A spatiotemporal tourism demand prediction method based on graph convolution and attention mechanism

Abstract

This invention provides a method for predicting spatiotemporal tourism demand based on graph convolution and attention mechanisms. The method comprises the following steps: data acquisition; spatial connection analysis between scenic spots: based on the acquired data, determining the spatial connections between scenic spots that can be explicitly represented, and summarizing the possible relevant factors of implicit spatial connections; spatiotemporal information representation: constructing dynamic multidimensional graphs and spatiotemporal sequences for explicit and implicit spatial connections, respectively; high-dimensional spatial connection embedding: based on the spatiotemporal information representation, using a graph convolution embedding layer and a spatial self-attention embedding layer to extract high-dimensional explicit and implicit spatial connection representations, respectively, and aggregate them into a high-dimensional spatial connection representation; high-dimensional spatiotemporal feature extraction: using a bidirectional convolutional long short-term memory layer and a temporal self-attention layer to gradually extract short-term and long-term high-dimensional spatiotemporal features; and spatiotemporal tourism demand prediction. This invention can significantly improve the performance of spatiotemporal tourism demand prediction. The analysis method and prediction model have important guiding significance and application value.

Description

Technical Field

[0001] The present invention relates to the technical field of spatiotemporal sequence prediction and tourism demand prediction, and in particular to a spatiotemporal tourism demand prediction method based on graph convolution and attention mechanism. Background Art

[0002] Over the past few decades, the tourism industry has experienced rapid growth worldwide, generating significant economic and social benefits. Accurate tourism demand forecasting can help tourists plan their itineraries and help tourism practitioners rationally allocate tourism resources, playing an increasingly important role in enhancing the travel experience and enabling smart tourism.

[0003] The tourism demand forecasting problem is typically modeled as a time series forecasting problem, a spatial modeling problem (or cross-sectional analysis), or a spatiotemporal prediction problem. Time series forecasting uses observed time series data to predict the development trend of a target time series over a period of time. Spatial modeling, on the other hand, uses data collected from different objects at a specific moment (i.e., cross-sectional data) to establish spatial connections between different objects, then makes assumptions about the explanatory variables involved to form an "ex ante forecast." Tourism demand forecasting based on time series forecasting and spatial modeling ignores spatial information, or more importantly, temporal information, making it difficult to obtain accurate tourism demand forecasts. Spatiotemporal forecasting considers both spatial and temporal information of tourism demand and is considered the most reasonable problem formulation for tourism demand forecasting. Spatiotemporal prediction models can handle tourism demand forecasting for multiple regions within a specified spatial structure because spatial information is aggregated into the model to provide accurate cross-regional tourism demand forecasts. However, existing spatiotemporal tourism demand forecasting methods only consider static spatial connections between pre-specified regions (see Non-Patent Literatures 1-3), but fail to consider multiple and dynamic spatial connections.

[0004] In recent years, deep learning-based spatiotemporal prediction models (see Non-Patent Literatures 4-6), particularly those using graph convolution and attention mechanisms, have achieved excellent results in predicting traffic flow and travel demand, demonstrating that deep learning models can learn complex spatiotemporal features from data to make accurate spatiotemporal predictions. However, to date, no spatiotemporal prediction model for tourism demand using graph convolution and attention mechanisms exists.

[0005] Prior art literature

[0006] Non-patent literature

[0007] Non-Patent Literature 1: YANG Y, ZHANG H. Spatial-temporal forecasting of tourism demand [J / OL]. Annals of Tourism Research, 2019, 75: 106-119. DOI: 10.1016 / j.annals.2018.12.024.

[0008] Non-patent literature 2: Jiao X, Li G, Chen J L. Forecasting international tourism demand: a local spatiotemporal model [J / OL]. Annals of Tourism Research, 2020, 83: 102937. DOI: 10.1016 / j.annals.2020.102937.

[0009] Non-patent literature 3: Long W, Liu C, Song H. Pooling in tourism demand forecasting [J / OL]. Journal of Travel Research, 2019, 58(7): 1161-1174. DOI: 10.1177 / 0047287518800390.

[0010] Non-patent literature 4: MA X, DAI Z, HE Z, et al. Learning traffic as images: A deep convolutional neural network for large-scale transportation network speed prediction [J / OL]. Sensors, 2017, 17(4): 818. DOI: 10.3390 / s17040818.

