A scenic spot passenger flow prediction method based on multi-source information fusion

Abstract

The application provides a scenic spot passenger flow prediction method based on multi-source information fusion, and belongs to the field of deep learning. The method first collects multi-source heterogeneous data including weather, holidays and passenger flow of adjacent scenic spots, and carries out pretreatment, feature extraction and dimension reduction on the multi-source heterogeneous data, so that more compact and effective feature representation is obtained; then, a feature-level attention mechanism and a time attention mechanism are introduced to highlight the importance of key features and key time slices; finally, a long short-term memory network (LSTM) is combined to model the time sequence features, so that dynamic and accurate prediction of the passenger flow of the scenic spot is realized, thereby providing strong support for the visitor management and resource scheduling of the scenic spot.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of deep learning, and particularly relates to a scenic spot passenger flow prediction method based on multi-source information fusion. BACKGROUND

[0002] With the development of tourism, the passenger flow of various scenic spots increases rapidly. The problems of passenger flow oversaturation and unreasonable resource allocation are increasingly prominent. The existence of these problems not only affects the touring experience of tourists, but also poses a potential threat to the life and property safety of tourists. Therefore, how to effectively manage the passenger flow of scenic spots has become a problem that needs to be solved in the field of public safety at present.

[0003] As a key technology for improving the management level of scenic spots, passenger flow prediction has been widely used in various scenic spots. By predicting the passenger flow of scenic spots in real time, management personnel can master the passenger flow change in each period in advance, so as to realize the optimized deployment of multiple management measures such as reservation flow control, resource allocation, personnel scheduling and security warning, thereby improving the touring experience of tourists and effectively ensuring the safe operation of scenic spots. The current common passenger flow prediction methods can be mainly divided into traditional statistical methods, machine learning methods and deep learning methods.

[0004] Among them, the traditional statistical method mainly analyzes and models the historical data to realize the prediction of passenger flow. This method is easy to implement, but it is difficult to capture the complex nonlinear relationship and dynamic time sequence characteristics in the passenger flow data in the application process, so there are obvious deficiencies in the prediction accuracy. In comparison, the machine learning method improves the nonlinear modeling ability on the basis of the traditional statistical method, and can to some extent mine the potential patterns and characteristic relationships in the passenger flow data, thereby improving the prediction effect, but there are still bottlenecks in processing long time sequence dependence or multi-scale characteristics. In recent years, the deep learning method has been widely used in passenger flow prediction tasks. This method effectively overcomes the deficiencies of traditional statistical methods and machine learning methods in nonlinear modeling ability and time sequence dependence processing, and significantly improves the passenger flow prediction accuracy. However, the existing deep learning-based methods mainly focus on long-time-scale passenger flow prediction (such as annual, monthly and daily), which is difficult to meet the demand of high-time-resolution prediction (such as hourly) for daily operation of scenic spots. At the same time, most of the researches ignore the influence of weather factors and passenger flow of adjacent scenic spots on the passenger flow of the target scenic spot, which limits the prediction accuracy and practicability of the model. SUMMARY

[0005] The application provides a scenic spot passenger flow prediction method based on multi-source information fusion, which specifically comprises the following steps:

[0006] Step 1, collecting passenger flow, weather conditions and holiday information from the scenic spot management system, the meteorological service platform and the calendar system, and constructing a multi-source heterogeneous data set;

[0007] Step 2, pre-processing and standardization operation is performed on the collected multi-source heterogeneous data to construct a standardized data set;

[0008] Step 3, using the self-encoder to extract features and reduce dimension of the surrounding scenic spot passenger flow information, remove redundant information, and obtain a more expressive low-dimensional feature representation;

[0009] Step 4, introduce a feature-level attention mechanism to dynamically evaluate and calculate the importance weight of each type of feature in the prediction task, and construct a weighted fusion unified feature sequence;

[0010] Step 5, the weighted fused features and the historical passenger flow information of the research scenic spot are spliced and input into the long short-term memory network, and the time attention mechanism is introduced to highlight the contribution of key time points to the prediction result;

[0011] Step 6, connect the full connection layer to output the final passenger flow prediction result.

