Airport terminal security check passenger number hourly prediction method

Abstract

The invention discloses an airport terminal security check passenger number hourly prediction method comprising the following steps: 1) preprocessing security check passenger flow characteristic data and terminal historical security check passenger data, and merging the preprocessed security check passenger flow characteristic data and terminal historical security check passenger data into a passenger security check number prediction data set; 2) constructing an NBMT-OneNet-based hourly security check passenger number prediction model; 3) training the hourly security check passenger number prediction model by using the passenger security check number prediction data set; and 4) obtaining security check passenger data and security check passenger flow characteristic data of the terminal to be predicted in the past T time period, and inputting the data into the hourly security check passenger quantity prediction model to obtain security check passenger data in the future T'time period. According to the characteristics of hourly airport security check data, the N-BEATS model and the MTGNN model are used for capturing the complex correlation between the time sequence trend and the variable, and the accuracy of the prediction result is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of airport security passenger quantity prediction, and particularly relates to an airport security passenger quantity hourly prediction method. BACKGROUND

[0002] Under the background of the increasing number of passengers, airport upgrading is imminent, and airport intelligent development has become an inevitable trend of global aviation development, and optimizing ground service process is an important component thereof. Ground service refers to various services provided by airports, airlines or agents for aircraft, passengers and consignees within the airport. As an important part of airport operation and management, the operation quality of ground service process directly affects the safety and operation efficiency of the airport. The scale and resource allocation of ground service are closely related to air passenger flow, and security check is an important part of ground service process. Undoubtedly, airport security passenger flow prediction not only has important reference value for route planning, flight state adjustment and operation decision, but also helps to optimize the allocation of limited resources and provide better ground service. Lack of accurate security passenger flow prediction may cause low operation efficiency, poor passenger experience and increased safety risks. Therefore, accurate security passenger flow prediction results have important value for improving the operation efficiency and service quality of the airport.

[0003] The existing models for airport passenger flow prediction still have some deficiencies. For example, the influence of temperature, holidays, flight plans and major events on passenger flow is not fully considered. And existing researches mainly focus on medium and long-term passenger flow prediction, without combining with actual operation needs of the airport to predict real-time passenger flow, which limits the limitations of these models in actual application. SUMMARY

[0004] The purpose of the present application is to provide an airport security passenger quantity hourly prediction method, which comprises the following steps:

[0005] 1) Obtain historical security passenger data of the airport, and security passenger flow feature data in the same period as the security passenger flow data;

[0006] 2) Preprocess the security passenger flow feature data and the historical security passenger data of the airport, and combine the preprocessed security passenger flow feature data and the historical security passenger data of the airport into a passenger security quantity prediction data set;

[0007] 3) Construct a NBMT-OneNet-based hourly security passenger quantity prediction model;

[0008] 4) Train the hourly security passenger quantity prediction model using the passenger security quantity prediction data set;

[0009] 5) Obtain the security passenger data and security passenger flow feature data of the terminal in the past T time period, and input them into the hourly security passenger quantity prediction model to obtain the security passenger data in the future T' time period.

[0010] Further, in step 1), the security passenger flow feature data includes airport operation features, time features, and weather features.

[0011] The airport operation features include hourly historical check-in data and boarding data.

[0012] The time features include daily patterns, weekly patterns, and holidays.

[0013] The weather features include hourly historical weather, temperature, wind speed, and precipitation.

[0014] Further, in step 2), the preprocessing includes data alignment, missing value processing, and LOF anomaly processing.

[0015] Further, in step 3), the hourly security passenger quantity prediction model includes an N-BEATS model and an OCP module.

[0016] The N-BEATS model is used for time-dependent modeling.

[0017] The MTGNN is used for cross-variable-dependent modeling.

[0018] The OCP module fuses the outputs of the N-BEATS and MTGNN models to obtain the security passenger quantity prediction result.

[0019] Further, in step 3.1), the N-BEATS model includes multiple stack modules; each stack module includes multiple blocks connected in series.

[0020] The input of the N-BEATS model is the security passenger flow data, and the output is the sum of the outputs of all stack modules.

[0021] In each stack module, the input x l of the first block is the original input sequence, and the output includes the predicted value for the future window H and the reconstructed value for the block input The input of the remaining blocks is the residual of the previous block input minus the reconstructed value of the previous block output.

[0022] Each block includes a fully connected network and a mapping layer.

[0023] The fully connected network includes multiple fully connected layers and linear layers, which are used to map the input to the forward unfolding coefficient and the backward expansion coefficient The mapping layer uses the mapping function to transform the forward expansion coefficients and the backward expansion coefficient Convert to prediction results;

[0024] The output of the lth block is as follows:

[0025] h l,1 =FC l,1 (x l )(1)

[0026] h l,2 =FC l,2 (h l,1 )(2)

[0027] h l,3 =FC l,3 (h l,2 )(3)

[0028] h l,4 =FC l,4 (h l,3 )(4)

[0029]

[0030] Where LINEAR represents the linear projection layer; FC represents the nonlinear standard fully connected layer with RELU; represents the prediction basis vector, is the backtracking basis vector, yes The i-th element of h l,1 、h l,2 、h l,3 、h l,4 Represents the output of different fully connected layers;

[0031] After each stack module obtains the output of each block, it merges the output of each block through a double-layer residual structure to obtain:

[0032]

[0033] Where, is the merged output.

[0034] Furthermore, the input of the MTGNN model is the security inspection passenger flow data X={z1,z2,z3,...,z P}, the output is the number of passengers Y to be screened in the Qth time step in the future;

[0035] The MTGNN model comprises a graph learning layer, m graph convolution modules, m time convolution modules and an output module;

[0036] The graph learning layer mines the potential relationship between nodes, calculates an adjacency matrix of the graph, and passes the adjacency matrix as input to all the graph convolution modules;

[0037] The graph convolution modules and the time convolution modules are arranged alternately, and respectively capture dimensional and time dependencies;

[0038] Residual connections are added between the outputs of the graph convolution modules and the inputs of the time convolution modules, and a skip connection is added after each time convolution module;

[0039] The output module maps hidden features to a target dimension to generate a final prediction result.