[0011] Non-patent document 5: YU B, YIN H, ZHU Z. Spatio-temporal graph convolutional networks: a deep learning framework for traffic forecasting[C] / / Proceedings of the 27th International Joint Conference on Artificial Intelligence. Stockholm, Sweden: AAAI Press, 2018: 3634-3640.

[0012] Non-Patent Literature 6: Xu M, Dai W, Liu C, et al. Spatial-temporal transformer networks for traffic flow forecasting [J / OL]. arXiv:2001.02908 [cs, eess], 2021. http: / / arxiv.org / abs / 2001.02908. Summary of the Invention

[0013] In response to the shortcomings of existing tourism demand forecasting methods, the present invention provides a spatiotemporal tourism demand forecasting method based on graph convolution and attention mechanism. Graph convolution and spatial self-attention are used to learn the explicit spatial connections and implicit spatial connections between attractions respectively, and bidirectional convolution long short-term memory and temporal self-attention layer are used to gradually extract short-term high-dimensional spatiotemporal features and long-term high-dimensional spatiotemporal features, thereby improving the effect of spatiotemporal tourism demand forecasting.

[0014] The time-space tourism demand forecasting method of the present invention is characterized by comprising the following steps:

[0015] S1, data acquisition step, based on experience and data availability, obtains useful data for the prediction of spatiotemporal tourism demand for scenic spots;

[0016] S2, a step of analyzing the spatial connections between scenic spots, analyzing the spatial connections between scenic spots in the scenic area based on the obtained useful data, determining the spatial connections between scenic spots that can be explicitly expressed, and summarizing the relevant factors of the implicit spatial connections that are difficult to be explicitly expressed;

[0017] S3, a spatiotemporal information representation step, constructing a spatiotemporal information representation for the explicit spatial connections and implicit spatial connections between the scenic spots, specifically including:

[0018] S301. Construction of a multi-dimensional dynamic graph associated with explicit spatial connections between scenic spots. Consider a scenic area S with N scenic spots S = {s1, s2, ..., s N}, construct a dynamic multi-dimensional graph G = (V, E), which consists of a set of nodes V = {V1, V2, ..., V N} and d disjoint subsets representing different explicit dynamic connections constitute the edge set E={E1,E2,...,E d}composition,

[0019] S302: Constructing a spatiotemporal sequence associated with the implicit spatial connections between scenic spots. Considering c variables that may be related to the implicit spatial connections, the spatiotemporal sequence X can be expressed as:

[0020]

[0021] Where T is the time series length of X, and N is the number of scenic spots;

[0022] S4, a high-dimensional spatial connection embedding step, extracting a high-dimensional explicit spatial connection representation from the dynamic multidimensional graph using a graph convolutional embedding layer based on the spatiotemporal information representation, extracting a high-dimensional implicit spatial connection representation from the spatiotemporal sequence using a spatial self-attention embedding layer, and finally aggregating the high-dimensional explicit spatial connection representation and the high-dimensional implicit spatial connection representation into a high-dimensional spatial connection representation;

[0023] S5, high-dimensional spatiotemporal feature extraction step, applying a bidirectional convolutional long short-term memory layer and a temporal self-attention layer on the high-dimensional spatial connection feature sequence to gradually extract short-term high-dimensional spatiotemporal features and long-term high-dimensional spatiotemporal features.

[0024] S6, a prediction step, predicting the spatiotemporal tourism demand based on the extracted high-dimensional spatiotemporal features.

[0025] As an optional method of the spatiotemporal tourism demand prediction method of the present invention, each dimension of the dynamic multidimensional graph represents an explicit spatial connection between scenic spots, wherein the nodes in the node set V contain historical tourism demand attributes and periodic tourism demand attributes of scenic spots.

