Tourist destination passenger flow dispersion method based on multi-modal data fusion analysis

Abstract

The invention relates to the technical field of big data processing, in particular to a tourist destination passenger flow dispersion method based on multi-modal data fusion analysis. The method is suitable for tourist destination passenger flow guidance in holidays and vigorous seasons; by integrating various types of data and combining the means of correlation analysis, index mapping, prediction and early warning and the like, accurate analysis and effective dispersion of the passenger flow of the tourist destination are realized. Through the BiLSTM-Attention multi-mode deep learning prediction method, the model prediction precision can be effectively improved, benefited from the fact that a BiLSTM unit can identify feature information of input time series data in different directions at the same time, an Attention mechanism can distribute corresponding weights according to the importance degree of an input sequence at different time points; meteorological data and network keywords significantly improve the prediction capability of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of big data processing, and in particular to a method for guiding passenger flow at tourist destinations based on multimodal data fusion analysis; the method is suitable for guiding passenger flow at tourist destinations during holidays and peak tourist seasons; and the method achieves accurate analysis and effective guidance of passenger flow at tourist destinations by integrating various types of data and combining correlation analysis, indicator mapping, prediction and early warning, and other means. Background Art

[0002] In today's booming tourism industry, the passenger flow of tourist destinations is increasing day by day, especially during holidays, peak tourist seasons or popular scenic spots (venues). Peak passenger flow often leads to a decline in tourist experience, increased safety hazards, and uneven resource allocation. Traditional passenger flow diversion methods often rely on manual monitoring and limited data analysis, making it difficult to fully and accurately predict and respond to changes in passenger flow. Therefore, how to effectively divert passenger flow, improve tourist experience, and ensure the safety of scenic spots (venues) has become an urgent problem to be solved in tourism management.

[0003] With the rapid development of big data and artificial intelligence technologies, multimodal data integration and analysis technologies have provided new solutions for the flow of tourists to tourist destinations. By integrating structured data (such as historical passenger flow data, reservation data, etc.), video data (such as real-time video streams obtained by surveillance cameras in scenic spots), and historical data (such as tourist flow and weather data in the same period of previous years), a comprehensive and rich data set can be constructed to provide more accurate data support for passenger flow guidance. Summary of the invention

[0004] Tourism demand time series data is often a superposition or coupling of factors such as long-term trend changes, seasonal changes, cyclic changes, and irregular changes. Through multimodal data fusion analysis, the passenger flow of tourist destinations can be accurately predicted and monitored, so as to formulate reasonable passenger flow diversion strategies, reduce congestion and waiting time of tourists during the tour, and improve the overall satisfaction and tour experience of tourists; achieve real-time grasp and dynamic adjustment of passenger flow distribution in tourist destinations, ensure the rational allocation and utilization of scenic spot (venue) resources, avoid resource waste and overcrowding, and improve the operational efficiency and service quality of scenic spots (venues); provide a scientific basis for the planning and management of tourist destinations, promote the rational use and sustainable development of tourism resources, and also help promote the upgrading and transformation of the tourism industry.