[0040] Further, the graph learning layer of the MTGNN model captures hidden relationships in the security passenger flow sequence by adaptively learning the adjacency matrix, that is:

[0041] M1 = tanh (aE1 Θ1) (11)

[0042] M2 = tanh (aE2 Θ2) (12)

[0043]

[0044] for i = 1, 2, 3,.., N (14)

[0045] idx = argtopk (A[i, :]) (15)

[0046] A[i, -idx] = 0 (16)

[0047] In the formula, E1 represents an original node feature matrix; E2 represents a target node feature matrix; Θ1 and Θ2 are two independent learnable weight matrices; a is a hyperparameter for controlling the saturation rate of an activation function; argtopk(·) is used to return the index of the top k maximum values of a vector; A is an adjacency matrix; M1 and M2 are feature matrices for capturing node potential relationships; i is a node index in the adjacency matrix; idx is the top k key neighbor index of node i;

[0048] The graph convolution module is used to fuse node and neighbor information to process spatial dependencies in the graph;

[0049] The graph convolution module comprises two mix-hop propagation layers;

[0050] The output of the graph convolution module is as follows:

[0051]

[0052] H out = Δ(H (2) H (1) ) = H 2 -H 1 (19)

[0053] where β is a hyper-parameter to control the ratio of keeping the original state of root node; k is the propagation depth; H in represents the input hidden state output by the previous layer; H (0) = H in ; is the normalized adjacency matrix; A is the original adjacency matrix, I is the unit matrix, is the normalized degree matrix; H out represents the output hidden state of the current layer; the parameter matrix W (k) serves as a feature selector; H 2 , H 1 represent the hidden state with propagation depth 2, the hidden state with propagation depth 1; H (k) is the hidden state with propagation depth k;

[0054] The time convolution module extracts high-order time sequence features through a standard dilated one-dimensional convolution filter;

[0055] The time convolution module includes a filtering layer, a gating layer, a dilated initial layer, a skip connection layer and an output layer;

[0056] Wherein, the filtering layer uses the hyperbolic tangent activation function to capture the time sequence pattern, and the gating layer adopts the S-shaped activation function to dynamically adjust the amount of information transmitted by the filter to the next module; the dilated initial layer is filtered by the filter;

[0057] The skip connection layer is used to normalize the information so that the sequence length input to the output layer is the same. The output layer includes two 1x1 convolution layers, which output values of a specified dimension.

[0058] The dilated initial layer of the time convolution module outputs z' as follows:

[0059] z' = concat(z*f 1×2 , z*f 1×3 , z*f 1×6 , z*f 1×7 ) (20)

[0060]

[0061] where d is the dilation factor. t is the time point currently calculated; s is the position index of the filter; z ∈ R T is a given one-dimensional input sequence; f 1×2 , f1×3 , f 1×6 , f 1×7 are filters of different sizes.

[0062] Further, the OCP module maintains a long-term weight w of the hourly security passenger flow prediction in the day through an exponential gradient descent algorithm, sets different short-term weights b through the RvS framework, and then fuses the long-term weight w and the short-term weight b into a model optimal weight and fuses the weight and the prediction result; wherein represents a prediction result of a time-dependent model, represents a prediction result of a cross-variable-dependent model, represents a prediction result of a cross-variable-dependent model, represents a prediction result of a cross-variable-dependent model, represents a prediction result of a cross-variable-dependent model;

[0063] The exponential gradient descent algorithm is as follows:

[0064] Δ = {w t | w t,i ≥ 0 and ||w t ||1 = 1} (22)

[0065] In the formula, t is a time step indicator; Δ is a gradient; w t,i is a weight.

[0066] Further, in the OCP module, the long-term weight w, the short-term weight b, and the model weight are as follows:

[0067]

[0068] In the formula, l t,i represents a loss of f i at time step t; is a fusion factor; f i (x) represents a prediction output of the current round of the prediction model on the input sample;

[0069]

[0070] Further, the hourly security passenger flow prediction model is trained using a passenger security quantity prediction data set, and the training target is as follows:

[0071]

[0072] In the formula, is a loss; y is an output; θ rl is a parameter for defining a policy selection action.

[0073] The technical effect of the present application is self-evident, and the beneficial effects of the present application are as follows:

[0074] The present application proposes a per-hour airport security passenger flow prediction algorithm based on NBMT-OneNet. This algorithm is designed to capture the characteristics of per-hour airport security data, using N-BEATS and MTGNN models to capture the complex correlations between time trends and variables, effectively improving the accuracy of the prediction results.

[0075] The present application evaluates the proposed algorithm on real airport security data and compares the test results with the current SOTA online learning method in terms of different performance indicators, confirming the effectiveness and superiority of the proposed method.

[0076] In short, to meet the real-time response demand, the present application proposes a per-hour security passenger flow prediction model based on NBMT-OneNet to address the slow response speed and insufficient dynamic adaptability of traditional methods. The model decouples the time-dependent model and the cross-variable-dependent model, using N-BEATS and MTGNN models to capture the complex correlations between time trends and variables. Then, the OCP module is used to calculate the long-term and short-term weights, and the final prediction result is obtained by combining them. Experimental results show that the MSE, MAE and R 2 The performance of the present application is superior to other algorithms for predicting airport passenger security data, and can be used in actual airport big data platforms to provide a basis for optimizing real-time security processes. BRIEF DESCRIPTION OF DRAWINGS

[0077] Fig. 1 The overall framework of the present application is shown in Figure 1.

[0078] Fig. 2 (a)-(d) are the prediction result curve graphs of the present application on October 13, 15, 14 and 20, 2023. DETAILED DESCRIPTION

[0079] The present application will be further described below in conjunction with examples, but should not be understood as limiting the above-mentioned subject matter of the present application to the following examples. Various substitutions and modifications can be made according to ordinary technical knowledge and conventional means without departing from the above-mentioned technical idea of the present application, and all such substitutions and modifications should be included within the scope of protection of the present application.

[0080] Example 1:

[0081] Referring to Figs. 1-2 , an airport terminal security passenger number per-hour prediction method, comprising the following steps:

[0082] 1) Obtain airport historical security passenger data and security passenger flow feature data in the same period as the security passenger flow data;

[0083] 2) Preprocess the security passenger flow feature data and the airport historical security passenger data, and combine the preprocessed security passenger flow feature data and the airport historical security passenger data into a passenger security quantity prediction data set;

[0084] 3) Construct a per-hour security passenger quantity prediction model based on NBMT-OneNet;

[0085] 4) Train the per-hour security passenger quantity prediction model using the passenger security quantity prediction data set;

[0086] 5) Obtain security passenger data and security passenger flow feature data of a past T time period of the airport to be predicted, and input them into the per-hour security passenger quantity prediction model to obtain security passenger data of a future T' time period.

[0087] In step 1), the security passenger flow feature data includes airport operation features, time features, and weather features;

[0088] The airport operation features include per-hour historical check-in data and boarding data;

[0089] The time features include daily patterns, weekly patterns, and holidays;

[0090] The weather features include per-hour historical weather, temperature, wind speed, and precipitation.