[0012] Compared with the prior art, the application has the following beneficial effects:

[0013] The application fully considers various influencing factors such as the passenger flow of adjacent scenic spots, weather and holiday information in the process of scenic spot passenger flow prediction, and significantly improves the prediction accuracy of scenic spot passenger flow by comprehensively considering these factors.

[0014] The application proposes a scenic spot passenger flow prediction model based on AE-LSTM-Attention combination. The model fully integrates the advantages of autoencoder (AE), long short-term memory network (LSTM) and attention mechanism (Attention), and improves the modeling ability of passenger flow changes in complex scenarios. Among them, the autoencoder (AE) is used to compress and de-redundant high-dimensional multi-source features, and extract representative low-dimensional features; the long short-term memory network (LSTM) is used to model the time evolution law of passenger flow data and improve the modeling ability of long-term dependent information; the attention mechanism (Attention) is used to dynamically allocate the weights of different features and key moments, further improve the response ability and prediction accuracy of the model in complex change scenarios, so as to realize efficient and accurate prediction of scenic spot passenger flow.

[0015] The application realizes the system integration from multi-source data collection, preprocessing, feature extraction to deep model prediction by constructing a complete data processing and modeling process, has good scalability and universality, and can be widely used in passenger flow prediction scenarios of different types of scenic spots, and provides scientific basis and technical support for scenic spot tourist management, service scheduling and safety warning. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1The flow chart of the multi-source data preprocessing.

[0017] Figure 2 The flow chart of the multi-source data preprocessing.

[0018] Figure 3 The schematic diagram of the encoding and decoding process of the autoencoder.

[0019] Figure 4 The flow chart of the scenic spot passenger flow prediction.

[0020] Figure 5 The schematic diagram of the AE-LSTM-Attention model structure.

[0021] Figure 6 The comparison chart of the original flow and the predicted flow of the scenic spot.

[0022] Figure 7 The comparison chart of the prediction results of the model and the baseline model. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings of the present application. Obviously, the described embodiments are only a part of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0024] Please refer to Figures 1 to 7 The present application provides a scenic spot passenger flow prediction method based on multi-source information fusion, which aims to realize accurate and efficient prediction of scenic spot passenger flow by fusing multi-source heterogeneous data, and specifically includes the following steps:

[0025] Step 1, collect passenger flow, weather conditions and holiday information from the scenic spot management system, meteorological service platform and calendar system, and the specific steps of constructing the multi-source heterogeneous data set include:

[0026] Step 1.1, from the scenic spot management system, collect the cumulative passenger flow information of the target scenic spot and its adjacent scenic spots at the entrance and exit of the gate from July 1, 2022 to July 1, 2024, including a total of 70176 data;

[0027] Step 1.2, collect rainfall and temperature information from the meteorological service platform, including a total of 17544 data;

[0028] Step 1.3, call the holiday information query interface to obtain holiday information, including a total of 731 data.

[0029] Step 2, preprocessing and standardization of collected multi-source heterogeneous data, the specific steps of constructing standardized data set include:

[0030] Step 2.1, the time stamp of passenger flow, weather and holiday data is uniformly converted into standard format, and is aligned according to unified time granularity, taking hour as the basis, and the low frequency data such as holiday is up sampled and aligned to hour level;

[0031] Step 2.2, the units of indicators in different data sources are consistent, the temperature is unified to Celsius, and the rainfall is unified to millimeter;

[0032] Step 2.3, the collected real-time passenger flow data, weather information and holiday information are processed for missing value, among which the passenger flow, temperature and rainfall are filled by mean filling method, and the holiday information is filled by forward filling method;

[0033] Step 2.4, the multi-source data is normalized, for the scenic spot passenger flow data, Min-Max normalization method is adopted to map the passenger flow data to [0, 1] interval, and the normalization formula is recorded as:

[0034] ;

[0035] Among them, x represents the current passenger flow, and respectively represent the maximum and minimum passenger flow, for temperature, Z-score standardization method is adopted for processing, and the formula is recorded as:

[0036] ;

[0037] Among them, x is the current temperature, is the mean temperature of the region, is the standard deviation of the temperature corresponding to the region, for rainfall information, first log standardization is carried out to compress the right biased distribution of data, and then Z-score standardization is carried out, and the log transformation formula is recorded as:

[0038] ;

[0039] Among them, x is the current rainfall, c is a constant, which avoids the case of , for holiday information, binary coding method is adopted for normalization, and the coding rule is as follows:

[0040] ;

[0041] Among them, represents the normalized holiday data.

