Flight flow distribution scheme generation method based on GAN

Abstract

The invention provides a GAN-based flight flow distribution scheme generation method, which comprises the following steps: collecting flight flow data, and establishing an air flight network model; performing threat degree prediction to obtain an initial flow distribution scheme; constructing a generative adversarial network (GAN), and carrying out iterative training; using the trained confrontation generation network GAN to obtain a first air flight flow distribution scheme; performing air flight situation assessment according to the current flight flow data; in combination with a preset target, a second air flight flow distribution scheme is obtained through calculation; and selecting from the first air flight flow distribution scheme and the second air flight flow distribution scheme to obtain a final flight flow distribution scheme. According to the method, the problems that aircraft scenes are changeable, planning and decision making of flight network space-time configuration need to comprehensively consider many factors are solved, intelligent decisions are made without being affected by time and space limitations, and the management efficiency and safety of air flight flow are improved.

Description

Technical Field

[0001] The present invention relates to a method for generating a flight flow distribution scheme, in particular to a method for generating a flight flow distribution scheme based on GAN. Background Art

[0002] This section merely provides background information related to the present disclosure and is not necessarily prior art.

[0003] The current conventional strategy for flight flow distribution is to first identify the bottleneck points of route flight flow, and then establish an optimization model to optimize flow control. Traditional flow control strategies include ground holding model, air holding model, rerouting strategy, etc. These classic methods mostly use expert solutions to make artificial planning decisions, which are affected by time and space limitations, and the planning made cannot guarantee the optimal strategy.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0005] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for generating a flight flow allocation scheme based on GAN in view of the shortcomings of the prior art.

[0006] In order to solve the above technical problems, the present invention discloses a method for generating a flight flow allocation scheme based on GAN, the method comprising the following steps:

[0007] Step 1: collect current and historical flight flow data and establish an air flight network model; perform threat level prediction based on the target threat level evaluation index system and historical flight flow data, and obtain a preliminary flow distribution plan by combining the current flight flow data and the predicted threat level;

[0008] Step 2, constructing a generative adversarial network (GAN), using the collected current flight traffic data as the input of the generator in the generative adversarial network (GAN), using the preliminary traffic allocation plan as the input of the discriminator in the generative adversarial network (GAN), and iteratively training the generative adversarial network (GAN); using the trained generative adversarial network (GAN) to obtain the first air flight traffic allocation plan;

[0009] Step 3: Conduct an air flight situation assessment based on current flight traffic data;

[0010] Step 4, combining the results of the air flight situation assessment and the preset goals, calculating a second air flight flow allocation plan;

[0011] Step 5: Select a final flight flow allocation plan from the first flight flow allocation plan and the second flight flow allocation plan.

[0012] Further, the current and historical flight traffic data described in step 1 include:

[0013] Flight origin, flight destination, route, number of flights and estimated flight time.

[0014] Furthermore, the establishment of the air flight network model in step 1, that is, calculating according to the current and historical flight flow data to obtain the air flight network model, is specifically as follows:

[0015] Flight set: S = {S1, S2, ..., S k , …, S n}

[0016] Flight time set: t = {t1, t2, ..., t k ,…,t n}

[0017] Route density: A

[0018] Average flight time:

[0019] Sector capacity: B

[0020] Among them, S k represents the kth flight, n represents the number of flights, t k represents the flight time of the kth flight;

[0021] The calculation method of route density A is as follows:

[0022] Route density A = number of flights / preset time period

[0023] Average flight time The calculation method is as follows:

[0024]

[0025] Furthermore, the step 1 of obtaining a preliminary traffic allocation plan specifically includes:

[0026] Step 1-1, define the target threat evaluation index, as follows:

[0027] D threat =∑ω i ·μ i

[0028] Among them, D threat is the target threat level, ωi is the weight of the i-th factor indicator, μ i is the i-th factor indicator;

[0029] Step 1-2, the air flight network model is divided into three levels, namely, the target level, the criterion level and the indicator level, a fuzzy matrix is ​​constructed for the indicators of each level, the elements in the matrix are used to represent the relationship between the factor indicators, and the fuzzy hierarchy analysis method is used to obtain the weights of the factor indicators;

[0030] Steps 1-3, use linear programming method to build an optimization model;

[0031] Step 1-4, based on the optimization model constructed in step 1-3, obtain the preliminary traffic distribution plan as follows:

[0032] Air flight conditions: T = {T1, T2, T3}, where T1 represents relaxation, T2 represents moderate, and T3 represents congestion;

[0033] Decision variable information: R = {R1, R2, R3, R4, R5, R6}, where R1 represents sector traffic, R2 represents airport traffic, R3 represents route traffic, R4 represents sector airspace utilization, R5 represents route traffic saturation, and R6 represents the number of safety incidents;

[0034] Allocation scheme: x = {x1, x2, x3}, where x1 represents air holding, x2 represents diversion, and x3 represents trailing interval.