[0091] In step 2), the preprocessing includes data alignment, missing value processing, and LOF anomaly processing.

[0092] In step 3), the per-hour security passenger quantity prediction model includes an N-BEATS model and an OCP module:

[0093] The N-BEATS model is used for time-dependent modeling;

[0094] The MTGNN is used for cross-variable dependent modeling;

[0095] The OCP module fuses the outputs of the N-BEATS and MTGNN models to obtain security passenger quantity prediction results.

[0096] In step 3.1), the N-BEATS model includes multiple stack modules; each stack module includes multiple blocks connected in series;

[0097] The input of the N-BEATS model is security passenger flow data, and the output is the sum of the outputs of all stack modules;

[0098] In each stack module, the input x l of the first block is the original input sequence, and the output includes the predicted value for the future window H and the reconstructed value for the block input The input of the remaining blocks is the residual of the previous block input minus the reconstructed value of the previous block output.

[0099] Each block includes a fully connected network and a mapping layer.

[0100] The fully connected network includes multiple fully connected layers and linear layers for mapping the input to forward unfolding coefficients and backward unfolding coefficients The mapping layer converts the forward unfolding coefficients and backward unfolding coefficients to prediction results through a mapping function.

[0101] The output of the lth block is as follows:

[0102] h l,1 = FC l,1 (x l )(1)

[0103] h l,2 = FC l,2 (h l,1 )(2)

[0104] h l,3 = FC l,3 (h l,2 )(3)

[0105] h l,4 = FC l,4 (h l,3 )(4)

[0106]

[0107] In the formula, LINEAR represents a linear projection layer; FC represents a nonlinear standard fully connected layer with RELU; represents a prediction basis vector, is a backtracking basis vector, is the ith element of h l,1 , h l,2 , h l,3 , h l,4 represent the outputs of different fully connected layers.

[0108] Each stack module merges the output of each block by a double residual structure after obtaining the output of each block to obtain:

[0109]

[0110] wherein, is the merged output.

[0111] The input of the MTGNN model is the security passenger flow data X = {z1, z2, z3,..., zP of P historical time steps, and the output is the security passenger number Y at the Qth future time step. P};

[0112] The MTGNN model comprises a graph learning layer, m graph convolution modules, m time convolution modules, and an output module.

[0113] The graph learning layer mines the potential relationship between nodes, calculates the adjacency matrix of the graph, and passes the adjacency matrix as input to all graph convolution modules.

[0114] The graph convolution modules and the time convolution modules are arranged alternately, and respectively capture the dimensional and time dependencies.

[0115] Residual connections are added between the outputs of the graph convolution modules and the inputs of the time convolution modules, and a skip connection is added after each time convolution module.

[0116] The output module maps the hidden features to the target dimension to generate the final prediction result.

[0117] The graph learning layer of the MTGNN model captures the hidden relationship in the security passenger flow sequence by adaptively learning the adjacency matrix, i.e.

[0118] M1=tanh(αE1Θ1)(11)

[0119] M2=tanh(αE2Θ2)(12)

[0120]

[0121] for i=1,2,3,..,N(14)

[0122] idx=argtopk(A[i,:])(15)

[0123] A[i,-idx]=0(16)

[0124] In the formula, E1 represents a source node feature matrix; E2 represents a target node feature matrix; Θ1 and Θ2 are two independent learnable weight matrices; α is a hyperparameter for controlling the saturation rate of the activation function (generally taking a value of 0.1-1, and the value is determined through hyperparameter search (grid search, Bayesian optimization) or trial and error method, and the goal is to balance the convergence speed and expression ability of the model); argtopk(·) is used to return the index of the top k maximum values of a vector; A is an adjacency matrix; M1 and M2 are feature matrices capturing the potential relationship of nodes; i is the node index in the adjacency matrix; idx is the top k key neighbor index of node i;

[0125] The graph convolution module is used for fusing node and neighbor information to process spatial dependency in the graph.

[0126] The graph convolution module includes two mix-hop propagation layers.

[0127] The output of the graph convolution module is as follows:

[0128]

[0129] H out = Δ(H (2) H (1) ) = H 2 -H 1 (19)

[0130] In the formula, β is a hyperparameter for controlling the ratio of keeping the original state of the root node; k is the propagation depth; H in represents the input hidden state output by the previous layer; H (0) = H in . is a normalized adjacency matrix; A is an original adjacency matrix, and I is a unit matrix, is a normalized degree matrix; H out represents the output hidden state of the current layer; the parameter matrix W (k) serves as a feature selector; H 2 , H 1 represent the hidden state of the propagation depth of 2 and the propagation depth of 1; H (k) is the hidden state of the propagation depth of k; Δ is a difference symbol;

[0131] The time convolution module extracts high-order time sequence features through a standard dilated one-dimensional convolution filter.

[0132] The time convolution module includes a filtering layer, a gating layer, a dilated initial layer, a skip connection layer, and an output layer.

[0133] Among them, the filter layer uses the hyperbolic tangent activation function to capture the timing mode, and the gating layer uses the S-shaped activation function to dynamically adjust the amount of information passed by the filter to the next module; the expansion initial layer is filtered by the filter;

[0134] The skip connection layer is used to normalize the information, so that the sequence length input to the output layer is the same. The output layer includes two 1x1 convolution layers, which output values of a specified dimension.

[0135] The output z' of the expansion initial layer of the time convolution module is as follows:

[0136] z' = concat(z*f 1×2 ,z*f 1×3 ,z*f 1×6 ,z*f 1×7 )(20)

[0137]

[0138] In the formula, d is the expansion factor. t is the time point of the current calculation; s is the position index of the filter; z is a given one-dimensional input sequence; f T , f 1×2 , f 1×3 , f 1×6 , f 1×7 are filters of different sizes.

[0139] The OCP module maintains the long-term weight w of the hourly security check passenger flow prediction in the day through the exponential gradient descent algorithm, sets different short-term weights b through the RvS framework, and then fuses the long-term weight w and the short-term weight b into the model optimal weight and fuses the weight and the prediction result through ; wherein represents the prediction result of the time-dependent model, represents the prediction result of the cross-variable-dependent model, represents the optimal weight of the time-dependent model at t time, represents the prediction result of the cross-variable-dependent model at t time.

[0140] The exponential gradient descent algorithm is as follows:

[0141] Delta = {w t | w t,i >=0 and ||w t ||1=1}(22)

[0142] In the formula, t is a time step indicator; delta is a gradient; and w t,i is a weight.