[0005] The method for guiding passenger flow at tourist destinations based on multimodal data fusion analysis includes the following steps: S1. Data Collection Obtain the daily passenger flow data set of the scenic spot and set the time range for passenger flow data collection; The historical meteorological data and future meteorological forecast data of key places in the tourist destination within the time range are obtained through the Internet, and two kinds of meteorological data, weather status and temperature, are collected. The weather status variables include light rain, sunny, cloudy, and showers; the temperature includes the daily maximum temperature T max and the minimum temperature T min； Use Python to crawl keyword search values ​​related to tourist destinations, remove keywords with zero data, and form a backup network keyword data set; S2. Data preprocessing and data analysis Smooth the abnormal points of the above data set, and replace individual missing data with interpolation; transform the daily weather status data into dummy variables. When there is light rain, sunny, cloudy or overcast weather, the weather status has no effect on passenger flow, that is, dummy=0, otherwise dummy=1; The Pearson cross-correlation analysis was used to analyze the correlation between the multiple network keyword variables collected and the multi-order lagged variables of each keyword variable. The tourism demand variable is y t , the i-th predictor variable is x it , the specific calculation method is as follows: Among them, S xy is the sample covariance between tourism demand and the t-th order lagged variable of the ith predictor variable, S x and S y denote the sample standard deviations of the predicted and predictor variables, respectively; Covariance: Standard Deviation: This analysis method is used to find the optimal lag structure of each predictor variable and select variables with correlation coefficients greater than the threshold, thus achieving the purpose of variable selection; S3. Data prediction To compare the prediction capabilities, LSTM network, back-propagation neural network BPNN and AR model were selected as benchmark models; The LSTM network adds a hidden state Ct, and the hidden layer also contains an input gate it, an output gate ot, and a forget gate ft, making it easy for information to flow unchanged; Where W is the input weight matrix of the hidden unit; b is the bias vector; subscripts f, i, o represent the forget gate, input gate, and output gate; ⊗ represents the point-by-point product operation; σ() is the activation function; And set two activation functions, which have more stable gradients: The BiLSTM neural network is a variant of the standard LSTM network. It consists of a forward LSTM and a backward LSTM network, providing a way to obtain sequence information in both the forward and backward input directions. Attention calculates the weights of the feature vectors output from the BiLSTM network at different time steps and assigns higher weights to more important features. The Attention layer is expressed as: Among them, Zt represents the hidden state, b is the bias parameter, and v represents the output vector calculated based on the weighted sum of the hidden states; The feedforward of BPNN network is that the input information starts from the input layer, and the output of each neuron in each hidden layer is calculated and passed to the next layer until the output layer calculates the output result of the network. The feedforward is only used to calculate the output of the network, and the parameters of the network are not adjusted. The back propagation algorithm is used to adjust the network weights and thresholds when training the model. This process is a supervised learning process. The feedback is used to find partial derivatives and execute the gradient descent algorithm to obtain the minimum value of the loss function, so that the error between the expectation and the output is reduced as much as possible. AR Model The AR model uses the historical information of tourism demand to build a model and predict future tourism demand. Assuming that the passenger flow time series y is a stationary time series, the AR model is expressed as Among them, φ0 is a constant term, φ i is the estimated parameter of the lag term, e t The mean is 0 and the variance is σ 2 The lag term of the model is determined by the AIC and BIC criteria. The BiLSTM-Attention multimodal fusion model is used to calculate the predicted data. The model consists of only two basic units, namely the bidirectional long short-term memory neural network and the Attention mechanism. The BiLSTM and bidirectional LSTM model is composed of a forward LSTM and a backward LSTM network. It uses multiple LSTM layers to process the periodically changing passenger flow information. The Attention mechanism can assign different weights according to the importance of the multi-source data input set at different time steps. It can learn the long-term dependency characteristics of passenger flow and capture the changing trend of passenger flow from two different directions.

[0006] S4. Warning and guidance Assume that the daily passenger flow of the tourist destination is y and the maximum carrying capacity is υmax. If the passenger flow exceeds this value, the alarm level η is calculated based on the real-time passenger flow and the maximum carrying capacity υmax. The specific calculation method is as follows: 1. Real-time passenger flow of the day , η calculation: , η ranges from (1, +∞); 2. Real-time passenger flow of the day ,η calculation: 1,η value range 1; 3. Real-time passenger flow of the day , η calculation: , η takes the value range [0,1).

[0007] Compared with the prior art, the method of the present invention has the following beneficial effects: (1) The BiLSTM-Attention multimodal deep learning prediction method can effectively improve the prediction accuracy of the model. This is because the BiLSTM unit can simultaneously identify the feature information of the input time series data in different directions, and the Attention mechanism can assign corresponding weights according to the importance of the input sequence at different time points. Meteorological data and network keywords significantly improve the prediction ability of the model.

[0008] (2) A reasonable early warning framework was constructed in terms of warning degree forecasting and guidance, the warning degree was defined, the maximum daily carrying capacity and maximum capacity of the scenic area were calculated, and then the predicted value generated by the model was used to calculate the warning level. This prediction method can provide more timely and effective early warning information for tourism-related departments and tourists, which is of great significance to the public safety and sustainable development of scenic spots. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 Comparison chart of daily traffic trends of different keyword searches Figure 2 Maximum and minimum temperatures and daily passenger flow trends Figure 3 Weather condition dummy variables and daily passenger flow trends Figure 4 Comparison of prediction curves of various models at different prediction steps Figure 5 It is the BiLSTM-Attention multimodal fusion model; Figure 6 It is a flow chart of the implementation steps of the technical solution of the present invention. DETAILED DESCRIPTION