[0035] Furthermore, the adversarial generative network GAN described in step 2 specifically includes:

[0036] Generator and discriminator, where the generator is used to learn the distribution of real data and the discriminator is used to distinguish between real data and generated data;

[0037] The generator and the discriminator play a minimax game with the value function V(G,D), which is expressed as follows:

[0038]

[0039] Where z refers to random noise; p z (z) refers to the probability distribution of random noise z; x~p Data It refers to the probability distribution that the real data obeys; D(G(z)) refers to the probability that the discriminator believes that the image generated by the generator is a real image; Represents the real data distribution p Data expectations, Represents the input distribution p to the generator z expectations;

[0040] The training criterion of the discriminator is to maximize the value function V(G,D), and the training criterion of the generator is to minimize the value function V(G,D); the loss function L of the discriminator D , which is expressed as follows:

[0041]

[0042] The loss function L of the generator G , which is expressed as follows:

[0043]

[0044] During the training process of the generative adversarial network (GAN), the loss functions of the generator and the discriminator are optimized alternately, that is, the generator is fixed to optimize the discriminator, and then the discriminator is fixed to optimize the generator, and the cycle continues until the Nash equilibrium is reached.

[0045] Furthermore, the optimization model is constructed using the linear programming method described in steps 1-3, as follows:

[0046] Step 1-3-1, design the objective function as follows:

[0047] min Z=ρ1·f1(x1)+ρ2·f2(x2)+ρ3·f3(x3)

[0048] Wherein, Z is the objective function in the preliminary traffic allocation scheme, and the objective function includes: total delay cost; x1 represents the air holding strategy; x2 represents the rerouting strategy; x3 represents the trailing interval strategy; ρ1, ρ2 and ρ3 are weights, indicating the importance of each strategy;

[0049] Step 1-3-2, the design constraints are as follows:

[0050] The first constraint condition is to combine the weather influence factor W of the kth flight k , restrict the flights added to the sector under high load conditions to reduce the risk of airspace congestion:

[0051]

[0052] The second constraint is the sector capacity factor L allowed for the kth flight in bad weather. k , constraining the sector capacity allowed in bad weather:

[0053]

[0054] The third constraint is to constrain the total number of flights in a sector within a specific time period:

[0055]

[0056] Among them, x k represents the strategy of the kth flight, L max Indicates the maximum sector capacity, W max Indicates the maximum value of the impact of severe weather, C max Indicates the maximum sector traffic value.

[0057] Furthermore, the iterative training of the generative adversarial network GAN described in step 2 specifically includes:

[0058] Step 2-1, using the collected current flight traffic data as the input of the generator in the adversarial generative network GAN network, and obtaining a virtual traffic allocation plan through the generator;

[0059] Step 2-2, converting the preliminary traffic distribution plan into an image and inputting it into the discriminator as real data; wherein, the preliminary traffic distribution plan is converted into an image, taking RGB image as an example, and the specific method is as follows:

[0060] The values ​​of the six decision variable information R in the preliminary traffic allocation plan are combined in pairs, and then the combined variable values ​​are converted to the range of 0-255, mapped to different color channels of the RGB image, and the mapped RGB image is used to represent the preliminary traffic allocation plan;

[0061] Step 2-3, simultaneously input the virtual traffic allocation scheme obtained in step 2-1 into the discriminator;

[0062] Step 2-4, use the discriminator to discriminate the two input solutions, that is, use the FID evaluation index to calculate the similarity of the two solutions. If the similarity is greater than the threshold, return to step 2-1 for cyclic iterative training. If the similarity is less than the threshold, the training is completed.

[0063] Further, the air flight situation assessment as described in step 3 specifically includes:

[0064] The situation includes: aircraft position, speed, altitude, route, flight schedule, flight delays, weather conditions and air traffic management instructions;

[0065] The evaluation indicators include: sector flow, sector flow capacity ratio, route traffic saturation, controller load and severe weather impact;

[0066] Assessment methods include:

[0067] Step 3-1, constructing a flight traffic network knowledge graph, wherein the knowledge graph includes flights, weather, control factors and the interactive relationships between the factors;

[0068] Step 3-2, flight traffic network situation prediction, using the long short-term memory network LSTM and the Prophet method-based model to predict sector traffic and calculate the evaluation index;

[0069] Step 3-3, complete the air flight situation assessment by collecting flight plans and using the flight flow network situation prediction results.

[0070] Furthermore, the calculation described in step 4 obtains the second air flight flow allocation scheme, and the specific calculation method includes:

[0071] F=αD+βC+γL+μW+ωE

[0072] Among them, F represents the total delay cost; D is the flight density in the preset time period or area; C represents the congestion level in the preset time period or area; L is the airspace restriction factor, which reflects the degree of unavailability or restriction of a specific area; W represents the impact factor of weather conditions; E represents the adjustment factor for emergency or priority flights;

[0073] By adjusting the parameters, the flight density D in the preset time period or area, the congestion level C in the preset time period or area, the airspace restriction factor L, the weather condition influencing factor W and the adjustment factor E of the emergency or priority flight are obtained when the total delay cost F is the lowest; taking them as constraints, combined with the preliminary flow allocation plan, the control strategy is determined, and the second air flight flow allocation plan is obtained, which is described as follows:

[0074] Flight Status i,t , represents the state of flight i at time t, the state includes: waiting state, take-off preparation state, flight state, descent preparation state and landing state;

[0075] Description of location information i,t , represents the position of flight i at time t, represented by latitude and longitude, sector or route number;

[0076] Traffic distribution decision point D i,k , as follows:

[0077]

[0078] Where k is the sequence number of the flight decision point.

