Deep learning airport flow prediction method and system of embedded physical model

Abstract

The invention discloses a deep learning airport flow prediction method and system for an embedded physical model, and relates to the field of airport flow prediction.The method comprises the steps that airport information of a target area is obtained, and multi-scale feature data is obtained through division according to time periods; slicing the feature data, and executing multi-scale fusion to obtain a feature vector set; taking the set and single-point flow data as input, and extracting time information through expansion causal convolution and gating operation; and after spatial information is extracted by using the graph matrix and the adaptive matrix, residual connection and linear transformation are carried out to obtain airport flow prediction data. According to the method, the physical model, the graph neural network and the convolutional network are combined, so that the accuracy of predicting the airport flight traffic is enhanced.

Description

Technical Field

[0001] The present application relates to the technical field of airport traffic prediction, and in particular to a deep learning airport traffic prediction method and system with an embedded physical model. Background Art

[0002] As social travel demand grows, global airport flight traffic is rising, and civil aviation transportation is expected to continue to grow. However, airport resources cannot meet demand, resulting in congestion and delays, economic losses and safety hazards. Improving operational efficiency and rationally dispatching resources are the first choice, and the key lies in accurately predicting flight traffic. Predictions that are too high will lead to adjustments in air operations, while predictions that are too low will waste resources.

[0003] However, airport flight traffic exhibits high dynamics, nonlinearity and complex spatiotemporal correlations. Existing technologies only focus on the traffic at a single airport and do not take into account the complex spatial connections between airports, which leads to low prediction accuracy. Summary of the invention

[0004] The purpose of this application is to provide a deep learning airport traffic prediction method and system with an embedded physical model, which can more accurately predict airport flight traffic.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a deep learning airport traffic prediction method with an embedded physical model, and the deep learning airport traffic prediction method includes:

[0007] Obtain airport information of the target area within a set time; the airport information includes actual airport take-off and landing data and airport weather.

[0008] The airport information is divided according to fixed time periods to obtain multi-scale feature data; the multi-scale feature data is composed of single-point scale flow data of the airport and flow data on the route scale.

[0009] The multi-scale feature data is sliced, and when the number of slices is reached, the multi-scale feature data fusion process is sequentially performed on each slice to obtain a multi-scale fused feature vector set; the feature vector set is composed of a plurality of hidden variable outputs.

[0010] The multi-scale fused feature vector set and the single-point scale flow data are used as input data, and the time information of the input data is extracted through dilated causal convolution and gating operations.

[0011] According to the time information of the input data, a graph matrix and an adaptive matrix are used for extraction to obtain a time output after extracting spatial information through diffusion graph convolution.

[0012] The temporal output after the spatial information is extracted by the diffusion graph convolution is subjected to residual connection, and a linear transformation operation is performed to obtain airport traffic prediction data.

[0013] Optionally, a slicing operation is performed on the multi-scale feature data. When the number of slicing times is reached, a multi-scale feature data fusion process is sequentially performed on each slice to obtain a multi-scale fused feature vector set, specifically including:

[0014] The multi-scale feature data is sliced, and when the number of slices τ=H / L1, the multi-scale feature data fusion process is sequentially performed on each slice to obtain a multi-scale fused feature vector set.

[0015] Among them, τ is the number of slices collected by the counter, H is the length of the time dimension, and L1 is the set time interval.

[0016] Optionally, the multi-scale feature data fusion process includes linear fusion, feature fusion based on graph convolutional network and gated recurrent unit processing.

[0017] Optionally, a slicing operation is performed on the multi-scale feature data. When the number of slicing times is reached, a multi-scale feature data fusion process is sequentially performed on each slice to obtain a multi-scale fused feature vector set, specifically including:

[0018] The multi-scale feature data is sliced, and when the number of slices is reached, linear fusion, feature fusion based on a graph convolutional network, and gated recurrent unit processing are sequentially performed on each slice.

[0019] Optionally, when performing linear fusion on slices, specifically including:

[0020] Use linear layer pairs The feature channels are aggregated and recorded as Linear agg ; Linear agg is a linear layer named agg; is the τth slice of the single-point scale flow data.

[0021] Through the Linear layer att Linear Input agg Perform linear transformation to generate the time weight of the kth airport node at time t Among them, Linear att is a linear layer named att.

[0022] The time weight After normalization, we get exp(·) represents an exponential function, which is used to perform nonlinear transformation on weights; It is the normalized time weight corresponding to airport node k at time t.

[0023] Based on the formula Sure The time weight at time t.

[0024] According to the formula Linearly fuse the slices.

[0025] in, express The traffic data of the i-th node at time t, is the new feature data after fusion, for The time weight at time t, subscript 2k represents The kth node of .

[0026] Optionally, when performing graph convolutional network-based feature fusion on the slice, specifically including:

[0027] Based on the feature data after linear fusion, the diffusion graph convolution operation and the activation function ReLU are used to pass the graph matrix A τ temp right The feature data is learned and fused to obtain the spatial output; the formula expression of the spatial output is:

[0028]

[0029] in, represents the spatial output under the counter τ, the subscript spatial represents the spatial convolution output; ReLU is the activation function; step represents the number of steps of the diffusion convolution; represents the forward diffusion matrix, f represents forward propagation, τ is the time step, rowsum(·) is the row sum, and the superscript represents the matrix power series; represents the diffusion matrix of back propagation at time step τ, b represents back propagation, and the rest is the same as above; W τ1 Represents the forward weight matrix at the step number step, 1 represents forward; W step2 Represents the backward weight matrix at step number step, 2 represents backward.