[0010] In order to achieve the above purpose, the present invention takes Maiji Mountain Scenic Area in Tianshui City, Gansu Province as the research object and predicts the daily passenger flow it receives. This is achieved through the following technical scheme, and the specific steps are as follows: Data Collection According to the data provided by the Tianshui Maiji Mountain Scenic Area Management Committee, the daily passenger flow of Maiji Mountain Tourist Scenic Area fluctuates greatly. During the off-season, the passenger flow is relatively small, with daily passenger flow ranging from a few hundred people. During the peak season, the passenger flow increases, with daily passenger flow reaching tens of thousands. The periodic fluctuation characteristics between the off-season and peak-season are obvious, and it is of great significance to predict its daily passenger flow.

[0011] 1. Obtain the daily passenger flow data of Maiji Mountain Scenic Area (marked as y). The passenger flow data collection period is from March 1, 2023 to May 30, 2024. The passenger flow data is provided by the Tianshui Maiji Mountain Scenic Area Management Committee.

[0012] (II) The historical meteorological data and future meteorological forecast data of key places in tourist destinations are obtained free of charge through the weather network. The main meteorological data collected are weather status and temperature. Among them, the weather status variable (dummy) mainly includes various weather conditions such as light rain, cloudy, and showers. This data is a dummy variable; the temperature includes the daily maximum temperature T max and the minimum temperature T min .

[0013] The data collection period is from March 1, 2023 to May 30, 2024 (III) Use Python to crawl the Baidu Index and the keyword search values ​​related to food, accommodation, transportation, travel, shopping, entertainment and other elements of tourist destinations, remove the keywords with data of 0, and form a backup network keyword data set (x). The data collection time range is still from March 1, 2023 to May 30, 2024.

[0014] Data preprocessing and data analysis 1. Data preprocessing The data collected above are processed structurally to form standard data. As daily passenger flow data and network keyword data are very sensitive to emergencies, these abnormal points are smoothed and individual missing data are replaced by interpolation. The daily weather status data is converted into a dummy variable method. Specifically, if there is light rain, sunny, cloudy, overcast or other related conditions on a certain day, it is considered that the passenger flow has no effect or has a very small effect, that is, dummy=0, otherwise dummy=1.

[0015] 2. Data Analysis The linear correlation of the collected variables was analyzed to obtain the variables with strong correlation. For the 8 collected network keyword variables, we used Pearson cross-correlation analysis to explore the correlation between the predicted variables and the 0 to 120 order lagged variables of each keyword variable. Let the tourism demand variable be y t , the i-th predictor variable is x it , the specific calculation method is as follows: Among them, S xy is the sample covariance between tourism demand and the t-th order lagged variable of the ith predictor variable. x and S y represent the sample standard deviation of the predicted variable and the predictor variable respectively.

[0016] Covariance: Standard Deviation: This analysis method can be used to find the optimal lag structure of each predictor variable and select variables with correlation coefficients greater than the threshold, thereby achieving the purpose of variable selection. The custom threshold is as follows: Baidu Index keyword threshold X ≥ 0.80; Weather temperature maximum temperature threshold X ≥ 0.6; Weather temperature minimum temperature threshold X ≥ 0.7; Combine all predictor variables and predicted variables into a data set: The correlation analysis between the selected 8 keyword variables, meteorological data, and the 1st to 3rd order lagged variables of the predicted variables and the predicted variables is as follows: Note: ***, **, * are significant at the 1%, 5%, and 10% levels, respectively.

[0017] From the above table we can find that: ① The lag order of the keyword prediction variables obtained by the cross-correlation analysis is 1-3, and there is a significant correlation between these variables and the predicted variables (1% level), and the correlation coefficients are all above 0.80, which fully reflects the correlation between tourists' information search behavior and potential tourism demand. For example, most tourists will check travel guides and traffic information about 3 days in advance, search for hotel information about 2 days in advance, and check maps and attractions about 1 day in advance. Daily network search data more accurately reflects the dynamic characteristics of passenger flow and can predict changes in passenger flow in advance.

[0018] ②There is also a significant correlation between the maximum temperature and minimum temperature variables and the predicted variables, with correlation coefficients reaching 0.698 and 0.739 respectively, indicating that the temperature of the tourist destination among the meteorological factors is also an important influencing factor in tourists' travel decisions.