[0079]

[0080] Furthermore, the selection described in step 5 results in a final flight flow allocation solution, including:

[0081] Step 5-1, establishing a hierarchical structure showing the logical relationship of each level in the air flight network model;

[0082] Step 5-2, introduce experts and use fuzzy language to evaluate the importance of factor indicators at each level to obtain the fuzzy discriminant matrix;

[0083] Step 5-3, normalize the fuzzy discriminant matrix and calculate the weight of each factor index in the flow allocation decision;

[0084] Step 5-4, combine the weights with the corresponding evaluation values, conduct fuzzy comprehensive evaluation, and determine the optimal decision-making plan.

[0085] Beneficial effects:

[0086] 1. The present invention uses historical and existing data as a training set to train the GAN network, so that a realistic flight flow distribution plan that conforms to the actual situation can be generated. At the same time, once new data is available, the model can be updated by retraining the GAN to ensure its accuracy.

[0087] 2. The present invention proposes a deep neural network structure including two networks, wherein the two networks compete with each other, and the generation and discrimination capabilities are continuously improved in the confrontation training, integrating the idea of ​​game confrontation. It can effectively avoid the problem that the planning is greatly influenced by the decision makers, and obtain efficient and intelligent decisions. It can also take into account the complex factors involved in the flight flow distribution and generate a comprehensive flight flow distribution plan. In practical applications, using this flow control generation method, compared with the traditional distribution method, more accurate and intelligent strategy support can be obtained in the planning and management of air flight flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0089] Figure 1 It is a GAN-based flight flow distribution scheme generation method and system flow chart of the present invention.

[0090] Figure 2 It is an example diagram of the flight flow target threat assessment index system according to an embodiment of the present invention.

[0091] Figure 3 Schematic diagram of allocation scheme conversion according to an embodiment of the present invention.

[0092] Figure 4 It is a schematic diagram of an algorithm based on a GAN network according to an embodiment of the present invention.

[0093] Figure 5 2 is a diagram showing an example of a GAN-based network structure according to an embodiment of the present invention.

[0094] Figure 6 It is the flight situation assessment component of the embodiment of the present invention.

[0095] Figure 7 This is the knowledge graph construction process of an embodiment of the present invention.

[0096] Figure 8 It is a visualization of the flight traffic network knowledge graph according to an embodiment of the present invention.

[0097] Fig. 9 This is the indicator weight calculation process of the embodiment of the present invention.

[0098] Fig.10 It is a network architecture based on knowledge graph in an embodiment of the present invention.

[0099] Fig.11 It is a GAN applicable to the allocation scheme of the embodiment of the present invention.

[0100] Fig.12 The figure is a simulation diagram when the threshold is set to 2 in an embodiment.

[0101] Fig.13 This is a simulation diagram when the threshold is set to 3 in an embodiment. DETAILED DESCRIPTION

[0102] The present invention proposes a method for generating a flight flow distribution scheme based on GAN (Generative Adversarial Network, GAN), and proposes a deep neural network structure including two networks. It is a generative model in which the two networks compete with each other, and the generation and discrimination capabilities are continuously improved in adversarial training, integrating the idea of ​​game confrontation. It can effectively avoid the problem that the planning is greatly affected by the decision makers, and obtain efficient and intelligent decisions. It can also take into account the complex factors involved in the flight flow distribution and generate a comprehensive flight flow distribution scheme. In practical applications, using this flow control generation method, compared with the traditional distribution method, more accurate and intelligent strategy support can be obtained in the planning and management of air flight flow.

[0103] The overall idea of ​​the technical solution of the present invention is as follows:

[0104] The air flight flow distribution plan is generated by iteratively training the GAN network, so as to achieve the purpose of generating the corresponding optimal flow distribution strategy for any air flight flow data input. The method described is to generate distribution plans for different air flight flow scenario environments by iteratively training the GAN network. It mainly includes: acquiring data on the current air flight situation and deriving a preliminary flow distribution plan; using the data as the input of the GAN generator for iterative training; combining the current air flight situation and the existing flight data to conduct an air flight situation assessment; combining the results of the situation assessment with the preset goals, designing a calculation formula to generate the optimal air flight flow distribution plan.

[0105] The technical solution mainly includes the following steps:

[0106] Collect current relevant air flight traffic data, establish an air flight network model, make situation predictions through the target threat evaluation index system, and derive a preliminary traffic allocation plan based on historical records;

[0107] The collected data is used as the input of the GAN network generator, and the preliminary allocation scheme is used as the real data input of the discriminator to iteratively train the GAN;

[0108] Conduct air flight situation assessment based on current air flight conditions and existing flight data;

[0109] Combining the situation assessment results with the preset goals, a calculation formula is designed to generate the best air traffic flow allocation plan;

[0110] The best solution is obtained from the solutions generated by GAN and those generated by the formula.

[0111] Embodiment 1:

[0112] like Figure 1 As shown, a method for generating a flight flow allocation scheme based on GAN includes the following steps:

[0113] S1 collects current relevant air flight traffic data, establishes an air flight network model, makes situation predictions through the target threat evaluation index system, and derives a preliminary traffic allocation plan based on historical records;

[0114] S2 uses the collected data as the input of the GAN network generator, and the preliminary allocation scheme as the real data input of the discriminator to iteratively train the GAN;

[0115] S3 combines the current air flight situation with the existing flight data to conduct an air flight situation assessment;

[0116] S4 combines the situation assessment results and preset goals to design a calculation formula to generate the best air flight flow allocation plan;

[0117] S5 derives the best solution from those generated by GAN and those generated by the formula.