[0042] Step 3, feature extraction and dimensionality reduction of the surrounding scenic spot passenger flow information using autoencoder, removing redundant information, obtaining more expressive low-dimensional feature representation, the specific steps include:

[0043] Step 3.1, arrange the obtained surrounding scenic spot passenger flow data in time sequence, and construct the original input sequence;

[0044] Step 3.2, build an autoencoder model, the encoder part is composed of multiple layers of fully connected neural network, used to compress the input high-dimensional time series data into low-dimensional latent feature representation, the decoder is responsible for reconstructing the original input from the low-dimensional feature, realizing self-supervised learning, in this embodiment, the encoder and the decoder are both composed of 4 layers of fully connected network;

[0045] Step 3.3, train the autoencoder model using layer-by-layer unsupervised pre-training method, optimize the parameters of the encoder and the decoder, and minimize the reconstruction error as the training target, the loss function is recorded as:

[0046] ;

[0047] Wherein, is the original input data sample, is the reconstruction result, and T is the total number of samples;

[0048] Step 3.4, input the passenger flow data into the trained autoencoder, and obtain the corresponding low-dimensional feature sequence through the encoder, recorded as:

[0049] ;

[0050] Wherein, T is the length of time series, that is, the time step number of passenger flow data, and Z is the feature representation of the entire passenger flow time series in low-dimensional space, which can be used as the input of subsequent prediction or analysis.

[0051] Step 4, introduce feature-level attention mechanism, dynamically evaluate and calculate the importance weight of each type of feature in the prediction task, and construct the specific steps of the weighted fusion of the unified feature sequence include:

[0052] Step 4.1, linearly transform the weather information rainfall , temperature , holiday information And the passenger flow characteristics of surrounding scenic spots (extracted by autoencoder) Project it into the unified latent space for subsequent weighted fusion operation, for example, the embedding expression formula of rainfall is recorded as:

[0053] ;

[0054] Wherein, This is the weight matrix. For bias terms, An embedded representation of weather features;

[0055] Step 4.2: Introduce a learnable attention query vector v, and calculate the attention score through dot product operation. Taking weather features as an example, the score calculation formula is denoted as:

[0056] ;

[0057] in, This is the transpose of the attention vector, used to score the input features. The larger the value, the more important the feature; similarly, it can be calculated... , , To facilitate comparison of the relative importance of different features, the scores above are subjected to Softmax normalization to obtain the final weight coefficients, denoted as:

[0058] ;

[0059] This normalization process ensures that the sum of the weights of all features is 1, thus forming a probability distribution;

[0060] Step 4.3: Based on the weight coefficients output by the attention mechanism, perform a weighted summation of the original three types of features to construct a unified fused feature sequence, denoted as:

[0061] ;

[0062] After fusion It not only retains the effective information of various features, but also highlights the feature categories that contribute the most to the prediction through the attention mechanism, thereby enhancing the model's ability to perceive environmental changes, holiday impacts, and regional linkages.

[0063] Step 5, which involves concatenating the weighted and fused features with historical visitor flow information of the research area and inputting the result into a Long Short-Term Memory (LSTM) network, and introducing a time attention mechanism to highlight the contribution of key time points to the prediction results, includes the following specific steps:

[0064] Step 5.1: Weighted fusion feature sequence Studying the historical visitor flow sequence of scenic spots The sequences are concatenated to form a new input sequence. The data is fed into a Long Short-Term Memory (LSTM) network to obtain the hidden states at each time step. ;

[0065] Step 5.2, for each time step Calculate its attention score to the final prediction target. This is achieved through a trainable attention function, denoted by the following formula:

[0066] ;

[0067] Where W, v, and b are learnable parameters;

[0068] Step 5.3: Attention scores for all time steps. Attention weights are obtained by normalization using the Softmax function. ;

[0069] Step 5.4: The hidden states at each time step are weighted and summed according to the attention weights to obtain the aggregated context vector c, denoted by the following formula:

[0070] ;

[0071] in, This represents the hidden state vector at time step t. Let be the attention weight corresponding to the t-th time step.