[0143] The long-term weight w, the short-term weight b and the model weight in the OCP module are as follows:

[0144]

[0145] In the formula, l t,i represents the loss of f i at time step t. is a fusion factor; f i (x) represents the prediction output of the current round prediction model for the input sample;

[0146]

[0147] The passenger security quantity prediction dataset is used to train the hourly security passenger quantity prediction model, and the training target is as follows:

[0148]

[0149] In the formula, l t,i , respectively represent the loss in the long-term weight training process and the loss in the short-term weight training process. y is the output; and θ rl is a parameter for defining policy selection action.

[0150] Embodiment 2:

[0151] An airport terminal security passenger quantity hourly prediction method, comprising the following steps:

[0152] 1) Obtain terminal historical security passenger data and security passenger flow feature data in the same period as the security passenger flow data;

[0153] 2) Preprocess the security passenger flow feature data and the terminal historical security passenger data, and combine the preprocessed security passenger flow feature data and the terminal historical security passenger data into a passenger security quantity prediction dataset;

[0154] 3) Construct an hourly security passenger quantity prediction model based on NBMT-OneNet;

[0155] 4) Use the passenger security quantity prediction dataset to train the hourly security passenger quantity prediction model;

[0156] 5) Obtain security passenger data and security passenger flow feature data of a past T period of the terminal to be predicted, and input them into the hourly security passenger quantity prediction model to obtain security passenger data of a future T' period.

[0157] Embodiment 3:

[0158] An airport terminal security passenger quantity hourly prediction method, the technical content is same with embodiment 2, further, in step 1), the security passenger flow characteristic data includes airport operation characteristics, time characteristics, weather characteristics;

[0159] The airport operation characteristics include hourly historical check-in data and boarding data;

[0160] The time characteristics include daily mode, weekly mode, holiday;

[0161] The weather characteristics include hourly historical weather, temperature, wind speed, precipitation.

[0162] Embodiment 4:

[0163] An airport terminal security passenger quantity hourly prediction method, the technical content is same with any one of embodiments 2-3, further, in step 2), the preprocessing includes data alignment, missing value processing, LOF anomaly processing.

[0164] Embodiment 5:

[0165] An airport terminal security passenger quantity hourly prediction method, the technical content is same with any one of embodiments 2-4, further, in step 3), the hourly security passenger quantity prediction model includes N-BEATS model, OCP module:

[0166] The N-BEATS model is used for time-dependent modeling;

[0167] The MTGNN is used for cross-variable-dependent modeling;

[0168] The OCP module fuses the outputs of the N-BEATS and MTGNN models to obtain security passenger quantity prediction results.

[0169] Embodiment 6:

[0170] An airport terminal security passenger quantity hourly prediction method, the technical content is same with any one of embodiments 2-5, further, in step 3.1), the N-BEATS model includes multiple stack modules;Each stack module includes multiple blocks connected in series;

[0171] The input of the N-BEATS model is security passenger flow data, and the output is the sum of the outputs of all stack modules;

[0172] In each stack module, the input x l of the first block is the original input sequence, and the output includes the predicted value for the future window H and the reconstructed value for the block input The input of the rest of the blocks is the residual of the previous block input minus the reconstruction of the previous block output;

[0173] Each block includes a fully connected network and a mapping layer;

[0174] The fully connected network includes a plurality of fully connected layers and a linear layer, for mapping the input to forward unfolding coefficients and backward unfolding coefficients The mapping layer converts the forward unfolding coefficients and backward unfolding coefficients to a prediction result through a mapping function;

[0175] The output of the lth block is as follows:

[0176] h l,1 = FC l,1 (x l )(1)

[0177] h l,2 = FC l,2 (h l,1 )(2)

[0178] h l,3 = FC l,3 (h l,2 )(3)

[0179] h l,4 = FC l,4 (h l,3 )(4)

[0180]

[0181] In the formula, LINEAR represents a linear projection layer; FC represents a nonlinear standard fully connected layer with RELU; represents a prediction basis vector, is a backtracking basis vector, is the ith element of h l,1 , h l,2 , h l,3 , h l,4 represent the outputs of different fully connected layers;

[0182] Each stack module combines the output of each block after obtaining the output of each block through a double-layer residual structure, to obtain:

[0183]

[0184] In the formula, is the combined output.

[0185] Embodiment 7:

[0186] An airport terminal security passenger quantity hourly prediction method, the technical content is the same as any one of embodiments 2-6, further, the input of the MTGNN model is the historical P time step security passenger flow data X={z1, z2, z3,..., z P}, and the output is the security passenger quantity Y of the future Q time step;

[0187] The MTGNN model comprises a graph learning layer, m graph convolution modules, m time convolution modules, and an output module;

[0188] The graph learning layer mines the potential relationship between nodes, calculates the adjacency matrix of the graph, and passes the adjacency matrix as input to all graph convolution modules;

[0189] The graph convolution modules and the time convolution modules are alternately arranged, and respectively capture the dimension and time dependence;

[0190] Residual connections are added between the outputs of the graph convolution modules and the inputs of the time convolution modules, and a skip connection is added after each time convolution module;

[0191] The output module maps the hidden features to the target dimension to generate the final prediction result.

[0192] Embodiment 8:

[0193] An airport terminal security passenger quantity hourly prediction method, the technical content is the same as any one of embodiments 2-7, further, the graph learning layer of the MTGNN model captures the hidden relationship in the security passenger flow sequence by adaptively learning the adjacency matrix, that is:

[0194] M1=tanh(αE1Θ1)(11)

[0195] M2=tanh(αE2Θ2)(12)

[0196]

[0197] for i=1,2,3,..,N(14)

[0198] idx=argtopk(A[i,:])(15)

[0199] A[i,-idx]=0(16)

[0200] In the formula, E1 represents a source node feature matrix; E2 represents a target node feature matrix; Θ1 and Θ2 are two independent learnable weight matrices; α is a hyperparameter for controlling the saturation rate of an activation function; argtopk(·) is used to return the index of the top k maximum values of a vector; A is an adjacency matrix; Θ1 is used to transform the source node feature matrix E1; and Θ2 is used to transform the target node feature matrix E2.

[0201] The graph convolution module is used for fusing node and neighbor information to process spatial dependency in the graph.

[0202] The graph convolution module includes two mix-hop propagation layers.

