions in traditional models. Structure pruning eliminates low-coupling connections, simplifies the network scale, and enables the model to have higher operating efficiency and deployment lightweight characteristics while maintaining prediction accuracy.

[0159] S50: Input the initial tourist prediction value into the trained index correction model to obtain the target tourist prediction value.

[0160] In this embodiment, the training steps of the index correction model are roughly the same as those of the first neural network model, except that the input data is changed to the historical initial tourist prediction value, and the output data is changed to the historical target tourist prediction value. In this embodiment, the training method of the index correction model includes:

[0161] The input layer in the index correction model receives the historical initial tourist prediction value, and the output layer in the index correction model outputs the historical target tourist prediction value;

[0162] When training the index correction model, select the mean square error loss function as the loss function, and minimize the loss function through the gradient descent method;

[0163] Update the weight parameters of the index correction model, and through iterative training, obtain the trained index correction model.

[0164] As Figure 6 shown, Figure 6 Compare the actual tourist numbers A(t), the initial tourist prediction values 0(t), and the target tourist prediction values P*(t) at 20 time indices. The solid line A(t) and the dotted line P*(t) basically coincide, while the dashed line 0(t) overestimates or underestimates many times. 0(t) is output by the first neural network model; after being dynamically corrected by the index correction model driven by the first variation factor and the second variation factor, P*(t) is obtained. The improvement in the curve fitting degree shows that: the index correction model effectively compensates for the decoupling phenomenon between platform popularity and actual passenger flow, reduces the overall mean square error, and provides more accurate real-time tourist predictions for decision-making scenarios such as scenic spot parking space allocation and ticket warning.

[0165] In this embodiment, first, the spatio-temporal cross-attention mechanism is used to analyze the correlation between traffic flow data and tourism platform data to obtain a spatio-temporal correlation matrix, and the historical number of tourists is obtained. The spatio-temporal correlation matrix and the historical number of tourists are input into the trained first neural network model to obtain an initial tourist prediction value. The traffic flow data is decomposed by time series to obtain a first variation factor, and the semantic emotion features of the tourism platform review data are extracted to obtain a second variation factor. Then, based on the first variation factor and the second variation factor, a variation network based on a graph neural network is constructed, and the preset second neural network model is adaptively topologically optimized based on the variation network to obtain an index correction model. Finally, the initial tourist prediction value is input into the trained index correction model to obtain a target tourist prediction value. By introducing the spatio-temporal cross-attention mechanism, this embodiment establishes the correlation between traffic flow data and tourism platform data, and can reveal the interference degree of traffic status on platform popularity. Further, the variation network constructed based on the first variation factor and the second variation factor can capture the abnormal fluctuations of vehicle flow and user emotion changes in a scenic area during a specific period, so as to adaptively topologically optimize the preset second neural network model and improve the model's adaptability to local fluctuations. Finally, the index correction model dynamically corrects the initial tourist prediction value, effectively alleviating the problem that the platform popularity is decoupled from the actual number of tourists during peak periods or complex traffic situations.

[0166] Embodiment 2

[0167] Please refer to Figure 3 As shown, based on the same inventive concept, this embodiment discloses and provides an artificial intelligence-based tourism index prediction system. For the details not described in this embodiment, please refer to the relevant parts in Embodiment 1. The system includes:

[0168] Correlation analysis module: used to analyze the correlation between traffic flow data and tourism platform data based on the spatio-temporal cross-attention mechanism to obtain a spatio-temporal correlation matrix;

[0169] Index prediction module: used to obtain the historical number of tourists, and input the spatio-temporal correlation matrix and the historical number of tourists into the trained first neural network model to obtain an initial tourist prediction value;

[0170] First processing module: used to decompose the traffic flow data by time series to obtain a first variation factor, and extract the semantic emotion features of the tourism platform data to obtain a second variation factor;

[0171] Second processing module: used to construct a variation network based on a graph neural network according to the first variation factor and the second variation factor, and adaptively topologically optimize the preset second neural network model based on the variation network to obtain an index correction model;

[0172] Index calibration module: It is used to input the initial tourist prediction value into the trained index calibration model to obtain the target tourist prediction value.