[0118] When formulating the preliminary air traffic allocation plan, we comprehensively used historical data and current air traffic information, carried out historical data integration and preprocessing, historical traffic analysis, demand forecasting, capacity assessment and other steps, and finally the preliminary traffic allocation. We used the linear programming method to build an optimization model to allocate traffic:

[0119]

[0120] sA i x i ≤ b i ,i=1,2,…,n

[0121] x i ≥0

[0122] Where Z is the objective function (such as total delay cost), x i represents the decision variables (such as the take-off and landing time or route selection of each flight), c i is the relevant cost or benefit coefficient, A i and b i Various constraints are defined (such as sector capacity, time window, etc.). The specific process is simply to input data such as flight density, airspace capacity, and weather conditions, with the goal of minimizing delay costs, and to change the priority of the strategy by adjusting the weights. According to different conditions, the control strategy for flight selection is determined. The priorities are as follows: if the number of flights in the sector is close to the upper limit, air holding is preferred; in bad weather conditions, trailing interval is preferred; if the sector is loaded, the rerouting strategy is preferred.

[0123] like Figure 2 As shown, it is an example diagram of the flight flow target threat assessment index system of an embodiment of the present invention. The target threat level of the flight flow network refers to the assessment of the potential threats and risks faced by specific targets or elements (such as sector airspace, aircraft) in the flight flow network. The flight flow network includes aircraft, routes, airports, navigation facilities, communication systems, air traffic management facilities, etc., which together constitute a complex aviation system. The purpose of the target threat assessment is to identify and quantify the severity of the threat in order to formulate an effective risk management strategy to protect the safety and availability of the flight flow network.

[0124] The main operating characteristics include sector traffic, route traffic, airport traffic, sector airspace utilization, and route traffic saturation. Sector traffic refers to the aircraft traffic within the sector, which indicates the density of the number of aircraft in a specific airspace. Route traffic refers to the aircraft traffic on a specific route, which indicates the density of the number of aircraft on that route. Airport traffic: refers to the aircraft traffic at a specific airport, which indicates the density of the number of aircraft at that airport. Sector airspace utilization refers to the degree of utilization of air traffic within a specific airspace, which can reflect the degree of congestion in the area. Route traffic saturation refers to the proportional relationship between the number of aircraft carried by a specific route and its design capacity, which is used to measure the degree of congestion on the route.

[0125] The controller load mainly adopts the coordination load. Coordination Load: Coordination Load includes information sharing and coordination work within the air traffic management system and with other control departments, airports and relevant stakeholders. The impact of severe weather includes: flight delay rate, flight cancellation rate, and number of safety incidents.

[0126] Combining the quantification method of each key operation feature in the multi-dimensional flight flow spatiotemporal operation feature system, the present invention introduces the flight flow network target threat assessment index, which is defined as follows:

[0127] D threat =∑ω i ·μ i

[0128] =ω section ·μ secti +ω route ·μ route +ω manager ·μ manager +ω delay ·μ delay +ω cancel ·μ cancel +ω accident ·μ accident

[0129] Where D thre ω is the target threat level, which is an assessment of the potential threats and risks faced by the target (a sector of airspace). i is the weight of the feature. The weight can be customized or calculated by FAHP method (Fuzzy Hierarchy Analysis Method, Reference: Bian Xiaofeng, Wang Lin. Analysis of Fatigue Risk Factors of Controllers Based on Fuzzy Hierarchy Analysis Method [J]. Science and Technology Innovation and Application, 2020, (03): 45-47.) as needed. section represents the weight of the sector traffic feature, ω route represents the weight of the route flow feature, ω managerepresents the weight of the coordinated regulatory load feature, ω delay represents the weight of the feature of flight delay rate, ω cancel represents the weight of the feature of flight cancellation rate, ω accident Represents the weight of the feature of number of security events. i are the characteristic variables, where μ section is the sector flow, μ route is the airway flow, μ manage To coordinate the control load, μ delay is the flight delay rate, μ cancel is the flight cancellation rate, μ accident The number of security incidents.

[0130] The fuzzy hierarchical analysis method is used to obtain the weights of the factor indicators. The specific method is as follows:

[0131] Step 1: Create a target threat assessment indicator system, create a hierarchical structure, and adjust the initial complex assessment object to make it hierarchical, systematic, and organized;

[0132] Step 2: Adjust the fuzzy complementary matrix R = (r ij ) n×n , convert it into a fuzzy consistent matrix;

[0133] Step 3: hierarchical sorting. The weight value to be calculated. is a condition that needs to be satisfied. Then, the weight vector The solution can be transformed into a least squares problem that satisfies the following conditions:

[0134]

[0135] The constraint condition for parameter a is The size ranking is the result of comparing the required indicators with each other relative to the previous layer.

[0136] ① Hierarchical synthesis. When there is a three-level hierarchy, the weight of the previous layer relative to the previous layer is calculated using the same method as above, and the result is W = (w1, w2, ..., w n ).

[0137] ② Overall hierarchical sorting. Calculate the overall importance of the top target layer T i , then follow The indicators are sorted according to their size to obtain the final result of the weight corresponding to each indicator.