[0030] Optionally, when performing gated recurrent unit processing on a slice, specifically including:

[0031] The spatial output is temporally learned using a GRU unit to obtain a reset gate output of the GRU unit and an update gate output of the GRU unit.

[0032] According to the formula and formula Determine the hidden state output of the current GRU unit.

[0033] in, represents the candidate hidden state at time step τ, ^ represents the candidate; tanh represents the activation function tanh; represents the weight matrix used by the algorithm to calculate the candidate hidden state in the counter τ step, W is the weight matrix, and h represents the weight matrix used for the candidate hidden state; h τ-1 represents the hidden variable output obtained when the counter is τ-1; It indicates that the counter τ is used to calculate the bias vector of the candidate hidden state, h represents the candidate hidden state; ⊙ represents the Hadamard product.

[0034] Optionally, the multi-scale fused feature vector set and the flight flow transfer sum data set between the airports are used as input data, and time information of the input data is extracted through dilated causal convolution and gating operations, specifically including:

[0035] According to the formula formula And the formula o=g(θ1·ouput a +b)⊙σ(θ2·ouput b +c), extracting the time information of the input data.

[0036] Among them, ouput a represents the output after convolution with the convolution kernel f1(t), a represents the result of the first set of convolution kernels; ouput b represents the output after convolution with the convolution kernel f2(t), b represents the first set of convolution kernel results; f1(s) represents the weight of the first convolution kernel at position s; f2(s) represents the weight of the second convolution kernel at position s; t represents the time step; d represents the step size of the convolution; K represents the size of the convolution kernel; g represents the activation function tanh; σ represents the activation function sigmoid; b represents the bias for the first set of convolution results; c represents the bias for the second set of convolution results; θ1 represents the weight for the first set of convolution results; θ2 represents the weight for the second set of convolution results; ⊙ represents the Hadamard product; * represents the convolution operation.

[0037] Optionally, according to the time information of the input data, the use of the map matrix A and the adaptive matrix A adp Extraction is performed to obtain the temporal output after extracting spatial information through diffusion graph convolution, which specifically includes:

[0038] According to the formula The temporal output is obtained after extracting spatial information through diffusion graph convolution.

[0039] Among them, s_output represents the temporal output after extracting spatial information through diffusion graph convolution; represents the value of the forward diffusion matrix at order k, f1 represents the forward diffusion matrix, represents the value of the backward diffusion matrix at order k, b1 represents the backward diffusion matrix, represents the value of the adaptive matrix at order k; W k1 represents the weight matrix used for the forward diffusion matrix at order k; W k2 represents the weight matrix used for the back diffusion matrix at order k; W k represents the weight matrix used for the adaptation matrix at order k.

[0040] In the second aspect, the present application provides a deep learning airport traffic prediction system with an embedded physical model, comprising:

[0041] The information acquisition module is used to obtain the airport information of the target area within a set time; the airport information includes the actual take-off and landing data of the airport and the airport weather.

[0042] The division module is used to divide the airport information according to fixed time periods to obtain multi-scale feature data; the multi-scale feature data is composed of the single-point scale flow data of the airport and the flow data on the route scale.

[0043] The slicing module is used to perform a slicing operation on the multi-scale feature data. When the number of slicing times is reached, the multi-scale feature data fusion process is performed on each slice in turn to obtain a multi-scale fused feature vector set; the feature vector set is composed of a number of hidden variable outputs.

[0044] The first extraction module is used to take the multi-scale fused feature vector set and the single-point scale flow data as input data, and extract the time information of the input data through dilated causal convolution and gating operations.

[0045] The second extraction module is used to extract the time information of the input data by using a graph matrix and an adaptive matrix to obtain a time output after extracting the spatial information through diffusion graph convolution.

[0046] The prediction module is used to perform residual connection on the time output after the spatial information is extracted through the diffusion graph convolution, and adopt linear transformation operation to obtain airport traffic prediction data.

[0047] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0048] The present application provides a method and system for predicting airport traffic flow with deep learning and embedded physical models, and the method includes: first, obtaining airport information of the target area within a set time, including actual take-off and landing data and weather conditions of the airport. In order to analyze the data more carefully, the system divides the airport information according to a fixed time period and generates multi-scale feature data, which includes the single-point scale traffic flow of the airport (such as the traffic flow at a specific time) and the traffic flow at the route scale (such as the total traffic flow on a certain route within a period of time); next, the system performs a slice operation on the multi-scale feature data. When the predetermined number of slices is reached, the multi-scale feature data fusion process is performed on each slice, that is, the data of different scales are integrated together to form a multi-scale fused feature vector set; in order to predict future traffic conditions, the time information in the input data is extracted by dilated causal convolution and gate operation; in addition to time information, the spatial information is extracted by using the graph matrix and adaptive matrix through diffusion graph convolution; in order to further improve the accuracy of the prediction, the system performs residual connection on the time output after extracting the spatial information through the diffusion graph convolution, and adopts a linear transformation operation, which can retain more original information, and the linear transformation can further process and optimize the data; finally, the airport traffic prediction data is output. This application combines physical models, graph neural networks, and convolutional networks to learn multi-scale spatiotemporal information, including single-point scale and route scale in spatial scale, thereby enhancing the accuracy of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 A flow chart of a deep learning airport traffic prediction method with an embedded physical model provided in Example 1 of the present application.

[0051] Figure 2 A flow chart of a deep learning airport aircraft traffic prediction algorithm with an embedded physical model provided in Example 1 of the present application.