[0203] The output of the graph convolution module is as follows:

[0204]

[0205] H out = Δ (H (2) H (1) ) = H 2 -H 1 (19)

[0206] In the formula, β is a hyperparameter for controlling the ratio of keeping the original state of the root node; k is the propagation depth; H in represents an input hidden state output by a previous layer; H (0) = H in . is a normalized adjacency matrix; A is an original adjacency matrix, I is a unit matrix, and D is a degree matrix; H out represents an output hidden state of a current layer; the parameter matrix W (k) serves as a feature selector.

[0207] The time convolution module extracts high-order time sequence features through a standard dilated one-dimensional convolution filter.

[0208] The time convolution module includes a filtering layer, a gating layer, a dilated initial layer, a skip connection layer, and an output layer.

[0209] The filtering layer uses a hyperbolic tangent activation function to capture time sequence patterns, and the gating layer uses an S-shaped activation function to dynamically adjust the amount of information passed from the filter to the next module; the dilated initial layer is filtered by a filter.

[0210] The skip connection layer is used to normalize information, so that the sequence length input to the output layer is the same. The output layer includes two 1×1 convolution layers, and outputs values of a specified dimension.

[0211] The output of the dilated initial layer of the time convolution module is as follows:

[0212] z=concat(z*f 1×2 ,z*f 1×3 ,z*f 1×6 ,z*f 1×7 )(20)

[0213]

[0214] Where d is the dilation factor, t is the current calculation time point, and s is the filter position index.

[0215] Example 9:

[0216] A method for hourly forecasting the number of passengers undergoing security checks at an airport terminal, with the same technical content as any one of Examples 2-8. Furthermore, the OCP module maintains the long-term weight w of the hourly security check passenger flow forecast through an exponential gradient descent algorithm, sets different short-term weights b through an RvS framework, and then fuses the long-term weight w and the short-term weight b into the final model weight. and through The weights are integrated with the prediction results; represents the prediction result of the time-dependent model, represents the prediction result of the cross-variable dependency model, represents the final weight of the time-dependent model at time t, Represents the prediction results of the cross-variable dependence model at time t;

[0217] The exponential gradient descent algorithm is as follows:

[0218] Δ={w t ∣w t,i ≥0and‖w t ‖1=1}(22)

[0219] Where t is the time step indicator.

[0220] Example 10:

[0221] A method for hourly forecasting the number of passengers undergoing security screening at an airport terminal, the technical content of which is the same as any one of Examples 2-9. Furthermore, in the OCP module, the long-term weight w, the short-term weight b, and the model weight are expressed as follows:

[0222]

[0223]

[0224] Where, l t,i represents f at time step t i losses; is the fusion factor; f i(x) represents the predicted output of the current round prediction model for the input sample;

[0225]

[0226] Embodiment 11: A method for predicting the number of passengers passing through security checks at an airport terminal every hour, according to any one of embodiments 2-10, further, using the passenger security check quantity prediction dataset to train the hourly security check passenger quantity prediction model, the training target is as follows:

[0227]

[0228] Embodiment 12:

[0229] A verification method for predicting the number of passengers passing through security checks at an airport terminal every hour, the content is as follows:

[0230] The present application takes the number of passengers passing through security checks at T3A terminal of Chongqing Jiangbei International Airport from February 23, 2023 to December 12, 2023 as an example for empirical analysis. Based on the NBMT-OneNet algorithm, the hourly security passenger flow prediction is carried out, which provides technical support for the real-time dynamic adjustment of airport terminal security services, reduces the operating cost while ensuring the efficiency of passenger traffic. Specifically, the steps include:

[0231] S100: Prediction target definition. Obtain the historical passenger security data of T3A terminal of Chongqing Jiangbei International Airport from February 23, 2023 to December 12, 2023. According to observation, it is found that the number of passengers passing through security checks in the time period of 0:00-3:00 and 22:00-23:00 is very small, and a small number of security personnel can meet the actual service demand, therefore, the prediction target is defined as: predicting the security passenger flow from 3:00 to 21:00 for a total of 19 hours;

[0232] S200: Feature data acquisition. Extract the time features, airport operation features and weather features in the same period as the security passenger flow data as the feature set, which helps the model to identify the short-term dependence and volatility of passenger security data and enhance the prediction ability of the model;

[0233] S300: Data preprocessing. After data alignment, missing value processing and LOF anomaly processing, the security passenger data and the feature set are combined into a passenger security quantity prediction dataset;

[0234] S400: Model construction. The application constructs a per-hour security passenger quantity prediction model based on NBMT-OneNet. The model is based on the Online ensembling Network (OneNet) foundation, and through the Neural Basis Expansion Analysis for Time Series (N-BEATS) and the Multivariate Time Series Forecasting with Graph Neural Networks (MTGNN) multivariate time series prediction model, independent modeling is carried out for time dependence and cross-variable dependence respectively. Subsequently, through the Online Convex Programming (OCP) module, the prediction results of the two models are fused to realize the per-hour accurate prediction of the security passenger flow;

[0235] S500: Per-hour security passenger quantity prediction using the prediction model. The passenger security data set is input into the prediction model constructed in S400, of which 150 days are used as the training set for training, 70 days are used as the verification set, and 72 days are used as the test set for testing.

[0236] S201: Collect per-hour historical check-in data and boarding data of T3A terminal of Chongqing Jiangbei International Airport as airport operation characteristics;

[0237] S202: According to the characteristics of passenger security data, daily mode (i.e. which period of the day), weekly mode (i.e. whether it is in the middle of the week or the weekend), and holiday (i.e. whether it is a holiday) are introduced as time characteristics.

[0238] S203: Use the API provided by Open-Meteo to obtain per-hour historical weather, temperature, wind speed, and precipitation under the location of the airport as weather characteristics.

[0239] S401: Model the time dependence using N-BEATS. The N-BEATS model is composed of multiple stack modules, and each stack is composed of multiple blocks in series. The input x l of the first block is the original input sequence, and the output is divided into two parts, one is the predicted value for the future window H , and the other is the reconstruction value for the block input The input of the rest of the blocks is the residual of the previous block input minus the reconstruction of the previous block output. The length of the input window is set to be a multiple of the prediction range H, usually between 2H and 14H. In this way, each block not only serves as a supplement to the previous block to continuously fit the residual information that has not been captured before, but also can be regarded as a decomposition of the time series information. Different blocks in different stacks will fit different parts of the time series, and finally the output of N-BEATS is the sum of the outputs of each stack.