[0019] The dynamic change trend between the three keyword variables "Maiji Mountain", "Maiji Mountain Grottoes" and "Maiji Mountain Hot Spring" among the eight keywords and the actual daily passenger flow received was investigated.

[0020] from Figure 1 It can be seen intuitively that online information search and passenger flow show periodic linkage fluctuations, and in each cycle they show complex nonlinear characteristics. Specifically, in May and October, which are Labor Day and National Day respectively, passenger flow will increase dramatically. The dynamic fluctuation characteristics of each keyword are somewhat different, indicating that different keyword queries reflect tourists' potential travel needs from different aspects. Introducing different keywords with higher correlation coefficients can help provide richer dynamic information and improve the model's prediction ability.

[0021] Figure 2 and Figure 3 The dynamic fluctuation trend between the temperature and weather conditions virtual variables of Maiji Mountain and the actual passenger flow is presented respectively. Figure 2 It can be found that the fluctuation trend between the maximum temperature and the minimum temperature is basically the same, indicating that the daily temperature difference in Maiji Mountain is relatively stable, and the fluctuation range of the daily maximum temperature is larger than that of the minimum temperature. Both the minimum temperature and the maximum temperature show a strong correlation with the measured variables, and the fluctuation trend reflects the changes in passenger flow to a certain extent. Figure 3 It is the dynamic fluctuation between the daily weather status and the measured variable. As can be seen from the figure, in most cases the weather status variable reflects the status of the peak and trough periods of passenger flow.

[0022] Data prediction In order to compare the prediction ability, LSTM network, back propagation neural network (BPNN) and AR model are selected as benchmark models. Among them, BiLSTM is based on the BiLSTM-Attention model without the Attention mechanism. This model is introduced to illustrate whether the Attention mechanism in the BiLSTM-Attention model can effectively use information of different importance at different time points in prediction. BPNN is a shallow learning method. The purpose of introducing this model is to verify whether the deep learning method can effectively improve the prediction accuracy. The AR model is a simple linear regression model that only uses historical passenger flow information for prediction. The purpose of introducing this model is to illustrate the effectiveness of other data sources in prediction, and at the same time illustrate the prediction ability of human intelligence nonlinear methods in complex data situations.

[0023] 1. LSTM model LSTM provides a solution to fuse memory units, which allows the network to learn previously forgotten hidden states and update hidden states based on new information. In particular, in addition to hidden units, LSTM networks also include input gates, forget gates, input adjustment gates, and a memory cell unit, which fuses the previous memory cell states. These memory cells are adjusted by the forget gate and input adjustment gate together with the previous hidden body, and the hidden body is adjusted by the input gate. These additional memory units enable the LSTM architecture to recognize dependencies in distant times. This is because the LSTM algorithm can learn information that can be remembered and forgotten through the training process, thereby ensuring that the LSTM network has long-term memory capabilities.

[0024] The LSTM network adds a hidden state Ct, and the hidden layer also contains an input gate it, an output gate ot, and a forget gate ft, making it easy for information to be passed on unchanged.

[0025] Where W is the input weight matrix of the hidden unit; b is the bias vector; subscripts f, i, o represent the forget gate, input gate, and output gate; ⊗ represents the point-by-point product operation; σ() is the activation function.

[0026] And set two activation functions, which have more stable gradients: 2. BiLSTM model The BiLSTM neural network is a variant of the standard LSTM network. It consists of a forward LSTM and a backward LSTM network, providing a way to obtain sequence information in both the forward and backward input directions. The structure of the BiLSTM is a stack of two unidirectional LSTM networks, one of which is used for forward propagation and the other for backward propagation. Generally, multiple LSTM layers are used to process passenger flow information that undergoes periodic alternations, where each BiSTM network consists of two LSTM layers for forward and backward propagation. In this way, the BiLSTM neural network model can capture more information from two different directions, thereby improving the model's prediction performance.

[0027] 3. Attention Mechanism Attention calculates the weights of the feature vectors output from the BiLSTM network at different time steps and assigns higher weights to more important features. The Attention layer can be expressed as: Among them, Zt represents the hidden state, b is the bias parameter, and v represents the output vector calculated based on the weighted sum of the hidden states.