[0138] like Figure 3The figure is a schematic diagram of the allocation scheme conversion of an embodiment of the present invention. It is necessary to process the data a priori and write the initial data into a set form:

[0139] Air flight situation T = {T1, T2, T3}

[0140] Among them, T1 represents relaxation, T2 represents moderate, and T3 represents congestion.

[0141] Decision variable information R = {R1, R2, R3, R4, R5, R6}

[0142] Among them, R1 represents sector traffic, R2 represents airport traffic, R3 represents route traffic, R4 represents sector airspace utilization, R5 represents route traffic saturation, and R6 represents the number of safety incidents.

[0143] Allocation x = {x1, x2, x3}

[0144] Among them, x1 represents ground holding, x2 represents route air holding, and x3 represents diversion strategy.

[0145] Next, we need to use the preliminary allocation scheme as the real data input of the GAN network discriminator in the form of RGB images. This is achieved by using the channel separation method. It is known that RGB images have three color modes, namely red, green and blue. The values ​​of the above six decision variables R are combined in pairs and then the variable values ​​are converted to the range of 0-255, and then mapped to different channels of the RGB image, such as R 1、 The value of R2 is mapped to the red channel, R 3、 The value of R4 is mapped to the green channel, R 5、 The value of R6 is mapped to the blue channel, and finally an RGB image is used to represent different allocation schemes.

[0146] Because the GAN network model can generate more than just images, its input form is also different. In addition to the RGB images mentioned above, there are also forms such as text. Combined with the decision variables mentioned above, such as sector traffic, airport traffic, airway traffic saturation, and the number of safety incidents, they need to be converted into text before being input into the GAN network for training. The general steps are as follows:

[0147] Step 1 requires collecting relevant flight flow data and normalizing them to ensure that indicators of different scales and ranges can be compared and processed under the same framework. After that, outliers may need to be processed to ensure the accuracy and consistency of the data.

[0148] Step 2 obtains the relevant values ​​of the decision variables mentioned above;

[0149] Step 3 maps the numerical features to text. The numerical range of each dimension needs to be divided into several categories, and a label is defined for each category (the specific numerical division needs to be based on the actual data). It is expressed as follows:

[0150] ■Sector traffic: high i1, medium i2, low i3. n ∈(0,1),n=1,2,3,i1>i2>i3

[0151] ■ Airport traffic: busy j1, normal j2, light load j3. n ∈(0,1),n=1,2,3,j1>j2>j3

[0152] ■Route saturation: severe k1, moderate k2, slight k3. n ∈(0,1),n=1,2,3,k1>k2>k3

[0153] ■Safety incidents: frequent r1, occasional r2, rare r3. n ∈(0,1),n=1,2,3,r1>r2>r3

[0154] Step 4 combines the text labels of each feature in a certain order to form a sequence, which represents a specific air traffic condition and, of course, an allocation scheme. For example, a sequence may be: "i1-j1-k2-r2", and its corresponding allocation scheme is airborne waiting for the route.

[0155] like Figure 4 FIG. 1 is a schematic diagram of an algorithm based on a GAN network according to an embodiment of the present invention. The training of the GAN network is performed according to the following steps:

[0156] Step 1: Input the collected values ​​of the six decision variables R as training samples of the generator to generate an allocation plan;

[0157] Step 2: The preliminary traffic allocation scheme obtained after threat assessment and combined with historical information is converted into an RGB image using the channel separation method as the reference data for the discriminator in the GAN network;

[0158] Step 3: Input the allocation plan generated by the generator and the preliminary allocation plan into the discriminator for discrimination;

[0159] In step 4, the Fréchet Inception Distance (FID) is used as the criterion for judging whether the model training is complete. When the FID value is closer to 1, it means that the similarity between the generated sample and the real sample is worse, so we return to step 1 to retrain. The closer it is to 0, the more similar it is, and the model training is complete.

[0160] like Figure 5 As shown, it is an example diagram of the network structure based on GAN of an embodiment of the present invention. The structure of the network is divided into a generator and a discriminator. In the generator, latent variables with a shape of 200×25 are sampled from a standard normal distribution and then converted to a dimension suitable for convolution through a fully connected layer. They are then input into a one-dimensional upsampling convolution subnetwork, and the final allocation scheme is generated after three upsamplings. In the discriminator network, the generated trajectory is input into a one-dimensional convolution layer, and then a one-dimensional downsampling convolution subnetwork is used to reduce the resolution.

[0161] The activation function of the entire GAN network is a LeakyReLU, which not only retains the advantage of fast calculation speed, but also overcomes the disadvantage that neurons do not participate in updating when the input value is negative. The Dropout mechanism is added to both upsampling and downsampling networks to limit the interaction of hidden layer nodes and prevent model overfitting. The working principle of Dropout is to randomly "discard" (i.e. set to 0) the output of some neurons during training, thereby forcing the model to learn a more generalized feature representation.

[0162] Let x i is the input of the i-th neuron, ω i is the corresponding weight, b is the bias term, and f is the activation function. In the absence of Dropout, the output y of the neuron is i Calculated as:

[0163] y i =f(x i ω i +b)

[0164] After applying Dropout, during training, assuming that the probability of this neuron being selected for inactivation is p, its output becomes:

[0165]

[0166] The mask i Is a binary variable with a value of {0,1}, indicating whether the neuron is inactivated (if mask i =0, the neuron output is 0; if mask i = 1, then the output is normal, but in the forward propagation, it is not directly divided by p. In actual operation, the output of the neurons that are not inactivated is directly multiplied by ).