[0240] The internal of each block is composed of two parts, the first part is a fully connected network (FC) for mapping the input to the expansion coefficients, i.e. the forward expansion coefficient and the backward expansion coefficient contains four fully connected layers and a simple linear layer. Therefore, the specific formula of the lth block is as follows:

[0241] h l,1 = FC l,1 (x l ), h l,2 = FC l,2 (h l,1 ), h l,3 = FC l,3 (h l,2 ), h l,4 = FC l,4 (h l,3 ),

[0242]

[0243] LINEAR is a simple linear projection layer. FC is a nonlinear standard fully connected layer with RELU. The and can be understood as low-dimensional vectors closely related to the prediction target after removing redundant information in the input.

[0244] The second part is to obtain the prediction result by mapping the and obtained in the previous stage through a mapping function, and the specific formula is as follows:

[0245]

[0246] where is the prediction basis vector, is the backtracking basis vector, is the i-th element of .

[0247] After obtaining the prediction result of the output of each block, it is input into a double-layer residual structure to combine and output, and the specific formula is as follows:

[0248]

[0249] In this way, the complex network structure and the convolution layer are shielded, and a fully connected network (FC) is used to model the time series data. Such a simplified architecture can make the model training more efficient and reduce the risk of overfitting.

[0250] S402: Model the cross-variable dependence using MTGNN. Given the security passenger flow data X = {z1, z2, z3,..., zP} of P time steps in the past, the target is to predict the value Y corresponding to the security passenger number at the future Qth step. In the MTGNN model, for an N-dimensional security passenger flow data set, let z P t ∈R N denote the value of the sequence at time step t, where z t [i] denotes the value of the ith variable at time step t, and N denotes the feature dimension and the number of nodes. MTGNN includes a graph learning layer, m graph convolution modules, m time convolution modules, and an output module. The graph learning layer is used to mine the potential relationship between nodes by calculating the adjacency matrix of the graph and passing it as input to all graph convolution modules. The graph convolution modules and the time convolution modules are arranged alternately, capturing the dimension and time dependence, respectively. In order to alleviate the problem of gradient disappearance, a residual connection is added between the output of the graph convolution module and the input of the time convolution module. At the same time, a skip connection is added after each time convolution module. Finally, the output module maps the hidden features to the target dimension to generate the final prediction result.

[0251] For the MTGNN algorithm as a whole, the graph formula is set as G(V, E), where V is the set of nodes, E is the set of edges, and N is used to represent the number of nodes; v ∈ V represents a node, e = (v, u) ∈ E represents an edge from u pointing to v, and the neighbor nodes of node v ∈ V are defined as N(v) = {u ∈ V | (v, u) ∈ E}; the adjacency matrix is a mathematical representation of the graph, denoted as A ∈ R N×N , if (v i , v j ) ∈ E, then A ij = c is greater than 0, and if , then A ij = 0.

[0252] ​The graph learning layer captures the hidden relationship in the security passenger flow sequence by adaptive learning of the adjacency matrix. Most studies construct the graph structure by distance metrics (such as dot product, Euclidean distance), resulting in high time and space complexity. To solve this problem, MTGNN only calculates the relationship of part of node pairs, reducing the calculation and memory overhead of small batch training. At the same time, in security passenger flow prediction, the invention hopes to identify the causal relationship between features, that is, the change of one feature state triggers the change of another feature state, so the learned relationship should have unidirectionality. The graph learning layer in MTGNN is used to extract unidirectional relationships, and the specific formula is as follows:

[0253] M1=tanh(αE1Θ1)

[0254] M2=tanh(αE2Θ2)

[0255]

[0256] for i=1,2,3,..,N

[0257] idx=argtopk(A[i,:])

[0258] A[i,-idx]=0

[0259] Where E1 and E2 represent randomly initialized node embeddings, Θ2 and Θ2 are model parameters, and alpha is a hyperparameter used to control the saturation rate of the activation function, and argtopk(·) returns the index of the top k maximum values of a vector. The asymmetry of the graph adjacency matrix proposed by MTGNN is realized by equation (4.6). The subtraction term and the ReLU activation function regularize the adjacency matrix, so that if A vu is positive, its diagonal corresponding item A uv will be zero. Equation (4.8) and equation (4.9) are strategies to make the adjacency matrix sparse while reducing the computational cost of subsequent graph convolution. For each node, select its top k nearest nodes as its neighbors. While preserving the weights of connected nodes, the weights of unconnected nodes are set to zero. In this way, when the model is trained in an online learning manner, the graph adjacency matrix can also adapt to new training data to update the model parameters.

[0260] The graph convolution module is mainly used to fuse node and neighbor information to process the spatial dependency in the graph. This module consists of two mix-hop propagation layers, which process the incoming and outgoing information of each node, respectively. The final incoming information is obtained by summing the outputs of the two mix-hop propagation layers. In the mix-hop propagation layer, there are two steps of information propagation and information selection. First, propagate the information horizontally, and then select the information vertically. Information propagation mainly balances the historical state and neighbor node information, and its steps are defined as follows:

[0261]

[0262] where β is a hyper-parameter controlling the ratio of keeping the root node original state. k is the propagation depth. H in denotes the input hidden state output by the previous layer, H (0) in . is the normalized adjacency matrix (A is the original adjacency matrix, I is the identity matrix, D is the degree matrix). Information selection is based on hidden states of different propagation depths, and multi-order neighborhood information is fused by weighted integration. The steps are defined as follows:

[0263]

[0264] where H out denotes the output hidden state of the current layer, and the parameter matrix W (k) is used as a feature selector. At the same time, the local and neighborhood information is dynamically balanced through equation (4.12), which not only retains the differential representation ability, but also avoids the risk of over-smoothing, and realizes more efficient cross-order feature fusion.

[0265] H out = Δ(H (2) H (1) ) = H 2 - H 1

[0266] The time convolution module extracts high-order time sequence features through a standard dilated one-dimensional convolution filter. This module contains two dilated inception layers (DilatedIncepation Layer), in which the filter layer uses the hyperbolic tangent activation function (Tanh) to capture the time sequence pattern, and the gating layer uses the sigmoid activation function (Sigmoid) to dynamically adjust the amount of information passed from the filter to the next module. In which a time inception layer composed of four filter sizes is used. For a given one-dimensional input sequence z∈R T and different sizes of convolution kernels, the calculation method is as follows:

[0267] z = concat(z * f 1×2 , z * f 1×3 , z * f 1×6 , z * f 1×7 )

[0268] The outputs of the four filters are truncated to the same length according to the largest filter, and concatenated in the channel dimension, represented by z★f 1×k is defined as:

[0269]

[0270] ​where d is the dilation factor. t is the time point of the current calculation. s is the position index of the filter. Finally, the results of multiple modules are spliced through a skip connection layer and an output layer. The skip connection layer is used to normalize the information so that the sequence length input to the output layer is the same. The output layer includes two 1x1 convolutional layers, and outputs a specified dimension value. Since the application mainly uses the MTGNN model for single-step prediction, the dimension is 1.