[0052] Figure 3 This is a RMSE test curve result diagram provided in Example 1 of the present application.

[0053] Figure 4 This is a MAE test curve result diagram provided in Example 1 of the present application.

[0054] Figure 5 A MAPE test curve result diagram provided in Example 1 of the present application.

[0055] Figure 6 A structural diagram of a deep learning airport traffic prediction system with an embedded physical model provided in Example 1 of the present application. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0057] With the increasing demand for social travel, airport flight traffic has been rising year by year around the world, and relevant data reports indicate that the civil aviation transportation industry will continue to grow rapidly for a long time in the future. However, with the surge in airport flight traffic, existing airport resources, such as runway capacity and airspace capacity, can no longer meet flight demand. This mismatch between demand and existing resources has caused a lot of congestion and delays, and has caused huge economic losses to airlines and passengers, and even has a huge hidden danger to airport traffic safety. Compared with the high cost of increasing airport capacity by expanding infrastructure, improving the operational efficiency of airports and air traffic management departments, conducting reasonable flight scheduling and resource allocation, and avoiding excess or shortage situations are the current top choices. One of the key steps is to accurately and reliably predict airport flight traffic and provide information support for air traffic management decisions. However, this depends largely on the accurate estimation of airport flight traffic. If the traffic is estimated too high, it will lead to adjustments in air operations, such as air delays and route changes; while underestimation will cause waste of air and ground resources, such as insufficient resource utilization. Therefore, accurately and reliably predicting airport flight traffic has become an urgent problem to be solved in air traffic management and has attracted more and more attention. Compared with the high cost of increasing airport capacity by expanding infrastructure, improving the operational efficiency of airports and air traffic management departments, conducting reasonable flight scheduling and resource allocation, and avoiding surplus or shortage situations are currently the first choice.

[0058] One of the key steps is to accurately and reliably predict airport flight traffic and provide information support for air traffic management decisions. However, this depends largely on the accurate estimation of airport flight traffic. If the traffic is overestimated, it will lead to adjustments in air operations, such as air delays and route changes; while underestimation will cause waste of air and ground resources, such as insufficient resource utilization. Therefore, accurately and reliably predicting airport flight traffic has become an urgent problem to be solved in air traffic management and has attracted more and more attention. Airport flight traffic shows high dynamics, nonlinearity and complex spatiotemporal correlation. Previous studies have only focused on the traffic of a single airport without considering the complex spatial connections between airports, which has led to unsatisfactory prediction accuracy. With the development of artificial intelligence technology, technologies such as graph neural networks have emerged, which can mine and fuse data features from historical data end-to-end and learn effective data distribution laws. However, most of the traffic prediction tasks based on deep learning networks currently focus on ground transportation and are less used in the aviation field. Therefore, research in this area is particularly important and has broad application prospects.

[0059] As a branch of machine learning, deep learning has shown great potential in traffic flow prediction. In recent years, a large number of studies have adopted the graph convolutional neural network (GCN) structure, which can effectively process non-Euclidean spatial data. By modeling the traffic network as a graph, the nodes in the graph represent traffic nodes, and the edges represent the connection or association between nodes, the graph convolution is used to capture the spatial topological structure in the traffic network. Zhao et al. proposed the temporal graph convolutional network (TGCN), which combines the graph convolutional network (GCN) and the gated recurrent unit (GRU) to capture spatial and temporal dependencies, and has shown significant results on two real-world datasets. Yu et al. proposed the spatio-temporal graph convolutional network (STGCN) framework, which is modeled on the graph and uses graph convolution and temporal convolution operations. It can effectively capture comprehensive spatiotemporal correlations and can be calculated in parallel, improving training efficiency. With the emergence of attention mechanism, more and more works adopt attention mechanism to obtain the importance of the relationship between node features. Guo et al. proposed the attention mechanism space-time graph convolution network (ASTGCN) model, which captures the dynamic spatiotemporal information in traffic data through the spatiotemporal attention mechanism. Cristea et al. proposed the Spatio-Temporal Weighted Attention (ST_WA) framework, which transforms the spatiotemporal agnostic traffic time series prediction model into a spatiotemporal aware model, encoding the time series as random variables and generating location-specific and time-varying model parameters to better capture the spatiotemporal dynamics, and introduces the window attention scheme to improve efficiency.

[0060] The above-mentioned traffic flow prediction tasks based on graph convolution all focus on ground traffic. However, with the development of the civil aviation industry in recent years, related research on air traffic flow prediction has also emerged. Yan et al. proposed an airport traffic flow prediction network, which uses the flight schedule to construct a special semantic graph to represent the airport network, combines the graph convolution operation with the gated recurrent unit, and extracts the transportation mode of airport traffic flow from the two dimensions of time and space. Du et al. introduced the causal transfer matrix between airports to model the interaction between airports, and constructed an adaptive causal graph in a data-driven way to cope with the heterogeneity of airports, and used graph convolutional neural networks for feature learning to achieve good results.

[0061] However, the current growth trend of airspace traffic flow prediction research is far from the trend of industry development, and there is a lack of traffic flow prediction research in the aviation field. Air traffic has its own unique operating laws and influencing factors, such as weather, and the existing prediction models often fail to fully consider these factors, resulting in low prediction accuracy. Therefore, it is necessary to conduct in-depth research on the characteristics of the aviation field and develop a prediction model suitable for air traffic to improve prediction accuracy and practicality. In addition, existing research does not sufficiently explore the multi-scale spatiotemporal characteristics of aviation. The air transportation system has multi-level and multi-scale spatiotemporal characteristics, including long-term and short-term scales in time, single-node scales and route scales in space. However, existing research often focuses on the mining of features at a single scale, and fails to fully consider the correlation and influence between multi-scale features. The mining of multi-scale spatiotemporal features is crucial to improving the accuracy and robustness of prediction models, and can more comprehensively grasp the dynamic changes of air traffic and improve the accuracy and reliability of predictions.