[0026] As an optional method of the spatiotemporal tourism demand forecasting method of the present invention, the step S4 of extracting a high-dimensional explicit spatial relationship representation from the dynamic multidimensional graph using a graph convolutional embedding layer includes:

[0027] S401, use the graph convolution embedding layer to extract high-dimensional explicit spatial connection representation, assuming that the historical tourism demand X his It is associated with n1 kinds of explicit spatial connections, and its graph embedding is expressed as:

[0028]

[0029] in By the i-th explicit spatial connection at time t and X his Obtained through graph convolution operation. Assuming q types of periodic tourism demand, periodic tourism demand With n p There are explicit spatial connections, and its graph embedding is expressed as:

[0030]

[0031] in By the i-th explicit spatial connection at time t and Obtained through graph convolution operations. The final high-dimensional explicit spatial connection representation Depend on With each group Connect and add a fully connected layer to get:

[0032]

[0033] Where ReLU is the linear rectification function, W exp and b exp are learnable parameters and biases.

[0034] As an optional method of the spatiotemporal tourism demand forecasting method of the present invention, the step S4 of using a spatial self-attention embedding layer to extract a high-dimensional implicit spatial relationship representation from the spatiotemporal sequence includes:

[0035] S402, use the spatial self-attention embedding layer to extract high-dimensional implicit spatial connection representation, assuming that the actual sequence input to the spatial self-attention embedding layer each time is a spatiotemporal sequence with a time sequence length of τ First, split it into τ segments Each segment is linearly mapped into three matrices Q, K, and V, and scaled dot product attention is calculated, and finally aggregated into the output of spatial self-attention. After two layers of normalization and a feedforward layer, a high-dimensional implicit spatial connection representation h is obtained. imp :

[0036] h imp =LayerNorm((ReLU(h ln W1+b1)W2+b2))+h ln ,

[0037] h ln =LayerNorm(h at )+h at ,

[0038]

[0039]

[0040] Where Attention(·,·,·) is the scaled dot product attention calculation.

[0041] As an optional method of the spatiotemporal tourism demand forecasting method of the present invention, the step S4 of aggregating the high-dimensional explicit spatial relationship representation and the high-dimensional implicit spatial relationship representation into a high-dimensional spatial relationship representation includes:

[0042] S403, aggregating the high-dimensional explicit spatial relationship representation and the high-dimensional implicit spatial relationship representation into a high-dimensional spatial relationship representation h emb ,Right now:

[0043] h emb =Concat(h exp ,h imp )W3+b3,

[0044] Where W3 and b3 are learnable parameters and biases.

[0045] As an optional method of the spatiotemporal tourism demand prediction method of the present invention, the useful data includes at least one of the historical tourism demand of each scenic spot in the scenic area, passenger flow information between each scenic spot, traffic conditions between each scenic spot, geographical location information of each scenic spot, search index of each scenic spot, weather conditions of each scenic spot (or scenic area), and holiday information of the scenic area.

[0046] The present invention has the following advantages and effects compared to the prior art:

[0047] 1. The present invention establishes a theory for distinguishing and modeling explicit and implicit spatial connections between scenic spots, which has important guiding significance for systematically analyzing the spatial connections between scenic spots.

[0048] 2. The spatiotemporal information representation method of the present invention can combine multiple dynamic spatial connections through two data forms: graph and spatiotemporal sequence.

[0049] 3. The graph convolution embedding layer of the present invention can embed explicit spatial connections between scenic spots in a multi-dimensional dynamic graph, and the spatial self-attention mechanism can embed implicit spatial connections between scenic spots in a spatiotemporal sequence, thereby comprehensively embedding high-dimensional dynamic spatial connections between scenic spots.

[0050] 4. The present invention can effectively extract complex high-dimensional local and long-term spatiotemporal features through bidirectional convolutional long short-term memory and temporal self-attention mechanism, thereby providing accurate spatiotemporal tourism demand forecasting.

[0051] 5. Compared with the time-space tourism demand prediction methods in the existing technical literature, the present invention can achieve the best time-space tourism demand prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is an overall flow chart of the technical solution of the present invention.

[0053] Figure 2 It is an architectural diagram of the spatiotemporal tourism demand prediction model of the present invention.

[0054] Figure 3 This is the architecture diagram of the spatial attention mechanism in the spatiotemporal tourism demand prediction model of the present invention.