[0271] S403: Fuse the prediction results of the N-BEATS and MTGNN models through the OCP module. In the OCP module, the application uses the exponential gradient descent algorithm (EGD) to maintain the long-term weight w of the hourly security passenger flow prediction in the day. At the same time, a set of different short-term weights b are set by the RvS framework on the basis of the long-term weight, which is used to capture the recent changes of the data. By combining w and b, the final dynamic prediction is realized.

[0272] EGD is a commonly used method for maintaining long-term weights, and its core idea is to define the decision space as a d-dimensional simplex Δ, as follows:

[0273] Δ={w t ∣w t,i ≥0and‖w t ‖1=1}

[0274] where t is a time step indicator, which is omitted in the following for simplicity. In the process of prediction result fusion, for a given hourly security data stream and its corresponding prediction target The application uses d fusion factors f with different parameters, and The goal of the fusion factor is to minimize the prediction error of the hourly security passenger flow, as shown below:

[0275]

[0276] where f i (x) represents the prediction output of the input sample by the current round of prediction model. According to the EGD algorithm, the center point of the simplex is selected as , and l t,i is expressed as the loss of f i at time step t, then the update rule of each w i is as follows:

[0277]

[0278]

[0279] Adapt to the dynamic changes of security passenger flow. Therefore, in order to effectively combine the long-term trend information of security passenger flow and the recent dynamic changes, the application learns a new short-term weight b at time step t according to the long-term weight w and the performance of each model in the historical short-term interval I = [l, t] through the RvS framework. In the application, l is set to t-1, and then the agent selects an action according to the policy defined by the parameters θ rl The input of the policy is the weight of each model in the historical interval and the actual statistical result of the security passenger flow. In training, the product of each predicted value and the fusion factor corresponding to it is spliced with the true result y. The policy network is designed as a two-layer multilayer perceptron (MLP) structure Then the short-term weight and the final overall weight will be as follows:

[0280]

[0281] Unlike the ordinary RvS framework, the hourly security passenger flow prediction cannot train the network through a simple classification task. The main reason is that the framework cannot obtain the real security passenger flow data at the next moment as a supervision signal. Therefore, the application trains the network by minimizing the training error caused by the new weight, ensuring that the framework can more accurately approximate the passenger flow change trend after each update. The mathematical expression of this optimization process, i.e. the minimization objective function, is as follows:

[0282]

[0283] In the inference process, as the concept drift changes gradually, the application generates a prediction using w t-1 +b t-1 , and trains the network after observing the real security passenger flow data.

[0284] Embodiment 13:

[0285] A verification of an airport terminal security passenger quantity hourly prediction method, the content is as follows:

[0286] The application takes the passenger security data of Chongqing Jiangbei Airport T3A terminal for 292 days as a case for empirical analysis, and the data is from the big data platform of Chongqing Jiangbei International Airport, and a total of 7008 records are included. To ensure the effectiveness of the data and the accuracy of the experimental results, the application processes the passenger security data and uses 150 days as the training set, 70 days as the validation set, and 72 days as the test set.

[0287] ​We use the mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), mean absolute percentage error (MAPE) and coefficient of determination (R 2 ) is used as the evaluation index, and its calculation formula is as follows:

[0288]

[0289] Where n represents the total sample size, y i Indicates the actual value of passenger security data, represents the predicted value of passenger security inspection data, Indicates the average value of passenger security screening data.

[0290] Table 1 Different model prediction evaluation indicators (based on MSE, MAE, RMSE, MAPE, R 2 )

[0291]

[0292] To verify the effectiveness of the proposed NBMT-OneNet algorithm for hourly forecasting of security check passenger flow during the day, the NBMT-OneNet algorithm is compared with FSNet, OnlineTCN, DER++, and the original OneNet. The results are shown in Table 2.

[0293] Table 2 Comparison of NBMT-OneNet algorithm effects

[0294]

[0295] Experimental results show that compared with the suboptimal algorithm DER++, the NBMT-OneNet algorithm has a slight decrease in MAPE, but the MSE index is improved by 17%, the MAE index is improved by 10.16%, and the RMSE index is improved by 8.91%.

[0296] In order to more comprehensively demonstrate the effectiveness of the NBMT-OneNet algorithm, the prediction results under two different scenarios, normal operation day and abnormal operation day, are compared and displayed. Fig. 2 As shown. Among them, Fig. 2 Figures (a) and 2(b) show the prediction performance of the NBMT-OneNet model on normal operating days, while Figures 2(c) and 2(d) show the prediction performance of NBMT-OneNet on abnormal operating days.

Claims

1. A method for hourly forecasting of the number of passengers undergoing security checks at an airport terminal, characterized in that: The following steps are involved: 1) Obtain historical terminal security passenger data and security passenger flow characteristic data within the same period as the security passenger flow data; 2) Preprocessing the security inspection passenger flow characteristic data and the terminal's historical security inspection passenger data, and merging the preprocessed security inspection passenger flow characteristic data and the terminal's historical security inspection passenger data into a passenger security inspection quantity prediction dataset; 3) Construct an hourly security check passenger number prediction model based on NBMT-OneNet; 4) Using the passenger security check number prediction dataset to train the hourly security check passenger number prediction model; 5) Obtain the security inspection passenger data and security inspection passenger flow characteristic data of the terminal to be predicted in the past T time period, and input them into the hourly security inspection passenger number prediction model to obtain the security inspection passenger data in the future T' time period.

2. The method for hourly forecasting of the number of passengers undergoing security checks at an airport terminal according to claim 1, characterized in that: In step 1), the security inspection passenger flow characteristic data includes airport operation characteristics, time characteristics, and weather characteristics; The airport operation characteristics include hourly historical check-in data and boarding data; The time characteristics include daily mode, weekly mode, and holidays; The weather characteristics include hourly historical weather, temperature, wind speed, and precipitation.

3. The method for hourly forecasting of the number of passengers undergoing security checks at an airport terminal according to claim 1, characterized in that: In step 2), the preprocessing includes data alignment, missing value processing, and LOF exception processing.

4. The method for hourly forecasting of the number of passengers undergoing security checks at an airport terminal according to claim 1, characterized in that: In step 3), the hourly security check passenger number prediction model includes the N-BEATS model and the OCP module: The N-BEATS model is used to model time dependence; The MTGNN is used to model cross-variable dependencies; The OCP module integrates the outputs of the N-BEATS and MTGNN models to obtain the prediction result of the number of security check passengers.