[0062] The purpose of this application is to provide a deep learning airport traffic prediction method and system with an embedded physical model, which can more accurately predict airport flight traffic.

[0063] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0064] Embodiment 1

[0065] like Figure 1 As shown, this embodiment provides a deep learning airport traffic prediction method with an embedded physical model, including:

[0066] Step 101: Acquire airport information of a target area within a set time; the airport information includes actual airport take-off and landing data and airport weather.

[0067] Step 102: Divide the airport information according to fixed time periods to obtain multi-scale feature data; the multi-scale feature data is composed of single-point scale flow data of the airport and flow data on the route scale.

[0068] Step 103: Slicing the multi-scale feature data. When the number of slices is reached, the multi-scale feature data fusion process is performed on each slice in turn to obtain a multi-scale fused feature vector set; the feature vector set is composed of a number of hidden variable outputs.

[0069] Step 104: taking the multi-scale fused feature vector set and the single-point scale flow data as input data, and extracting the time information of the input data through dilated causal convolution and gating operations.

[0070] Step 105: According to the time information of the input data, a graph matrix and an adaptive matrix are used for extraction to obtain a time output after the spatial information is extracted through diffusion graph convolution.

[0071] Step 106: Perform residual connection on the temporal output after the spatial information is extracted by the diffusion graph convolution, and use linear transformation operation to obtain airport traffic prediction data.

[0072] When executing step 101, the specific steps may be as follows:

[0073] Through crawlers, we obtain the actual and expected take-off and landing conditions and corresponding weather conditions of each airport in the target area over a period of time, and summarize and sum the take-off and landing numbers according to fixed time periods to obtain the total number of airport take-offs and landings in each period and the total number of flight flow transfers between airports. Finally, we perform maximum-minimum normalization on the obtained data, that is, find the minimum and maximum values ​​in the data set, perform linear transformation on each data, and convert it to between 0 and 1 by subtracting the minimum value and dividing it by the difference between the maximum and minimum values.

[0074] Specifically, the airport node network can be defined as G t =(V,E(t),A(t)), where V represents the set of airport nodes; E(t) represents the set of directed edges connecting the airport nodes at time t, which will change with time t; A(t) represents the adjacency weight matrix at time t, which will change with time t. In the airport network, the take-off and landing traffic of all airports at time t is recorded as F(t), and the weather data corresponding to each moment is recorded as W(t), where: F(t) = [f1(t), f2(t),…, f N (t)] T represents the take-off and landing traffic vector of all N airports at time t. Each f i (t) represents the take-off and landing traffic of the ith airport at time t. 11 (t),w 12 (t),…,w NM (t)] represents the weather data matrix of all airports at time t, and each w ij (t) represents the jth weather characteristic data of the ith airport at time t, where N is the number of airports and M is the number of characteristics of the weather data.

[0075] The goal of this embodiment is to use the historical P time period traffic data F(t-P+1), F(t-P+2), ..., F(t) and the corresponding weather data W(t-P+1), W(t-P+2), ..., W(t) in the airport network to calculate the average value of the traffic flow in the airport network through a function Predict the take-off and landing traffic data for the next Q time periods F(t+1), F(t+2), ..., F(t+Q). Function The representation is as follows:

[0076]

[0077] Where: t represents the current time t; P represents the number of historical time periods used for prediction; Q represents the number of future time periods that need to be predicted; F(t) represents the take-off and landing traffic vector of all airports at time t; W(t) represents the weather data matrix of all airports at time t.

[0078] When executing step 102, the specific steps may be as follows:

[0079] Based on the airport information, multi-scale spatiotemporal feature data is established as the input of the model, which is divided into two types of data sets, denoted as x in and mat, where x in ∈R C×H×|V| , represents the single-point scale flow data of the airport, C represents the number of characteristic channels (including expected take-off and landing flow data, actual take-off and landing flow data and weather characteristic data), H represents the length of the time dimension, and |V| represents the number of airport nodes; Represents the traffic data on the route scale, H mat Indicates the length of the time dimension corresponding to mat, and H = 2H mat , |V|Same as above.

[0080] When executing step 103, the specific steps may be as follows:

[0081] x in and mat data input for slicing operations, x in Slice according to the time interval l1 in the time dimension, record x in Each slice of Where i represents the i-th slice and subscript 1 represents x in Slice mat in the time dimension according to the time interval l2, where l1 = 2l2, and record each slice of mat as Where i represents the i-th slice, and subscript 2 represents the slice of mat. The total of each slice data is recorded as:

[0082]

[0083] When the counter τ=1,2,..,H / L1, the multi-scale feature data fusion process is performed on each slice in turn. Specifically, Figure 2As shown in step 1 of the module, the multi-scale feature data fusion process sequentially performs linear fusion (a1), feature fusion based on graph convolutional network (b1), and gated recurrent unit processing (c1), as follows:

[0084] (a1) Using linear layers The feature channels of Linear agg is a linear layer, and the subscript agg indicates that the name of the linear layer is agg. Time weight calculation is performed on the time dimension, that is, through the linear layer Linear att (The subscript att represents the name of the linear layer is att) Perform linear transformation to generate the time weight of the kth airport node at time t And normalized to get Here, exp(·) represents an exponential function, which is used to perform nonlinear transformation on the weights; is the normalized time weight corresponding to airport node k at time t. The time length L1 is The time length L2 is twice that of the original value, so the sum of each adjacent two weights in L1 is used as the corresponding moment in L2. The time weight of According to time weight right The subscript 2k represents The kth node of . The traffic data of each node is obtained by summing the weights in the time dimension according to the following formula, as follows:

[0085]

[0086] in express The traffic data of the i-th node at time t, is the new feature data after fusion, and its shape is

[0087] After normalization and fusion, a more accurate adjacency matrix in the short term can be obtained in the time dimension, that is, is the adjacency matrix, the superscript represents the adjacency matrix after the fusion of the τth counter data, and the subscript represents the name of the adjacency matrix, which is temp. At the same time, in order to prevent the situation where the matrix is ​​sparse and the influence of its own nodes is not considered, the adjacency matrix A τ temp 'Add the diagonal elements to the identity matrix, that is, A τ temp =Aτ temp '+I, where I is the identity matrix.

[0088] (b1) Using the diffusion graph convolution operation and the activation function ReLU, through the graph matrix A τ temp right The feature data is learned and integrated to mine spatial information and obtain spatial output

[0089]

[0090] in, represents the spatial output under the counter τ, the subscript spatial represents the spatial convolution output; ReLU is the activation function; step represents the number of steps of the diffusion convolution; represents the forward diffusion matrix, f represents forward propagation, τ is the time step, rowsum(·) is the row sum, and the superscript represents the matrix power series; represents the diffusion matrix of back propagation at time step τ, b represents back propagation, and the rest is the same as above; W τ1 Represents the forward weight matrix at the step number step, 1 represents forward; W step2 Represents the backward weight matrix at step number step, 2 represents backward.

[0091] (c1) In this step, the GRU unit is used to perform temporal learning on the spatial convolution output. The GRU hidden variable output in the algorithm (c1) executed by a counter τ-1 is in Represents h τ-1 The shape of is L2×|V|×2 (if the counter τ is 1, h0 is randomly initialized). The current GRU unit input is The reset gate and update gate are calculated by the following formula to get r τ ,z τ , as follows:

[0092]

[0093]

[0094] Among them, r τ represents the reset gate output of the GRU unit with counter τ; z τ represents the update gate output of the GRU unit with counter τ; σ represents the activation function sigmoid; represents the weight matrix used to calculate the reset gate on the counter τ step, W is the weight matrix, and r represents the weight matrix used to reset the gate; represents the weight matrix used to calculate the update gate on the counter τ step, W is the weight matrix, z represents the weight matrix used to update the gate; h τ-1 represents the GRU latent variable output of executing the algorithm on the counter τ-1 step; represents the bias vector used to calculate the reset gate on the counter τ steps, b represents the bias vector, and r represents the reset gate; represents the bias vector used to calculate the update gate on the counter τ steps, b represents the bias vector, and z represents the bias vector used to update the gate.

[0095] After getting the reset gate and update gate outputs, the candidate hidden state needs to be calculated according to the following formula And how much previous hidden state information h is retained is determined based on the reset gate and update gate outputs τ-1 How much new information will Merge into the current state to get the hidden state output h of the current GRU unit τ , the specific calculation formula is as follows:

[0096]

[0097]

[0098] in, represents the candidate hidden state at time step τ, ^ represents the candidate; tanh represents the activation function tanh; represents the weight matrix used by the algorithm to calculate the candidate hidden state in the counter τ step, W is the weight matrix, and h represents the weight matrix used for the candidate hidden state; h τ-1 represents the hidden variable output obtained by executing algorithm (c1) when the counter is τ-1; represents the bias vector used to calculate the candidate hidden state on the counter τ, h represents the candidate hidden state; ⊙ represents the Hadamard product.

[0099] In addition, after executing (c1), you can also execute (d1), including:

[0100] (d1) If the model is in the training stage, the value of counter τ remains unchanged and the algorithm (d1) is executed. If it is in the inference stage, the counter τ is increased by 1 and (a1) is executed.

[0101] If the model is in the training stage, the embedded physical model output is generated to predict the airport inflow at a time t after L2. The embedded physical model is as follows:

[0102]

[0103] in, is the inflow of airport i at time t, L2 is the historical time window, Δt is a moment in L2, is the route flow attenuation factor between airport i and airport j at time Δt, is the number of flights scheduled to take off from airport j at time Δt and expected to land at airport i.

[0104] Since the hidden variable h τ It aggregates multi-step temporal information, including the mutual influence between airport nodes and their own internal factors.

[0105] Therefore, the latent variables are linearly transformed through the linear layer to obtain the influencing factors of the take-off airport and the landing airport, respectively recorded as The subscripts represent takeoff and landing, respectively, and the superscripts correspond to the airport numbers. The attenuation factor in the above physical model is obtained by multiplying these two influencing factors at each moment Δt in the time dimension and sliced ​​through the flow transfer matrix According to the above physical model, the physical model output under the current counter τ is obtained ( Each corresponding element in the formula is substituted ), denoted as It should be noted here The subscript represents the prediction result of the physical model, and the superscript represents the prediction result at the moment after L2 obtained by executing the algorithm under the counter τ.