[0055] Figure 4 It is a multi-dimensional dynamic graph representation diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to more clearly demonstrate the technical principles, technical solution details and advantages of the present invention, a more detailed and clear description is provided below in conjunction with the embodiments and drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments described in this specification, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0057] The time-space tourism demand forecasting method of the present invention has the following technical implementation steps: Figure 1 As shown in the figure, the architecture of the spatiotemporal tourism demand prediction model is as follows Figure 2 As shown, each step is described in detail as follows:

[0058] Step S1, data acquisition, based on experience and data availability, obtain useful data that can be obtained for the spatiotemporal tourism demand forecast of a certain scenic spot, including but not limited to the historical tourism demand of each scenic spot in the scenic spot, passenger flow information between scenic spots, traffic conditions between scenic spots, geographical location information of each scenic spot, search index of each scenic spot, weather conditions of each scenic spot (or scenic spot), holiday information of the scenic spot, etc.

[0059] Step S2: Analyze the spatial connections between scenic spots. Based on the acquired useful data, analyze the spatial connections between scenic spots in the scenic area and determine the spatial connections that can be explicitly expressed, such as spatial distance connections, transportation connections, and passenger flow connections. In addition to explicit spatial connections, there are other connections between scenic spots that are difficult to express explicitly, such as spillover effects. The possible relevant factors (variables) of these implicit spatial connections are summarized.

[0060] Step S3: Spatiotemporal Information Representation: A spatiotemporal information representation is constructed for the explicit and implicit spatial connections between the attractions. Each dimension of the dynamic multidimensional graph represents an explicit spatial connection between attractions. Nodes contain attributes related to historical and cyclical tourism demand, and edge weights reflect the strength of the explicit connection between attractions. Both node attributes and edge weights change dynamically. A spatial sequence containing the historical and cyclical tourism demand data for all attractions, along with covariates, is constructed to extract implicit connections between attractions. These spatial sequences are stacked along the time dimension to form a spatiotemporal sequence.

[0061] Specifically include:

[0062] Step S301: Construction of a multi-dimensional dynamic graph associated with explicit spatial connections between scenic spots. Consider a scenic area S with N scenic spots = {s1, s2, ..., s N}, construct a dynamic multi-dimensional graph G = (V, E), which consists of a set of nodes V = {V1, V2, ..., V N} and d disjoint subsets representing different explicit dynamic connections constitute the edge set E={E1,E2,...,E d}composition.

[0063] Step S302: Constructing a spatiotemporal sequence associated with the implicit spatial connections between scenic spots. Considering c variables that may be related to the implicit spatial connections, the spatiotemporal sequence X can be expressed as:

[0064]

[0065] Where T is the time series length of X, and N is the number of scenic spots.

[0066] Step S4: high-dimensional spatial connection embedding. Based on the spatiotemporal information representation, a graph convolutional embedding layer is used to extract a high-dimensional explicit spatial connection representation from the dynamic multidimensional graph. A spatial self-attention embedding layer is used to extract a high-dimensional implicit spatial connection representation from the spatiotemporal sequence. Finally, the high-dimensional explicit spatial connection representation and the high-dimensional implicit spatial connection representation are aggregated into a high-dimensional spatial connection representation. Specifically, the steps include:

[0067] Step S401: Use the graph convolution embedding layer to extract high-dimensional explicit spatial relationship representation. Assume that the historical tourism demand X his It is associated with n1 kinds of explicit spatial connections, and its graph embedding is expressed as:

[0068]

[0069] in By the i-th explicit spatial connection at time t and X his Obtained through graph convolution operation. Assuming q types of periodic tourism demand, periodic tourism demand With n p There are explicit spatial connections, and its graph embedding is expressed as:

[0070]

[0071] in By the i-th explicit spatial connection at time t and Obtained through graph convolution operations. The final high-dimensional explicit spatial connection representation Depend on With each group Connect and add a fully connected layer to get:

[0072]

[0073] Where ReLU is the linear rectification function, W exp and b exp are learnable parameters and biases.