[0028] 4. BPNN model The feedforward of the BPNN network is that the input information starts from the input layer, and the output of each neuron in the layer is calculated through the neurons of each hidden layer and passed to the next layer until the output layer calculates the output result of the network. The feedforward is only used to calculate the output of the network, and the parameters of the network are not adjusted. The back propagation algorithm is used to adjust the network weights and thresholds when training the model. This process is a supervised learning process. Feedback is used to find partial derivatives and execute the gradient descent algorithm to obtain the minimum value of the loss function, so that the error between the expectation and the output is reduced as much as possible.

[0029] 5. AR Model The AR model is a common method for processing time series in statistics. The model uses the historical information of tourism demand to build a model and predict future tourism demand. Assuming that the passenger flow time series y is a stationary time series, the AR model can be expressed as Among them, φ0 is a constant term, φ i is the estimated parameter of the lag term, e t The mean is 0 and the variance is σ 2 The lag term of the model is determined by the AIC and BIC criteria. (VI) BiLSTM-Attention multimodal fusion model The BiLSTM-Attention multimodal fusion model is used to calculate the predicted data. The model consists of only two basic units, namely the bidirectional long short-term memory neural network and the Attention mechanism. The BiLSTM and bidirectional LSTM model is composed of a forward LSTM and a backward LSTM network, and uses multiple LSTM layers to process the passenger flow information that changes periodically. The Attention mechanism can assign different weights according to the importance of the multi-source data input set at different time steps. It can learn the long-term dependency characteristics of passenger flow, capture the changing trend of passenger flow from two different directions, and make full use of the dynamic feature information of the tourist destination to improve the prediction performance of the model.

[0030] (VII) Parameter settings Before the experiment, the basic parameters of the selected benchmark model were set. To avoid overfitting, except for BPNN, the hidden layer of most benchmark models selected Dropout with a random selection probability of 0.5 for regularization, which achieved or approached the optimal effect in most networks. The prediction step size of each model was set to h=1, 3, 6. To avoid missing the local optimal solution and ensure the learning complexity, the learning rate of all models was set to 0.01. Considering the small number of training samples, the Batch Size of the training samples was set to 20. In terms of activation function, the hidden layer of BPNN uses the tansig function as the activation function, and the output layer uses the pure linear function Pureline as the activation function; the output layers of DNN, DBN and LSTM models all use the ReLU function as the activation function, and all other layers use the tanh function as the activation function. The specific parameter settings of each model are as follows: (VIII) Prediction results The daily passenger flow of Maiji Mountain Scenic Area from June 1, 2024 to June 30, 2024 was predicted by using historical passenger flow, hot search keywords on the Internet, and meteorological data (three dimensions: daily maximum temperature, daily minimum temperature, and daily weather conditions).

[0031] from Figure 4 It can be seen that the prediction curves of each model are close to the actual passenger flow curve, but the prediction curve of the BiLSTM-Attention model deviates the closest to the actual passenger flow curve, and the degree of deviation at each point is relatively stable, while the prediction curves of the BPNN and AR models deviate far from the actual passenger flow curve, and the degree of deviation increases over time. Therefore, the BiLSTM-Attention model can better fit the actual daily passenger flow, while the BPNN and AR models are relatively poor at fitting actual data. The prediction results of other benchmark models are between BiLSTM-Attention and BPNN.

[0032] The comparison of the MAPE and PBIAS values ​​of each model at different prediction steps is shown in the following table: The table shows the following information: ① In terms of the MAPE index, the index value of BiLSTM-Attention at each step is the smallest and the performance is relatively stable. The index values ​​of the three benchmark models LSTM, BPNN and AR at a single prediction step (h=1, 3, 6) increase in turn. This means that the prediction method can accurately identify the bidirectional feature information of the time series input, control the importance of different time nodes of the input sequence, and then assign corresponding weights, so as to have the best prediction performance. ② In terms of the PBIAS index, the index value of the BiLSTM-Attention model performs best when the time domain is predicted one step ahead, and the index value of the BiLSTM model performs best when the time domain is predicted three and six steps ahead, and only when the time domain is predicted one step ahead, the values ​​of the BiLSTM and BPNN models on this index are negative. But overall, the index values ​​of each model at a single prediction step are very close to 0, indicating that the average deviation of the fitting values ​​of each model from the actual passenger flow value is small.