[0106] Before executing step 104, the method further includes:

[0107] A fixed graph matrix A is constructed based on the flow transfer of the number of flights between each airport, as follows:

[0108]

[0109] Among them, a ij Represents the element value between node i and node j in the graph matrix A; F ij represents the traffic from airport i to airport j; N(i) represents the set of airports with flights to airport i; m represents an airport in the set N(i).

[0110] At the same time, the adaptive matrix A is introduced adp , by randomly initializing two learnable parameters E1,E2∈R |V|×c Represent the source node embedding and the target node embedding respectively (where c represents the dimension of the node embedding). Then, the two node embedding matrices are multiplied to obtain the spatial dependency weights between the source node and the target node. Then, the ReLU function is used to activate and eliminate weak connections, and the SoftMax function is used to normalize the adaptive adjacency matrix to obtain the final adaptive matrix A. adp , as follows:

[0111]

[0112] When executing steps 104-106, the specific steps may be as follows:

[0113] like Figure 2 As shown in step 2 of the module, when the counter τ = H / L1, the hidden variable output obtained after executing algorithm (c1) is On the feature channel, in For splicing, the input is recorded as input, and it passes through two space-time modules in sequence. Each space-time module executes the following (a2)-(d2) algorithm, and counts each module, recorded as st=1,2 respectively.

[0114] (a2) First, perform temporal convolution. The temporal information of the input is extracted through the following dilated causal convolution and gating operation.

[0115]

[0116]

[0117] o=g(θ1·ouput a +b)⊙σ(θ2·ouput b +c).

[0118] Among them, ouput a represents the output after convolution with the convolution kernel f1(t), a represents the result of the first set of convolution kernels; ouput b represents the output after convolution with the convolution kernel f2(t), b represents the first set of convolution kernel results; f1(s) represents the weight of the first convolution kernel at position s; f2(s) represents the weight of the second convolution kernel at position s; t represents the time step; d represents the step size of the convolution; K represents the size of the convolution kernel; g represents the activation function tanh; σ represents the activation function sigmoid; b represents the bias for the first set of convolution results; c represents the bias for the second set of convolution results; θ1 represents the weight for the first set of convolution results; θ2 represents the weight for the second set of convolution results; ⊙ represents the Hadamard product; * represents the convolution operation.

[0119] (b2) The result o in (a2) is diffused through the graph convolution using the graph matrix A and the adaptive matrix A adp Extract spatial information, specifically:

[0120]

[0121] Among them, s_output represents the temporal output after extracting spatial information through diffusion graph convolution; represents the value of the forward diffusion matrix at order k, f1 represents the forward diffusion matrix, represents the value of the backward diffusion matrix at order k, b1 represents the backward diffusion matrix, represents the value of the adaptive matrix at order k; W k1 represents the weight matrix used for the forward diffusion matrix at order k; W k2 represents the weight matrix used for the back diffusion matrix at order k; W k represents the weight matrix used for the adaptation matrix at order k.

[0122] (c2) Residual connection. Since the time dimension and the number of features will change after the spatiotemporal convolution, the input needs to be aligned and cropped when performing the residual connection to obtain the residual signal res of the algorithm executed under the current counter. The dimension of the residual signal is the same as the dimension of s_output. The process is as follows:

[0123] res = Crop[Align(input)].

[0124] Among them, Crop(·) is the cropping operation, and Align(·) is the alignment operation.

[0125] The output and residual signals are added, i.e. res+s_output, to perform a residual connection.

[0126] (d2) If the counter st=1, then the counter is incremented by 1, and input=res+s_output is set, and the process continues with (a1).

[0127] If the counter st = 2, a linear transformation operation is performed on the output result res + s_output (c2), and its dimension is transformed into the dimension required by the final prediction data, which is recorded as the final output H L .

[0128] If the model usage scenario is a training scenario, step 107 is also included, otherwise the result H L The result required for the final prediction.

[0129] Specifically, step 107 is: if the model usage scenario is a training scenario, it is necessary to prepare label data corresponding to the training, and record the physical model output H P The label is R P , output H L The label is R L , which should have the same shape as the corresponding output dimension.

[0130] like Figure 2 As shown in step 3 of the module, firstly, the physical model output of each counter τ Stacking in the time dimension, we get P p The loss function is calculated with the real label and the loss is recorded as Loss p , the loss function used is MAE loss, as follows:

[0131]

[0132] Where N is the number of samples, y i is the true value, is the predicted value of the model.

[0133] Similarly, for H L Also calculate the loss function and get Loss l , take the weighted average of the two loss functions to get the final total loss Loss t =0.5*Loss p +0.5*Loss l , and use this total loss for gradient backpropagation.

[0134] After the training is completed, steps 101-106 can be used to predict the airport traffic.

[0135] In addition, this embodiment also carried out a series of targeted experiments, the specific details of which are described as follows:

[0136] The routine using this system / method has the following effects:

[0137] The minimum mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) are commonly used as measurement standards in related fields, and "this algorithm" in each chart is the algorithm of this paper, while other legends are other existing algorithms. Algorithm 1 is a spatiotemporal graph convolutional network (STGCN): a network architecture that uses graph convolutional neural networks and causal convolutional networks to capture spatial and temporal correlations respectively; Algorithm 2 is an attentional spatiotemporal graph convolutional network (ASTGCN): a spatiotemporal graph convolutional network that introduces an attention mechanism to dynamically learn spatiotemporal relationships, combining graph convolutional neural networks and causal convolutional networks; Algorithm 3 is a multi-task graph neural network (MTGNN): a model that uses a learnable graph structure to model the correlation between multiple variables. It combines dilated convolutions and jump connections to capture spatiotemporal dependencies; Algorithm 4 is a graph convolutional network (TGCN): a network architecture that uses graph convolutional neural networks to capture spatial correlations and gated recurrent units to capture temporal correlations; Algorithm 5 is a graph convolutional network (TGCN): a network architecture that uses graph convolutional neural networks to capture spatial correlations and gated recurrent units to capture temporal correlations. The results of various indicators on the experimental test set are as follows:

[0138] Among them, Figure 3 As shown, the x-axis represents the test time period, the unit is 30 minutes, and the y-axis represents the minimum root mean square error RMSE of the test, the unit is times (times). The definition of the RMSE value is as follows:

[0139]

[0140] In the formula, y i is the actual observed value, is the corresponding predicted value and n is the number of samples.