[0074] Step S402: Use the spatial self-attention embedding layer to extract high-dimensional implicit spatial connection representation. The spatial self-attention architecture is shown in the figure below. Figure 3 As shown, it is assumed that the actual sequence input to the spatial self-attention embedding layer each time is a spatiotemporal sequence with a time sequence length of τ First, split it into τ segments Each segment is linearly mapped into three matrices Q, K, and V, and scaled dot product attention is calculated, and finally aggregated into the output of spatial self-attention. After two layers of normalization and a feedforward layer, a high-dimensional implicit spatial connection representation h is obtained. imp :

[0075] h imp =LayerNorm((ReLU(h ln W1+b1)W2+b2))+h ln ,

[0076] in

[0077] h ln =LayerNorm(h at )+h at ,

[0078]

[0079]

[0080] Where Attention(·,·,·) is the scaled dot product attention calculation, that is:

[0081]

[0082] Step S403: Aggregate the high-dimensional explicit spatial relationship representation and the high-dimensional implicit spatial relationship representation into a high-dimensional spatial relationship representation h emb ,Right now:

[0083] h emb =Concat(h exp ,h imp )W3+b3,

[0084] Where W3 and b3 are learnable parameters and biases.

[0085] Step S5, high-dimensional spatiotemporal feature extraction, on the high-dimensional spatial connection feature sequence, apply the bidirectional convolution long short-term memory layer and the temporal self-attention layer to gradually extract short-term high-dimensional spatiotemporal features and long-term high-dimensional spatiotemporal features.

[0086] Step S6: predicting the spatiotemporal tourism demand. Predicting the spatiotemporal tourism demand based on the extracted high-dimensional spatiotemporal features.

[0087] Example

[0088] This example considers the application of the present invention to tourism demand forecasting for the Wanshan Islands in Zhuhai, China. The Wanshan Islands are a scenic island area in Zhuhai, primarily comprising Guishan Island, Wailingding Island, Dong'ao Island, and Wanshan Island. The Xiangzhou and Hengqin ports in downtown Zhuhai are the primary entry and exit ports for tourists to the Wanshan Islands. Therefore, this example considers tourism demand forecasting for these six attractions.

[0089] S1. Data acquisition. The collected data include:

[0090] 1) Historical tourism demand and inter-attraction passenger flow data were obtained from ferry ticketing data. The time range is January 1, 2019, to December 31, 2019, with a total of 11 observations per day (there is no ferry service between 20:00 and 8:00 (the next day), so the daily observations include hourly observations from 9:00 to 19:00). Daily, weekly, and monthly periodic tourism demand and periodic inter-attraction passenger flow data are the average of past daily, weekly, and monthly tourism demand and passenger flow, thus maintaining the same data granularity as the historical tourism demand and inter-attraction passenger flow data.

[0091] 2) Baidu search engine search index data for six scenic spots every day from January 1, 2019, to December 31, 2019, weather data from the China Meteorological Data Service Center, and Chinese holiday data; the daily data is expanded to a time index of 11 observations; in this embodiment, all six regions share the same holiday information;

[0092] 3) Geographical location information of the six attractions and the availability of ferry services between the six attractions; the data is extended to a time index of 11 observations per day from January 1, 2019 to December 31, 2019.

[0093] The dataset is split into training, validation, and test sets at a ratio of 8:1:3 (a total of 12 months of available data).

[0094] S2. Spatial connection analysis between scenic spots: Based on the acquired data, determine the spatial connections between scenic spots that can be explicitly expressed, including spatial distance connections between scenic spots, ferry transportation connections between scenic spots, and passenger flow connections between scenic spots (including short-term flow connections (i.e., within the last observation period), daily passenger flow connections, weekly passenger flow connections, and monthly passenger flow connections); the collected historical tourism demand X htd , cyclical tourism demand (X dtd 、X wtd 、X mtd ), Weather Information X wea , search index information X si 、Holiday InformationX hol Used to learn implicit spatial connections between scenic spots.

[0095] S3, spatiotemporal information representation, specifically including:

[0096] S301, for the explicit spatial connection between scenic spots, the dynamic multi-dimensional map is constructed as follows Figure 4 As shown in Figure 1, the multidimensional dynamic graph consists of six dimensions representing different explicit spatial connections. Each dimension contains six nodes and some edges connecting the nodes. The attributes of the nodes and the weights of the edges are dynamically changing. The adjacency matrix corresponding to this six-dimensional graph is represented as follows:

[0097]

[0098]

[0099]

[0100]

[0101]

[0102]

[0103] where dis(L i ,L j ) function represents the computing node and The distance between.