[0173] In the attached drawings of the embodiments of the present invention, only the structures related to the embodiments of the present invention are involved. For other structures, reference may be made to the general design. Without conflict, the features in the same embodiment and different embodiments of the present invention can be combined with each other. The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A tourism index prediction method based on artificial intelligence, characterized in that: include: Based on the spatiotemporal cross-attention mechanism, the correlation between traffic flow data and tourism platform data is analyzed to obtain the spatiotemporal correlation matrix; Obtain the historical number of tourists, input the spatiotemporal correlation matrix and the historical number of tourists into the trained first neural network model, and obtain the initial tourist prediction value; The traffic flow data is decomposed into time series to obtain the first change factor, and the tourism platform data is extracted from semantic emotional features to obtain the second change factor; According to the first change factor and the second change factor, a change network based on a graph neural network is constructed, and based on the change network, an adaptive topology optimization is performed on a preset second neural network model to obtain an indicator correction model; The initial tourist prediction value is input into the trained indicator correction model to obtain the target tourist prediction value.

2. The tourism index prediction method based on artificial intelligence according to claim 1 is characterized in that: Traffic flow data includes parking space occupancy rate, vehicle detention time and road congestion index. The methods of analyzing the correlation between traffic flow data and tourism platform data through spatiotemporal cross attention mechanism and obtaining spatiotemporal correlation matrix include: S101: aligning the timestamps of the vehicle flow data and the tourism platform data, encoding the spatial features of the vehicle flow data, and generating a spatial embedding vector; encoding the temporal features of the tourism platform data, and generating a temporal embedding vector; S102: construct a spatiotemporal cross attention mechanism, and calculate a bidirectional attention weight matrix with the spatial embedding vector as Query, the temporal embedding vector as Key-Value, and the temporal embedding vector as Query and the spatial embedding vector as Key-Value, respectively, where Query is the query matrix, Key is the key matrix, and Value is the value matrix; S103: concatenating the bidirectional attention weight matrix through a preset fully connected layer to obtain a spatiotemporal correlation matrix; S104: Obtain the stability value of the spatiotemporal correlation matrix and calculate the Pearson correlation coefficient between the traffic flow data and the tourism platform data; S105: Determine whether the stability value is less than a preset stability threshold value, if so, proceed to S106, if not, output the spatiotemporal correlation matrix; S106: Determine whether the Pearson correlation coefficient is less than a preset correlation coefficient threshold. If so, return to step S101 and adjust the encoding parameters. If not, output the spatiotemporal correlation matrix.

3. The tourism index prediction method based on artificial intelligence according to claim 2 is characterized in that: Methods for encoding the spatial features of traffic flow data and generating spatial embedding vectors include: Obtain a target area map corresponding to the traffic flow data, divide the target area map into fixed-size grids, and obtain W grids; Bind each grid with the corresponding parking space occupancy rate, vehicle detention time and road congestion index to obtain a traffic flow characteristic map with spatial location attributes; Calculate the Euclidean distance between each grid and use the Euclidean distance as the weight of each edge in the traffic flow characteristic graph, where the nodes in the traffic flow characteristic graph are grids and the edges are the spatial connection relationships between grids; The traffic flow feature map is encoded through a preset graph convolutional network, and feature aggregation and updating in the spatial dimension are performed to obtain a spatial embedding vector.

4. The tourism index prediction method based on artificial intelligence according to claim 2 is characterized in that: Methods for obtaining the stability value of the spatiotemporal correlation matrix include: Get the two spatiotemporal correlation matrices generated by the current time period H and the previous time period H-1 respectively, that is, and ; calculate and The rate of change of the mean and variance of the global feature distribution; Determine whether the average change rate is less than a preset average change threshold, and whether the variance change rate is less than a preset variance change threshold; If so, the cosine similarity between the two spatiotemporal correlation matrices is calculated and the cosine similarity is used as the stability value; If not, the KL divergence between the two spatiotemporal correlation matrices is calculated and the KL divergence is used as the stability value.