[0033] Based on the above performance, the following conclusions are drawn: ①The BiLSTM-Attention model has a bidirectional LSTM structure and utilizes the dynamic feature information of the input sequence from different directions. The model also recognizes the importance of the input sequence to the prediction at different times and assigns different weights, so it can fit the actual passenger flow data well. ②The BiLSTM benchmark model does not introduce the Attention mechanism, and its prediction ability is slightly weaker. ③The LSTM benchmark model only uses the information of the time series data in one direction and does not introduce the Attention mechanism, so its prediction ability is not as good as the BiLSTM model and the BiLSTM-Attention model. ④Although BPNN has good nonlinear prediction ability, it is a shallow learning method and is prone to local optimality and overfitting, so the prediction ability of the model is poor. ⑤The AR model only uses historical passenger flow information. Although the historical passenger flow meets the stability requirements, the model does not make full use of the dynamic feature information of other data sources, so the AR model performs the worst in all cases.

[0034] Warning and guidance Assuming that the daily passenger flow of a tourist destination scenic spot (venue) is y and the maximum carrying capacity is υmax, if the passenger flow exceeds this value, it will cause serious congestion in the scenic spot (venue), affecting sustainable development and even causing unpredictable economic losses. The alarm degree η is calculated based on the real-time passenger flow and the maximum carrying capacity υmax. The specific calculation method is as follows: 1. Real-time passenger flow of the day , η calculation: , η takes the value range of (1,+∞).

[0035] 2. Real-time passenger flow of the day ,η calculation: 1,η value range 1.

[0036] 3. Real-time passenger flow of the day , η calculation: , η takes the value range [0,1).

[0037] It is understood that the maximum daily capacity of Maiji Mountain Scenic Area is 9,600 people. Therefore, the maximum daily passenger capacity of Maiji Mountain Scenic Area is determined to be 9,000 people. The warning level of Maiji Mountain Scenic Area is defined as follows: The BiLSTM-Attention model is used to predict the passenger flow of Maiji Mountain Scenic Area from June 1 to June 10, 2024.

[0038] According to the classification of warning levels, when the passenger flow is less than the maximum capacity, the warning degree η is greater than or equal to 0 and less than 1, there is no abnormality in the passenger flow, no alarm occurs, the warning level is level three, and the color of the signal light is defined as green. If the value of η is greater than or equal to 1 and less than 1.5, it indicates that the tourist flow may reach the maximum carrying capacity of the scenic area. The warning level is level two, there is a mild alarm, and the signal light is yellow. If the value of η is greater than or equal to 1.5, the passenger flow exceeds the maximum carrying capacity, there is a severe alarm, the warning level is level one, and the signal light is red. As can be seen from the above table, there will be a level two alarm warning in Maiji Mountain Scenic Area on June 9, 2024. When an alarm occurs, customize the following SMS template: Tourists can receive text messages to view intelligent diversion information, and then choose areas with less crowds for sightseeing based on diversion routes.

[0039] Glossary Tourism demand time series data: refers to the data on tourism industry demand recorded and analyzed in the context of time series analysis. These data are usually collected according to a fixed period (such as day, week, month, quarter or year).

[0040] Break-even capacity: refers to the minimum number of tourists required for a scenic spot to cover all its fixed operating costs (such as maintenance costs, employee salaries, management fees, and other costs that do not change with the number of tourists) and achieve a balance between income and expenditure in the tourism industry.

[0041] Covariance: refers to a statistical indicator in statistics that measures the degree of linear correlation between two random variables. It describes the tendency of two variables to change together: if two variables tend to increase or decrease at the same time, the covariance between them is positive; if one variable tends to decrease when the other variable increases, the covariance is negative; if there is no obvious correlation between the changes in the two variables, the covariance is close to zero.

[0042] Standard deviation: It is an important indicator in statistics to measure the degree of dispersion of data distribution. It represents the average square root value of the difference (deviation) between the data and its mean.

[0043] Alert degree: The abnormal degree of passenger flow, that is, the abnormal degree of alert obtained through the predicted value and the carrying capacity of the tourist destination.