[0104] S302: Constructing a spatiotemporal sequence associated with the implicit spatial connections between scenic spots. Here, seven variables that may be related to the implicit spatial connections are considered. Therefore, the spatiotemporal sequence X can be expressed as:

[0105]

[0106] Where T is the time sequence length of X, and N is the number of scenic spots (here N = 6).

[0107] S4, high-dimensional spatial connection embedding, specifically including:

[0108] S401, using graph convolutional embedding layer to extract high-dimensional explicit spatial connection representation, historical tourism demand X htd and cyclical tourism demand (X dtd 、X wtd 、X mtd ) There are four groups in total. Assume that the actual input time series length of the spatiotemporal prediction model is τ lag , then the input of the graph convolution embedding layer is specifically:

[0109]

[0110] Assuming that these four groups of tourism demands are all related to two explicit spatial connections: geographical distance connection and ferry transportation connection, and each is also related to the corresponding passenger flow connection, their graph embedding representation is:

[0111]

[0112] in

[0113]

[0114]

[0115]

[0116]

[0117]

[0118] Where W exp 、W set 、W set,dis 、W set,trans 、W set,set is a learnable parameter, D t and A t is the degree matrix and adjacency matrix of the explicit spatial connection, ReLU is the linear rectification function, b exp is a learnable bias.

[0119] S402, use the spatial self-attention embedding layer to extract high-dimensional implicit spatial connection representation, for time series length τ lag The space-time sequence First, split it into τ lag fragments Each segment is linearly mapped into three matrices Q, K, and V, and scaled dot product attention is calculated, and finally aggregated into the output of spatial self-attention. After two layers of normalization and a feedforward layer, a high-dimensional implicit spatial connection representation h is obtained. imp :

[0120] h imp =LayerNorm((ReLU(h ln W1+b1)W2+b2))+h ln ,

[0121] h ln =LayerNorm(h at )+h at ,

[0122] in

[0123]

[0124]

[0125]

[0126] Step S403: Aggregate the high-dimensional explicit spatial relationship representation and the high-dimensional implicit spatial relationship representation into a high-dimensional spatial relationship representation h emb ,Right now:

[0127] h emb =Concat(h exp ,h imp )W3+b3,

[0128] Where W3 and b3 are learnable parameters and biases.

[0129] S5. High-dimensional spatiotemporal feature extraction. On the high-dimensional spatial connection feature sequence, the bidirectional convolutional long short-term memory layer and the temporal self-attention layer are applied to gradually extract short-term high-dimensional spatiotemporal features and long-term high-dimensional spatiotemporal features.

[0130] S6. Spatiotemporal tourism demand prediction. Based on the extracted high-dimensional spatiotemporal features, we use a two-layer feedforward network as a regressor to predict the spatiotemporal tourism demand of six scenic spots in Wanshan Islands.

[0131] To help understand the technical effects of the present invention, the inventors used technical indicators to compare and evaluate the methods and models of the present invention with the models of the prior art.

[0132] The evaluation indicators of the model are mean absolute error (MAE) and root mean square error (RMSE). seq For tourism demand forecasting, MAE and RMSE are defined as follows:

[0133]

[0134]

[0135] This example compares the present invention with the spatiotemporal autoregressive model (STARMA) in non-patent documents 1 and 2, the panel model (Panel) in non-patent documents 1 and 3, the deep convolutional network model (DCNN) in non-patent document 4, the spatiotemporal graph convolutional network model (STGCN) in non-patent document 5, and the spatiotemporal Transformer model (STTN) in non-patent document 6. The experimental results are shown in Table 1.

[0136] Table 1

[0137]

[0138] It can be seen from Table 1 that the MAE performance and RMSE performance of the spatiotemporal tourism demand prediction model proposed in this invention are generally better than those of the models in other technical literatures.

[0139] In summary, the spatiotemporal tourism demand prediction method and model of the present invention significantly improves the spatiotemporal tourism demand prediction performance by distinguishing, comprehensively analyzing and jointly representing explicit and implicit spatial connections between scenic spots, as well as the spatiotemporal tourism demand prediction model based on graph convolution and attention mechanism. This analysis method and prediction model have important guiding significance and application value.