5. The method for hourly forecasting of the number of passengers undergoing security checks at an airport terminal according to claim 4, characterized in that: In step 3.1), the N-BEATS model includes multiple stack modules; each stack module includes multiple blocks connected in series; The input of the N-BEATS model is the security inspection passenger flow data, and the output is the sum of the outputs of all stack modules; In each stack module, the input x of the first block l is the original input sequence, and the output includes the predicted value of the future window H And the reconstruction value of the block input The input of the remaining blocks is the residual of the previous block input minus the reconstructed value of the previous block output; Each block includes a fully connected network and a mapping layer; The fully connected network includes multiple fully connected layers and linear layers for mapping the input into forward expansion coefficients and the backward expansion coefficient The mapping layer uses the mapping function to transform the forward expansion coefficients and the backward expansion coefficient Convert to prediction results; The output of the lth block is as follows: h l,1 =FC l,1 (x l )(1) h l,2 =FC l,2 (h l,1 )(2) h l,3 =FC l,3 (h l,2 )(3) h l,4 =FC l,4 (h l,3 )(4) Where LINEAR represents the linear projection layer; FC represents the nonlinear standard fully connected layer with RELU; represents the prediction basis vector, is the backtracking basis vector, yes The i-th element of h l,1 、h l,2 、h l,3 、h l,4 Represents the output of different fully connected layers; After each stack module obtains the output of each block, it merges the output of each block through a double-layer residual structure to obtain: Where, is the merged output.

6. The method for hourly forecasting of the number of passengers undergoing security checks at an airport terminal according to claim 4, characterized in that: The input of the MTGNN model is the security inspection passenger flow data X={z1,z2,z3,...,z P }, the output is the number of passengers Y to be screened in the Qth time step in the future; The MTGNN model includes a graph learning layer, m graph convolution modules, m time convolution modules and an output module; The graph learning layer mines the potential relationships between nodes, calculates the adjacency matrix of the graph, and passes the adjacency matrix as input to all graph convolution modules; The graph convolution modules and the temporal convolution modules are arranged alternately to capture dimensionality and temporal dependencies respectively; A residual connection is added between the output of the graph convolution module and the input of the temporal convolution module, and a skip connection is added after each temporal convolution module; The output module maps the hidden features to the target dimension and generates the final prediction results.

7. The method for hourly forecasting of the number of passengers undergoing security inspection at an airport terminal according to claim 6, characterized in that: The graph learning layer of the MTGNN model captures the hidden relationships in the security check passenger flow sequence by adaptively learning the adjacency matrix, namely: M1=tanh(αE1Θ1)(11) M2=tanh(αE2Θ2)(12) for i=1,2,3,..,N(14) idx=argtopk(A[i,:])(15) A[i,-idx]=0(16) Where E1 represents the original node feature matrix; E2 represents the target node feature matrix; Θ1 and Θ2 are two independent learnable weight matrices; α is a hyperparameter used to control the saturation rate of the activation function; argtopk(·) is used to return the indices of the top k maximum values ​​of a vector; A is the adjacency matrix; M1 and M2 are feature matrices that capture the potential relationships between nodes; i is the node index in the adjacency matrix; idx is the index of the top k key neighbors of node i; The graph convolution module is used to fuse node and neighbor information to process spatial dependencies in the graph; The graph convolution module includes two mix-hop propagation layers; The output of the graph convolution module is as follows: H out =H 2 -H 1 (19) Where β is a hyperparameter used to control the ratio of keeping the original state of the root node; k is the propagation depth; H in Represents the input hidden state output by the previous layer; H (0) =H in ; is the normalized adjacency matrix; A is the original adjacency matrix, I is the identity matrix, is the normalized degree matrix; H out Represents the output hidden state of the current layer; parameter matrix W (k) As a feature selector; H 2 、H 1 represents the hidden state with a propagation depth of 2 and a propagation depth of 1; H (k) is the hidden state with a propagation depth of k; The temporal convolution module extracts high-order temporal features through a standard dilated one-dimensional convolution filter; The temporal convolution module includes a filter layer, a gating layer, an expansion initial layer, a skip connection layer and an output layer; The filtering layer uses the hyperbolic tangent activation function to capture the timing pattern, and the gating layer uses the S-type activation function to dynamically adjust the amount of information passed by the filter to the next module; The expanded initial layer is filtered through the filter; The jump connection layer is used to normalize information so that the lengths of sequences input to the output layer are the same; the output layer includes two 1×1 convolutional layers that output values ​​of a specified dimension. The expanded initial layer output z' of the temporal convolution module is as follows: z'=concat(z*f 1×2 ,z*f 1×3 ,z*f 1×6 ,z*f 1×7 ) (20) Where d is the expansion factor. t is the time point of the current calculation; s is the position index of the filter; z∈R T is a given one-dimensional input sequence; f 1×2 、f 1×3 、f 1×6 、f 1×7 are filters of different sizes.

8. The method for hourly forecasting of the number of passengers undergoing security checks at an airport terminal according to claim 4, characterized in that: The OCP module maintains the long-term weight w of the hourly security check passenger flow forecast through the exponential gradient descent algorithm, sets different short-term weights b through the RvS framework, and then merges the long-term weight w and the short-term weight b into the final weight of the model. and through The weights are integrated with the prediction results; represents the prediction result of the time-dependent model, represents the prediction result of the cross-variable dependency model, represents the final weight of the time-dependent model at time t, Represents the prediction results of the cross-variable dependence model at time t; The exponential gradient descent algorithm is as follows: Δ={w t ∣w t,i ≥0and‖w t ‖1=1}(22) Where t is the time step indicator; Δ is the gradient; w t,i is the weight.

9. The method for hourly forecasting of the number of passengers undergoing security checks at an airport terminal according to claim 8, characterized in that: In the OCP module, the long-term weight w, short-term weight b, and model weight are as follows: Where, represents f at time step t i losses; is the fusion factor; f i (x) represents the prediction output of the current round prediction model for the input sample; is the normalization factor, η is the learning rate; y is the true result; is the prediction result; d represents the dimension; f rt It is a multi-layer perceptron structure.

10. The method for hourly forecasting of the number of passengers undergoing security inspection at an airport terminal according to claim 1, characterized in that: The passenger security check number prediction dataset is used to train the hourly security check passenger number prediction model. The training objectives are as follows: Where, is the loss; y is the output; θ rl Parameters for defining the policy selection action.

Images