[0141] Among them, Figure 4 As shown, the x-axis represents the test time period, the unit is 30 minutes, and the y-axis represents the mean absolute error MAE of the test, the unit is times (times). The definition of the MAE value is as follows:

[0142]

[0143] In the formula, y i is the actual observed value, is the corresponding predicted value and n is the number of samples.

[0144] Among them, Figure 5 As shown, the x-axis represents the test time period, the unit is 30 minutes, and the y-axis represents the mean absolute percentage error MAPE of the test, the unit is percentage (%), and the definition of the MAPE value is as follows:

[0145]

[0146] In the formula, y i is the actual observed value, is the corresponding predicted value and n is the number of samples.

[0147] Embodiment 2

[0148] like Figure 6 As shown, the present application provides a deep learning airport traffic prediction system with an embedded physical model, including:

[0149] The information acquisition module 601 is used to acquire the airport information of the target area within a set time; the airport information includes the actual take-off and landing data of the airport and the airport weather.

[0150] The division module 602 is used to divide the airport information according to fixed time periods to obtain multi-scale feature data; the multi-scale feature data is composed of the single-point scale flow data of the airport and the flow data on the route scale.

[0151] The slicing module 603 is used to perform a slicing operation on the multi-scale feature data. When the number of slicing times is reached, the multi-scale feature data fusion process is performed on each slice in turn to obtain a multi-scale fused feature vector set; the feature vector set is composed of a number of hidden variable outputs.

[0152] The first extraction module 604 is used to take the multi-scale fused feature vector set and the single-point scale flow data as input data, and extract the time information of the input data through dilated causal convolution and gating operations.

[0153] The second extraction module 605 is used to extract the time information of the input data by using a graph matrix and an adaptive matrix to obtain a time output after extracting the spatial information through diffusion graph convolution.

[0154] The prediction module 606 is used to perform residual connection on the time output after the spatial information is extracted through the diffusion graph convolution, and adopt a linear transformation operation to obtain airport traffic prediction data.

[0155] In summary, this application has the following technical effects:

[0156] This application proposes a deep learning airport aircraft flow prediction algorithm with an embedded physical model. The algorithm combines physical models, graph neural networks, and convolutional networks to learn multi-scale spatiotemporal information, including single-point scale and route scale in space, as well as long-term scale and short-term scale in time, so as to enhance the accuracy and comprehensiveness of the prediction. By introducing physical model regularization, the output of the physical model is combined with the output of the deep learning model to realize the constraint and optimization of the model.

[0157] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0158] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A deep learning airport traffic prediction method with an embedded physical model, characterized in that: include: Obtain airport information of the target area within a set time; the airport information includes actual take-off and landing data of the airport and airport weather; Dividing the airport information according to fixed time periods to obtain multi-scale feature data; The multi-scale feature data is composed of single-point scale flow data of the airport and flow data on the route scale; Slicing the multi-scale feature data, and when the number of slices is reached, performing a multi-scale feature data fusion process on each slice in turn to obtain a multi-scale fused feature vector set; the feature vector set is composed of a plurality of hidden variable outputs; Taking the multi-scale fused feature vector set and the single-point scale flow data as input data, and extracting the time information of the input data through dilated causal convolution and gating operations; According to the time information of the input data, a graph matrix and an adaptive matrix are used for extraction to obtain a time output after extracting spatial information through diffusion graph convolution; The temporal output after the spatial information is extracted by the diffusion graph convolution is subjected to residual connection, and a linear transformation operation is performed to obtain airport traffic prediction data.

2. The deep learning airport traffic prediction method with an embedded physical model according to claim 1 is characterized in that: The multi-scale feature data is sliced, and when the number of slices is reached, the multi-scale feature data fusion process is sequentially performed on each slice to obtain a multi-scale fused feature vector set, specifically including: Slicing the multi-scale feature data, when the number of slices τ=H / L1, performing a multi-scale feature data fusion process on each slice in turn to obtain a multi-scale fused feature vector set; Among them, τ is the number of slices collected by the counter, H is the length of the time dimension, and L1 is the set time interval.

3. The deep learning airport traffic prediction method with an embedded physical model according to claim 1 is characterized in that: The multi-scale feature data fusion process includes linear fusion, feature fusion based on graph convolutional network and gated recurrent unit processing.

4. The deep learning airport traffic prediction method with an embedded physical model according to claim 3 is characterized in that: The multi-scale feature data is sliced, and when the number of slices is reached, the multi-scale feature data fusion process is sequentially performed on each slice to obtain a multi-scale fused feature vector set, specifically including: The multi-scale feature data is sliced, and when the number of slices is reached, linear fusion, feature fusion based on a graph convolutional network, and gated recurrent unit processing are sequentially performed on each slice.