[0072] Step 6, connecting the fully connected layer to output the final passenger flow prediction result, includes the following steps:

[0073] Step 6.1: Input the context vector c into a fully connected (FC) neural network layer, which is responsible for mapping the abstract hidden state to the actual prediction numerical space. The specific mapping process uses a linear transformation method, and the calculation formula is denoted as:

[0074] ;

[0075] in, and These are the weight matrix and bias terms of the fully connected layer, respectively.

[0076] Step 6.2: Because the input data underwent normalization preprocessing before model training, the model's output prediction results... In reality, it is a dimensionless relative value, which needs to be restored to the actual passenger flow range through inverse normalization to restore the original data's dimensions. The calculation formula is denoted as:

[0077] ;

[0078] in, and These represent the maximum and minimum passenger flow values ​​in the training data, respectively.

Claims

1. A scenic spot passenger flow prediction method based on multi-source information fusion, characterized in that, The method comprises the following steps: Step 1, collecting the traffic flow, weather conditions and holiday information from the scenic spot management system, weather service platform and calendar system to construct a multi-source heterogeneous data set; Step 2, preprocessing and standardizing the collected multi-source heterogeneous data to construct a standardized data set; Step 3, using a self-encoder to extract features and reduce dimensions of the surrounding scenic spot traffic flow information, remove redundant information, and obtain a more expressive low-dimensional feature representation; Step 4, introducing a feature-level attention mechanism to dynamically evaluate and calculate the importance weight of each type of feature in the prediction task, and constructing a weighted fusion unified feature sequence; Step 4.1, the rainfall R = [R1, R2,..., R T ] T , temperature W = [W1, W2,..., W T ] T , holiday information H = [H1, H2,..., H T ] T and the surrounding scenic spot passenger flow characteristics Z = [Z1, Z2,..., Z T ] T By linear transformation projection to the unified potential space, it is convenient to carry out subsequent weighted fusion operation, and the rainfall embedding expression formula is recorded as: f R = tanh(W R · R + b R ); where W R is a weight matrix, b R is a bias term, and f R is an embedding representation of the rainfall amount; Step 4.2, introduce a learnable attention query vector v, and calculate the attention score through dot product operation, the rainfall feature score calculation formula is: e R = v T · f R ; where v T is the transpose of the attention vector, used to score the input features, with the score e R being larger indicating that the feature is more important, and similarly the temperature feature score e W , holiday feature score e H , and surrounding scenic spot visitor flow feature score e Z is obtained. For ease of comparison of the relative importance between different features, the above scores are subjected to Softmax normalization processing to obtain the final rainfall weight coefficient, denoted as: This normalization process ensures that the sum of the weights of each feature is 1, forming a probability distribution; Step 4.3, according to the weight coefficient output by the attention mechanism, the original four types of features are weighted and summed to construct a unified fusion feature sequence, denoted as: F' = a R • R + a W • W + a H • H + a Z • Z; wherein, α R , α W , α H and α Z respectively represent the attention weight coefficients corresponding to rainfall, temperature, holiday information and the passenger flow of surrounding scenic spots. The fused F' not only retains the effective information of various features, but also highlights the feature categories that contribute most to the prediction through the attention mechanism, enhancing the model's perception of environmental changes, holiday impacts and regional linkages. Step 5, concatenate the weighted and fused features with the historical traffic flow information of the research scenic spot, and input it into the long short-term memory network, and introduce a time attention mechanism to highlight the contribution of key time points to the prediction result; Step 5.1, concatenate the weighted fusion feature sequence F' = {F'1, F'2,..., F' T} and the historical visitor flow sequence X = {X1, X2,..., X T} of the scenic spot under study to form a new input sequence X' = [X, F'], and input into the long short-term memory network to obtain the hidden state h t of each time step. T} and the historical visitor flow sequence X = {X1, X2,..., X T} of the scenic spot under study to form a new input sequence X' = [X, F'], and input into the long short-term memory network to obtain the hidden state h t of each time step. t ; Step 5.2, for each time step h t Compute its attention score e on the final prediction target t , implemented by a trainable attention function, formula written as: e t = v T tanh(Wh t + b); Where W, v, b are learnable parameters; Step 5.3, normalizing the attention scores for all time steps by a Softmax function to obtain attention weights a t ;​ Step 5.4, according to the attention weight, the hidden states of each time step are weighted and summed to obtain the aggregated context vector c, and the formula is: wherein h t denotes the hidden state vector at the t-th time step, α t is the attention weight corresponding to the t-th time step; Step 6, connect the full connection layer to output the final traffic flow prediction result. 2.The scenic spot passenger flow prediction method based on multi-source information fusion according to claim 1, characterized in that, In step 1, the specific steps of collecting the traffic flow, weather conditions and holiday information from the scenic spot management system, weather service platform and calendar system to construct a multi-source heterogeneous data set include: Step 1.1, collect the cumulative traffic flow information of the target scenic spot and adjacent scenic spots at the entrance and exit every hour from July 1, 2022 to July 1, 2024; Step 1.2, collect rainfall and temperature information from the weather service platform; Step 1.3, call the holiday information query interface to obtain holiday information. 3.The scenic spot passenger flow prediction method based on multi-source information fusion of claim 1, characterized in that, In step 2, the specific steps of preprocessing and standardizing the collected multi-source heterogeneous data to construct a standardized data set include: Step 2.1, unify the timestamps of the traffic flow, weather and holiday data into a standard format and perform time alignment; Step 2.2, unify the units of indicators in different data sources; Step 2.3, perform missing value processing on the collected real-time traffic flow data, weather information and holiday information, wherein the traffic flow, temperature and rainfall are filled using the mean filling method, and the holiday information is filled using the forward filling method; Step 2.4, normalize the multi-source data, for the scenic spot traffic flow data, use the Min-Max normalization method to map the traffic flow data to the [0, 1] interval, the normalization formula is: where q represents the current passenger flow, q max and q min respectively represent the maximum and minimum passenger flow, and for the temperature, the Z-score standardization method is used for processing, and the formula is recorded as: Where w is the current temperature, μ is the mean temperature, and σ is the standard deviation corresponding to the temperature, for rainfall information, first log-standardize to compress the right-skewed distribution of the data, then perform Z-score standardization, the log transformation formula is: r' = log(r + c); Wherein, r is the current rainfall, c is a constant, to avoid the situation of r+c=0, for holiday information, normalized by binary coding, coding rules as follows: wherein, represents normalized holiday data.