[0167] Batch Normalization (BatchNorm) reduces internal covariate shift and accelerates the training process by normalizing the input of each layer of the neural network. It also has a certain regularization effect. By introducing additional learnable parameters (scale γ and offset β) and processing small batches of training samples, it indirectly increases the generalization ability of the model, thereby helping to prevent overfitting. The specific steps are as follows:

[0168] In each forward propagation process, for a mini-batch of input data, S1 first calculates the mean μ of the batch data in each feature dimension B and variance

[0169] S2 uses these statistics to normalize each sample in the mini-batch to have zero mean and unit variance:

[0170]

[0171] where x i is the sample, and ò is a small constant (such as 1e-5) used to avoid division by zero errors.

[0172] In order to retain the expressive power of the network, S3 introduces two learnable parameters γ (scale) and β (bias) to adjust the normalized data:

[0173]

[0174] This way, the network can learn whether to keep the original scale and offset, or change them to better suit the task requirements.

[0175] Figure 6 As shown, the flight situation assessment composition of an embodiment of the present invention. Situation refers to the actual operating conditions and dynamic changes in the air traffic network at a specific time or over a period of time. This includes the aircraft's position, speed, altitude, route, flight plan, flight delays, weather conditions, air traffic management instructions and other relevant information. Combined with the current air flight situation and existing flight data, the evaluation indicators include sector flow, sector flow capacity ratio, route traffic saturation, controller load, severe weather impact, etc. The flight situation assessment consists of three parts: knowledge graph, target threat level, and situation prediction.

[0176] Flight traffic network situation prediction. First, the sector traffic, sector flow capacity ratio, route traffic saturation, as well as the operational characteristic indicators such as controller load and the impact of severe weather are numerically calculated and predicted. For example, the sector traffic is predicted using the Long Short-Term Memory Neural Network (LSTM) and the Prophet algorithm-based model. The next step is bottleneck identification. Here, all flight plan data is obtained through centralized flight plan processing, and the transit time of key waypoints is obtained using the transit time prediction model. The traffic prediction value of the key waypoint is obtained by statistics; then the traffic prediction value is compared with the capacity limit value to identify the flight traffic bottleneck point.

[0177] Figure 7 As shown, it is the knowledge graph construction process of the embodiment of the present invention. For the construction of the knowledge graph of flight network operation. First, knowledge extraction is required. The source of knowledge is the data set. After knowledge extraction, ontological knowledge expression is formed; secondly, the newly acquired knowledge from different sources is integrated and merged into a new data set; after the quality of the fused new knowledge is evaluated, it can be added to the knowledge base to ensure its quality. Finally, the ontology is constructed first to form an ontology library with a strong hierarchical structure, and then the knowledge graph is generated.

[0178] Figure 8 As shown, the flight flow network knowledge graph visualization of the embodiment of the present invention is shown. After the knowledge graph is constructed, it can also be visualized to facilitate a more intuitive understanding and operation of the data. The established flight flow knowledge graph is imported into the Neo4j graph database, and then the visualization operation is performed. The steps are as follows:

[0179] Enter the following in the Neo4j console:

[0180] ①CREATE INDEX ON:Resource(uri)

[0181] ②CALL semantics.importRDF('file: / / / D: / Project / 5Airport / work / Knowledge Graph / Sector Knowledge Graph-v2.owl','RDF / XML',{})

[0182] ③MATCH(n)

[0183] WHERE n.uri CONTAINS "<http: / / www.semanticweb.org / 86150 / ontologies / 2023 / 9 / untitled-ontology-13> "

[0184] SET n.uri=substring(n.uri,71)

[0185] RETURN

[0186] After configuring the style of the nodes, the knowledge graph can be visualized.

[0187] Fig. 9 The following is a flow chart showing the indicator weight calculation process of an embodiment of the present invention. After defining the flight traffic network threat assessment indicator, the most important step is to calculate the indicator weight, where the weight calculation process is as follows:

[0188] (1) Create a target threat assessment indicator system, create a hierarchical structure, and adjust the initial complex assessment object to make it hierarchical, systematic, and organized;

[0189] (2) Adjust the fuzzy complementary matrix R = (r ij ) n×n , convert it into a fuzzy consistent matrix;

[0190] (3) Hierarchical sorting. The weight value to be calculated. is a condition that needs to be satisfied. Then, the weight vector The solution can be transformed into a least squares problem that satisfies the following conditions:

[0191]

[0192] The constraint condition for parameter a is The size ranking is the result of comparing the required indicators with each other relative to the previous layer.

[0193] Fig.10 As shown, it is a network architecture based on knowledge graph of an embodiment of the present invention. FAHP is a method for multi-criteria decision making, which can help evaluate the threat level of different targets. In FAHP, a hierarchy needs to be established first to reflect the relationship between different levels. Then, experts provide comparative data to determine the relative importance of each level. Finally, weights are calculated based on these data and evaluated.