5. The deep learning airport traffic prediction method with embedded physical model according to claim 4 is characterized in that: When performing linear fusion on slices, it includes: Use linear layer pairs The feature channels are aggregated and recorded as Linear agg ; Linear agg is a linear layer named agg; is the τth slice of the single-point scale flow data; Through the Linear layer att Linear Input agg Perform linear transformation to generate the time weight of the kth airport node at time t Linear att is a linear layer named att; The time weight After normalization, we get exp(·) represents an exponential function, which is used to perform nonlinear transformation on weights; is the normalized time weight corresponding to airport node k at time t; Based on the formula Sure Time weight at time t; According to the formula Linearly fuse the slices; in, express The traffic data of the i-th node at time t, is the new feature data after fusion, for The time weight at time t, subscript 2k represents The kth node of .

6. The deep learning airport traffic prediction method with embedded physical model according to claim 5 is characterized in that: When performing graph convolutional network-based feature fusion on slices, it specifically includes: Based on the feature data after linear fusion, the diffusion graph convolution operation and the activation function ReLU are used to pass the graph matrix A τ temp right The feature data is learned and fused to obtain the spatial output; the formula expression of the spatial output is: in, represents the spatial output under the counter τ, the subscript spatial represents the spatial convolution output; ReLU is the activation function; step represents the number of steps of the diffusion convolution; represents the forward diffusion matrix, f represents forward propagation, τ is the time step, rowsum(·) is the row sum, and the superscript represents the matrix power series; represents the diffusion matrix of back propagation at time step τ, b represents back propagation, and the rest is the same as above; W τ1 Represents the forward weight matrix at the step number step, 1 represents forward; W step2 Represents the backward weight matrix at step number step, 2 represents backward.

7. The deep learning airport traffic prediction method with embedded physical model according to claim 6 is characterized in that: When performing gated recurrent unit processing on a slice, it includes: Using a GRU unit to perform temporal learning on the spatial output, to obtain a reset gate output of the GRU unit and an update gate output of the GRU unit; According to the formula and formula Determine the hidden state output of the current GRU unit; in, represents the candidate hidden state at time step T, ^ represents the candidate; tanh represents the activation function tanh; represents the weight matrix used by the algorithm to calculate the candidate hidden state in the counter τ step, W is the weight matrix, and h represents the weight matrix used for the candidate hidden state; h τ-1 represents the hidden variable output obtained when the counter is τ-1; It indicates that the counter τ is used to calculate the bias vector of the candidate hidden state, h represents the candidate hidden state; ⊙ represents the Hadamard product.

8. The deep learning airport traffic prediction method with embedded physical model according to claim 1 is characterized in that: The multi-scale fused feature vector set and the flight flow transfer sum data set between the airports are used as input data, and the time information of the input data is extracted through dilated causal convolution and gating operations, specifically including: According to the formula formula And the formula o=g(θ1·ouput a +b)⊙σ(θ2·ouput b +c), extracting the time information of the input data; Among them, ouput a represents the output after convolution with the convolution kernel f1(t), a represents the result of the first set of convolution kernels; ouput b represents the output after convolution with the convolution kernel f2(t), b represents the first set of convolution kernel results; f1(s) represents the weight of the first convolution kernel at position s; f2(s) represents the weight of the second convolution kernel at position s; t represents the time step; d represents the step size of the convolution; K represents the size of the convolution kernel; g represents the activation function tanh; σ represents the activation function sigmoid; b represents the bias for the first set of convolution results; c represents the bias for the second set of convolution results; θ1 represents the weight for the first set of convolution results; θ2 represents the weight for the second set of convolution results; ⊙ represents the Hadamard product; * represents the convolution operation.

9. The deep learning airport traffic prediction method with embedded physical model according to claim 1 is characterized in that: According to the time information of the input data, the use of the map matrix A and the adaptive matrix A adp Extraction is performed to obtain the temporal output after extracting spatial information through diffusion graph convolution, which specifically includes: According to the formula Get the temporal output after extracting spatial information through diffusion graph convolution; Among them, s_output represents the temporal output after extracting spatial information through diffusion graph convolution; represents the value of the forward diffusion matrix at order k, f1 represents the forward diffusion matrix, represents the value of the back diffusion matrix at order k, b1 represents the back diffusion matrix, represents the value of the adaptive matrix at order k; W k1 represents the weight matrix used for the forward diffusion matrix at order k; W k2 represents the weight matrix used for the back diffusion matrix at order k; W k represents the weight matrix used for the adaptation matrix at order k.

10. A deep learning airport traffic prediction system with an embedded physical model, characterized in that: include: Information acquisition module, used to obtain airport information of the target area within a set time; The airport information includes actual airport take-off and landing data and airport weather; A division module, used to divide the airport information according to fixed time periods to obtain multi-scale feature data; The multi-scale feature data is composed of single-point scale flow data of the airport and flow data on the route scale; A slicing module is used to perform a slicing operation on the multi-scale feature data. When the number of slicing times is reached, the multi-scale feature data fusion process is performed on each slice in turn to obtain a multi-scale fused feature vector set; the feature vector set is composed of a number of hidden variable outputs; A first extraction module is used to take the multi-scale fused feature vector set and the single-point scale flow data as input data, and extract the time information of the input data through dilated causal convolution and gating operations; A second extraction module is used to extract the time information of the input data by using a graph matrix and an adaptive matrix to obtain a time output after extracting the spatial information by diffusion graph convolution; The prediction module is used to perform residual connection on the time output after the spatial information is extracted through the diffusion graph convolution, and adopt linear transformation operation to obtain airport traffic prediction data.

Images