Claims

1. A tourist destination passenger flow diversion method based on multimodal data fusion analysis, characterized by The steps include: S1. Data Collection Obtain the daily passenger flow data set of the scenic spot and set the time range for passenger flow data collection; The historical meteorological data and future meteorological forecast data of key places in the tourist destination within the time range are obtained through the Internet, and two kinds of meteorological data, weather status and temperature, are collected. The weather status variables include light rain, sunny, cloudy, and showers; the temperature includes the daily maximum temperature T max and the minimum temperature T min； Use Python to crawl keyword search values ​​related to tourist destinations, remove keywords with zero data, and form a backup network keyword data set; S2. Data preprocessing and data analysis Smooth the abnormal points of the above data set, and replace individual missing data with interpolation; transform the daily weather status data into dummy variables. When there is light rain, sunny, cloudy or overcast weather, the weather status has no effect on passenger flow, that is, dummy=0, otherwise dummy=1; The collected multiple network keyword variables are analyzed by Pearson cross-correlation analysis to predict the correlation between the variables and the multi-order lagged variables of each keyword variable; let the tourism demand variable be y t , the i-th predictor variable is x it , the specific calculation method is as follows: Among them, S xy is the sample covariance between tourism demand and the t-th order lagged variable of the ith predictor variable, S x and S y denote the sample standard deviations of the predicted and predictor variables, respectively; Covariance: Standard Deviation: This analysis method is used to find the optimal lag structure of each predictor variable and select variables with correlation coefficients greater than the threshold, thus achieving the purpose of variable selection; S3. Data prediction To compare the prediction capabilities, LSTM network, back-propagation neural network BPNN and AR model were selected as benchmark models; The LSTM network adds a hidden state Ct, and the hidden layer also contains an input gate it, an output gate ot, and a forget gate ft, making it easy for information to flow unchanged; Where W is the input weight matrix of the hidden unit; b is the bias vector; subscripts f, i, o represent the forget gate, input gate, and output gate; ⊗ represents the point-by-point product operation; σ() is the activation function; And set two activation functions, which have more stable gradients: The BiLSTM neural network is a variant of the standard LSTM network. It consists of a forward LSTM and a backward LSTM network, providing a way to obtain sequence information in both the forward and backward input directions. Attention calculates the weights of the feature vectors output from the BiLSTM network at different time steps and assigns higher weights to more important features. The Attention layer is expressed as: Among them, Zt represents the hidden state, b is the bias parameter, and v represents the output vector calculated based on the weighted sum of the hidden states; The feedforward of BPNN network is that the input information starts from the input layer, and the output of each neuron in each hidden layer is calculated and passed to the next layer until the output layer calculates the output result of the network. The feedforward is only used to calculate the output of the network, and the parameters of the network are not adjusted. The back propagation algorithm is used to adjust the network weights and thresholds when training the model. This process is a supervised learning process. The feedback is used to find partial derivatives and execute the gradient descent algorithm to obtain the minimum value of the loss function, so that the error between the expectation and the output is reduced as much as possible. AR Model The AR model uses the historical information of tourism demand to build a model and predict future tourism demand. Assuming that the passenger flow time series y is a stationary time series, the AR model is expressed as Among them, φ0 is a constant term, φ i is the estimated parameter of the lag term, e t The mean is 0 and the variance is σ 2 The lag term of the model is determined by the AIC and BIC criteria. The BiLSTM-Attention multimodal fusion model is used to calculate the predicted data. The model consists of only two basic units, namely the bidirectional long short-term memory neural network and the Attention mechanism. The BiLSTM and bidirectional LSTM models are composed of forward LSTM and backward LSTM networks, and use multiple LSTM layers to process periodically changing passenger flow information. The Attention mechanism can assign different weights according to the importance of multi-source data input sets at different time steps. It can learn the long-term dependency characteristics of passenger flow and capture the changing trend of passenger flow from two different directions.

2. The method for guiding passenger flow at a tourist destination based on multimodal data fusion analysis according to claim 1 is characterized in that Also includes: S4. Warning and guidance Assume that the daily passenger flow of the tourist destination is y and the maximum carrying capacity is υmax. If the passenger flow exceeds this value, the alarm level η is calculated based on the real-time passenger flow and the maximum carrying capacity υmax. The specific calculation method is as follows:

1. Real-time passenger flow of the day , η calculation: , η ranges from (1, +∞); 2. Real-time passenger flow of the day ,η calculation: 1,η value range 1; 3. Real-time passenger flow of the day , η calculation: , η takes the value range [0,1).

Images