[0194] The general steps for determining the optimal allocation plan are:

[0195] Step 1 decomposes the decision-making problem of flight flow allocation into a hierarchical structure (goals, criteria, indicators). Each level has its corresponding factors. The indicator level corresponds to sector flow, sector flow capacity ratio, route traffic saturation, controller load, and bad weather impact. For each pair of factors in each layer, the relative importance between them is described by using fuzzy language and a fuzzy discriminant matrix is ​​constructed. The specific process of this step is as follows:

[0196] S1 clearly defines the quantitative standards of fuzzy language variables, and converts language descriptions such as "very important", "important", "average", "not very important", and "unimportant" into mathematical fuzzy sets. Usually, these language variables are mapped to membership functions with upper and lower limits. Very important: [0.8, 1]; important: [0.6, 0.8]; average: [0.4, 0.6]; not very important: [0.2, 0.4]; not important: [0, 0.2], where each interval represents a fuzzy set, expressing the uncertainty range of the decision maker's subjective evaluation of the relative importance of a factor;

[0197] S2 takes the indicator layer as an example. There are 5 factors that need to be evaluated in the indicator layer. Next, a 5×5 fuzzy discriminant matrix needs to be constructed, in which each element [i, j] represents the relative importance of the i-th factor relative to the j-th factor. The process is first that for each pair of factors (i, j), the decision maker selects a fuzzy linguistic variable according to his own judgment; secondly, according to the previously defined fuzzy linguistic scale, the selected linguistic variable is converted into the corresponding membership interval and filled in the corresponding position of the matrix; finally, this process is repeated until the entire fuzzy discriminant matrix is ​​filled.

[0198] Step 2 normalizes each fuzzy judgment matrix and converts it into a consistency matrix. Using the fuzzy discriminant matrix, the weight of each factor is calculated, which can be achieved through techniques such as the fuzzy eigenvector method.

[0199] Step 3 checks whether the fuzzy discrimination is consistent. If not, it needs to be adjusted. The weights of each sub-goal are combined with their evaluation values ​​to obtain a fuzzy comprehensive evaluation of each decision plan.

[0200] Step 4 ranks the decision options based on fuzzy comprehensive evaluation to determine the best option.

[0201] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0202] Fig.11 As shown in FIG. 1 , the classification of GANs applicable to the allocation scheme of the embodiment of the present invention is shown. Since the adversarial generative network appeared, it has been upgraded and iterated at a rapid pace. So far, there are many types of GANs. For the present invention, in addition to the WGAN-GP used above, cGAN and MM-GAN can also be applied.

[0203] The flight flow allocation scheme generation method based on conditional generative adversarial network (Conditional GAN, cGAN) has the following ideas:

[0204] S1 integrates historical flight data including past flight records, meteorological data, airspace restriction information, airport capacity, etc. with real-time situational conditions such as time and weather conditions.

[0205] S2 designs a generator and discriminator architecture based on a deep neural network with conditional input.

[0206] S3 is adjusted through joint training of the discriminator and generator, addition of conditional loss terms, and iterative training.

[0207] S4 uses domain expert knowledge and objective indicators (such as flight efficiency, delay rate, and airspace utilization) to evaluate the effectiveness and feasibility of the generated solutions, and compares the performance of the solutions generated by cGAN with existing methods or real data in a simulated environment.

[0208] The flight flow allocation scheme generation method based on multimodal generative adversarial network (MM-GAN) has the following ideas:

[0209] S1 collects historical flight data including flight path, time, altitude, speed, weather conditions, air traffic control information, etc., and performs data preprocessing;

[0210] S2 designs the model architecture, which includes multiple generators and a discriminator. Each generator is responsible for generating flight flow allocation plans under specific situational conditions such as different time periods. The discriminator not only needs to determine whether the generated plan is true, but also needs to distinguish the generated plans under different conditions to ensure that the plan output by each generator is not only of high quality, but also meets the corresponding condition settings;

[0211] S3 performs iterative training and optimization of the discriminator and then the generator;

[0212] S4 uses domain knowledge and actual operational indicators to evaluate the generated flight flow allocation plan, adjusts the model parameters based on the evaluation feedback, and performs multiple rounds of iterative optimization until satisfactory performance standards are achieved.

[0213] Embodiment 2:

[0214] In a specific embodiment, a flow control simulation is performed for a sector of a city in a certain region in East China. The simulation is performed based on a flight flow simulation system.

[0215] The main principle of traffic control is: first, build a route map through waypoints and route data, use the adjacency matrix to represent the distance between waypoints, and the distance calculation uses the spherical distance formula. Then, check the time and the number of aircraft in the sector. If the number of aircraft in a sector reaches the threshold, modify the weight of the edge passing through the sector in the route map and set it to infinity, thereby blocking the path. Subsequently, use the Dijkstra algorithm to recalculate the shortest path on the updated route map to ensure that the newly created aircraft route bypasses the overloaded sector. After completing the path calculation, restore the modified edge weights in the route map to maintain the accuracy of subsequent path planning. The aircraft automatically flies according to the new route.

[0216] During the simulation, the bottleneck threshold of the airspace is set to 2, and the simulation system is run. When the airspace reaches the preset bottleneck, 10 different moments can be selected to observe whether there is correct prompt information in the text output area.

[0217] like Fig.12 As shown, when both aircraft AC01 and aircraft AC02 are located in sector 11 of the city, the text output area prompts that the sector has reached the aircraft quantity threshold (2) and displays the current aircraft quantity. Then, the control strategy is operated.