4. The method of claim 1, wherein, In step 3, the self-encoder is used to extract features and reduce dimension of the surrounding scenic spot passenger flow information, remove redundant information, and obtain a more expressive low-dimensional feature representation. The specific steps include: Step 3.1, arrange the obtained surrounding scenic spot passenger flow data in time sequence to construct the original input sequence; Step 3.2, construct a self-encoder model, the encoder part is composed of multiple layers of fully connected neural network, used to compress the input high-dimensional time series data into low-dimensional latent feature representation, the decoder is responsible for reconstructing the original input from the low-dimensional feature, realizing self-supervised learning; Step 3.3, train the self-encoder model using layer-by-layer unsupervised pre-training method, optimize the parameters of the encoder and decoder, and minimize the reconstruction error as the training target, the loss function is denoted as: wherein X i is the original input data sample, is its reconstruction result, and N is the total number of samples. Step 3.4, input the surrounding scenic spot passenger flow data into the trained self-encoder, and obtain the corresponding low-dimensional feature sequence through the encoder, denoted as: Z = [Z1, Z2,..., Z T ] T ; Wherein, T is the time series length, that is, the time step number of passenger flow data, Z is the feature representation of the entire passenger flow time series in low-dimensional space, which can be used as input for subsequent prediction or analysis.

5. The method of claim 1, wherein, In step 6, the specific steps of connecting the fully connected layer to output the final passenger flow prediction result include: Step 6.1, input the context vector c into a fully connected neural network layer, which is responsible for mapping the abstract hidden state to the actual prediction value space, denoted as: where W fc and b fc are the weight matrix and bias term of the fully connected layer, respectively. Step 6.2, since the input data is normalized before model training, the prediction result output by the model is is actually a dimensionless relative value, which needs to be restored to the actual passenger flow range through inverse normalization operation, and the dimension of the original data is recovered. The calculation formula is recorded as: where q max and q min are the maximum and minimum values of the traffic in the training data, respectively.

Images