Claims

1. A method for predicting spatiotemporal tourism demand, characterized in that: The following steps are involved: S1, data acquisition step, based on experience and data availability, obtains useful data that can be obtained for the prediction of spatiotemporal tourism demand for scenic spots; S2, a step of analyzing the spatial connections between scenic spots, analyzing the spatial connections between scenic spots in the scenic area based on the useful data obtained, determining the spatial connections between scenic spots that can be explicitly expressed, and summarizing the relevant factors of the implicit spatial connections that are difficult to be explicitly expressed; S3, a spatiotemporal information representation step, constructing a spatiotemporal information representation for the explicit spatial connections and implicit spatial connections between the scenic spots, specifically including: S301, Construction of a multi-dimensional dynamic graph associated with explicit spatial connections between scenic spots. Consider a Scenic spots of attractions , building a dynamic multidimensional graph , which consists of a set of nodes representing multiple attractions and The edge set consists of disjoint subsets representing different explicit dynamic connections composition, S302, construction of spatiotemporal sequence associated with implicit spatial connections between scenic spots, considering variables that may be related to the implicit spatial relationship, the spatiotemporal sequence It can be expressed as: ,in for The time series length, is the number of attractions; S4, a high-dimensional spatial connection embedding step, extracting a high-dimensional explicit spatial connection representation from the dynamic multidimensional graph using a graph convolutional embedding layer based on the spatiotemporal information representation, extracting a high-dimensional implicit spatial connection representation from the spatiotemporal sequence using a spatial self-attention embedding layer, and finally aggregating the high-dimensional explicit spatial connection representation and the high-dimensional implicit spatial connection representation into a high-dimensional spatial connection representation; S5, a high-dimensional spatiotemporal feature extraction step, applying a bidirectional convolutional long short-term memory layer and a temporal self-attention layer to the high-dimensional spatial connection feature sequence to gradually extract short-term high-dimensional spatiotemporal features and long-term high-dimensional spatiotemporal features; S6, a prediction step, predicting the spatiotemporal tourism demand based on the extracted high-dimensional spatiotemporal features; The S4 high-dimensional spatial connection embedding step includes: S401, using graph convolution embedding layer to extract high-dimensional explicit spatial connection representation, assuming historical tourism demand and There are explicit spatial connections, and its graph embedding is expressed as: ， in Depend on Moment Explicit spatial connections and Obtained through graph convolution operation, assuming Periodic tourism demand, periodic tourism demand and There are explicit spatial connections, and its graph embedding is expressed as: ， in Depend on Moment Explicit spatial connections and The final high-dimensional explicit spatial connection representation is obtained through graph convolution operation Depend on With each group Connect and add a fully connected layer to get: ， in is a linear rectification function, and are learnable parameters and biases; and S402, use the spatial self-attention embedding layer to extract high-dimensional implicit spatial connection representation, assuming that the actual sequence input to the spatial self-attention embedding layer each time is a time series length of The space-time sequence , first split it into fragments , linearly map each fragment to , , Three matrices are used to perform scaled dot product attention calculations, and finally aggregated into the output of spatial self-attention. After two layers of normalization and a feedforward layer, a high-dimensional implicit spatial connection representation is obtained. : ， ， ， ,in Scaled dot product attention calculation.

2. The method for predicting spatiotemporal tourism demand according to claim 1, characterized in that: Each dimension of the dynamic multidimensional graph represents an explicit spatial connection between attractions, where the node set The nodes in include the historical tourism demand attributes and periodic tourism demand attributes of scenic spots.

3. The method for predicting spatiotemporal tourism demand according to claim 1, characterized in that: Aggregating the high-dimensional explicit spatial relationship representation and the high-dimensional implicit spatial relationship representation into a high-dimensional spatial relationship representation in step S4 includes: S403, aggregating the high-dimensional explicit spatial relationship representation and the high-dimensional implicit spatial relationship representation into a high-dimensional spatial relationship representation ,Right now: ,in and are learnable parameters and biases.

4. The method for predicting spatiotemporal tourism demand according to claim 1, wherein: The useful data includes at least one of the historical tourism demand of each scenic spot in the scenic area, passenger flow information between each scenic spot, traffic conditions between each scenic spot, geographical location information of each scenic spot, search index of each scenic spot, weather conditions of each scenic spot or scenic area, and holiday information of the scenic area.

Images