5. The tourism index prediction method based on artificial intelligence according to claim 3 is characterized in that: The tourism platform data includes user search volume, user comment data and ticket reservation volume. The methods for adjusting the encoding parameters include: Traverse the traffic flow data of each grid. If the parking space occupancy rate or road congestion index of at least one grid exceeds G times the standard deviation of the historical threshold value for the same period, it is determined that the traffic flow data is abnormal; Traverse the tourism platform data of each time period. If there is at least one time period in which the user search volume or ticket reservation volume exceeds the historical threshold value by G times the standard deviation, it is determined that the tourism platform data is abnormal; If only the traffic flow data is abnormal, the spatial feature encoding parameters are adjusted; if only the tourism platform data is abnormal, the temporal feature encoding parameters are adjusted; If the traffic flow data is abnormal and the tourism platform data is abnormal at the same time, the spatial feature coding parameters and the temporal feature coding parameters are adjusted at the same time; If no abnormal traffic flow data and tourism platform data are detected, the attention mechanism parameters are adjusted.

6. The tourism index prediction method based on artificial intelligence according to claim 1 is characterized in that: Methods for performing time series decomposition on vehicle flow data to obtain the first variation factor include: Determine whether each grid has missing traffic flow data within the target time window. If so, calculate the spatial similarity of adjacent grids and select adjacent grids whose spatial similarity is greater than the preset similarity threshold. Based on the mean of the traffic flow data of the neighboring grids, the missing traffic flow data are time-filled and aligned, and the filled grids are resampled according to the unified timestamp set to construct a multidimensional time series dataset. The multidimensional time series data set is processed by additive decomposition using the STL algorithm to obtain the residual term in the multidimensional time series data set; Calculate the residual variance value corresponding to the residual term and use the residual variance value as the first change factor.

7. The tourism index prediction method based on artificial intelligence according to claim 5 is characterized in that: The method of extracting semantic emotion features from tourism platform data to obtain the second change factor includes: Extract user review data from tourism platform data, preprocess the user review data, and construct the target review corpus; Based on the preset semantic encoding model, the target review corpus is contextually encoded to obtain the corresponding semantic text features; Semantic text features are used as the second variable factor.

8. The tourism index prediction method based on artificial intelligence according to claim 6 is characterized in that: According to the first change factor and the second change factor, a method for constructing a change network based on a graph neural network includes: Taking each grid in the target area as a node of the graph neural network, a topological initial graph structure is constructed. The edge weight between any two nodes in the topological initial graph structure is a function of their spatial distance. Embed the first change factor as a node dynamic attribute into the topological initial graph structure; Extracting the corresponding emotional intention vector from the second variable factor, and using the emotional intention vector as a dynamic weight adjustment factor of the associated edge in the topological initial graph structure to obtain a heterogeneous graph structure; Based on the preset multi-layer graph neural network, feature propagation and topological convolution operations are performed on the heterogeneous graph structure to generate a node coupling strength matrix, which is used as a variable network.

9. The tourism index prediction method based on artificial intelligence according to claim 8 is characterized in that: The method of performing adaptive topology optimization on the preset second neural network model based on the change network to obtain the index correction model includes: Align the hidden layer structure in the preset second neural network model with the node structure in the variable network, and establish a mapping relationship between the neural network structure and the grid; Obtaining a node coupling strength matrix between nodes in the change network, and adjusting the connection edge weights between neurons in the hidden layer of the second neural network model; If there is a node connection relationship whose coupling strength is lower than the preset pruning threshold, a structural pruning operation is performed in the neural network to remove the corresponding connection; If there are regional clusters with coupling strength higher than the average, local topological deepening is performed to obtain an index correction model.

10. A tourism index prediction system based on artificial intelligence, which is used to implement the tourism index prediction method based on artificial intelligence described in any one of claims 1 to 9, characterized in that: include: Correlation analysis module: used to perform correlation analysis on traffic flow data and tourism platform data based on the spatiotemporal cross attention mechanism to obtain the spatiotemporal correlation matrix; Index prediction module: used to obtain the historical number of tourists, input the spatiotemporal association matrix and the historical number of tourists into the trained first neural network model, and obtain the initial tourist prediction value; The first processing module is used to perform time series decomposition on the traffic flow data to obtain the first change factor, and to extract semantic emotional features from the tourism platform data to obtain the second change factor; The second processing module is used to construct a change network based on the graph neural network according to the first change factor and the second change factor, and to perform adaptive topology optimization on the preset second neural network model based on the change network to obtain an indicator correction model; Index correction module: used to input the initial tourist prediction value into the trained index correction model to obtain the target tourist prediction value.

Images