[0218] like Fig.13 As shown in the figure, after the control strategy is operated, when the bottleneck threshold is set to 3, the number of aircraft in the 11th sector of Shanghai will not exceed the preset threshold, so the newly created aircraft AC05 flies to the destination airport to the right along the preset path. The purple path in the figure is the current route of aircraft AC05.

[0219] After changing the bottleneck threshold to 2, when aircraft AC05 is created, since the number of aircraft in sector 11 of the city is 2, the bottleneck threshold has been reached, so the route of aircraft AC05 is replanned. As can be seen from the figure above, the replanned new route first flies to the left and then to the destination Hongqiao Airport. This simulation shows that the control strategy method is effective.

[0220] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, the invention content of a method for generating a flight flow allocation scheme based on GAN provided by the present invention and some or all steps in each embodiment can be executed. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0221] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on such an understanding, the technical solutions in the embodiments of the present invention can be essentially or partly contributed to the prior art in the form of computer programs, i.e., software products, which can be stored in a storage medium and include several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, an MCU or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0222] The present invention provides a method and idea for generating a flight flow allocation scheme based on GAN. There are many methods and approaches to implement the technical scheme. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A method for generating a flight flow allocation scheme based on GAN, characterized in that: The method comprises the following steps: Step 1: collect current and historical flight flow data and establish an air flight network model; perform threat level prediction based on the target threat level evaluation index system and historical flight flow data, and obtain a preliminary flow distribution plan by combining the current flight flow data and the predicted threat level; Step 2, constructing a generative adversarial network (GAN), using the collected current flight traffic data as the input of the generator in the generative adversarial network (GAN), using the preliminary traffic allocation plan as the input of the discriminator in the generative adversarial network (GAN), and iteratively training the generative adversarial network (GAN); using the trained generative adversarial network (GAN) to obtain the first air flight traffic allocation plan; Step 3: Conduct an air flight situation assessment based on current flight traffic data; Step 4, combining the results of the air flight situation assessment and the preset goals, calculating a second air flight flow allocation plan; Step 5: Select a final flight flow allocation plan from the first flight flow allocation plan and the second flight flow allocation plan.

2. The method for generating a flight flow allocation scheme based on GAN according to claim 1, characterized in that: Current and historical flight traffic data as described in Step 1, including: Flight origin, flight destination, route, number of flights and estimated flight time.

3. The method for generating a flight flow allocation scheme based on GAN according to claim 2, characterized in that: The establishment of the air flight network model in step 1, that is, calculating according to the current and historical flight flow data to obtain the air flight network model, is specifically as follows: Flight set: S = {S1, S2, ..., S k , …, S n } Flight time set: t = {t1, t2, ..., t k ,…,t n } Route density: A Average flight time: Sector capacity: B Among them, S k represents the kth flight, n represents the number of flights, t k represents the flight time of the kth flight; The calculation method of route density A is as follows: Route density A = number of flights / preset time period Average flight time The calculation method is as follows:

4. The method for generating a flight flow allocation scheme based on GAN according to claim 3, characterized in that: The preliminary traffic allocation plan described in step 1 specifically includes: Step 1-1, define the target threat evaluation index, as follows: D threat =Sow i ·m i Among them, D threat is the target threat level, ω i is the weight of the i-th factor indicator, μ i is the i-th factor indicator; Step 1-2, the air flight network model is divided into three levels, namely, the target level, the criterion level and the indicator level, a fuzzy matrix is ​​constructed for the indicators of each level, the elements in the matrix are used to represent the relationship between the factor indicators, and the fuzzy hierarchy analysis method is used to obtain the weights of the factor indicators; Steps 1-3, use linear programming method to build an optimization model; Step 1-4, based on the optimization model constructed in step 1-3, obtain the preliminary traffic distribution plan as follows: Air flight conditions: T = {T1, T2, T3}, where T1 represents relaxation, T2 represents moderate, and T3 represents congestion; Decision variable information: R = {R1, R2, R3, R4, R5, R6}, where R1 represents sector traffic, R2 represents airport traffic, R3 represents route traffic, R4 represents sector airspace utilization, R5 represents route traffic saturation, and R6 represents the number of safety incidents; Allocation scheme: x = {x1, x2, x3}, where x1 represents air holding, x2 represents diversion, and x3 represents trailing interval.

5. The method for generating a flight flow allocation scheme based on GAN according to claim 4, characterized in that: The adversarial generative network GAN described in step 2 specifically includes: Generator and discriminator, where the generator is used to learn the distribution of real data and the discriminator is used to distinguish between real data and generated data; The generator and the discriminator play a minimax game with the value function V(G,D), which is expressed as follows: Where z refers to random noise; p z (z) refers to the probability distribution of random noise z; x~p Data It refers to the probability distribution that the real data obeys; D(G(z)) refers to the probability that the discriminator believes that the image generated by the generator is a real image; Represents the real data distribution p Data Expectations, E z ~ pz(z) Represents the input distribution p to the generator z expectations; The training criterion of the discriminator is to maximize the value function V(G,D), and the training criterion of the generator is to minimize the value function V(G,D); the loss function L of the discriminator D , which is expressed as follows: The loss function L of the generator G , which is expressed as follows: During the training process of the generative adversarial network (GAN), the loss functions of the generator and the discriminator are optimized alternately, that is, the generator is fixed to optimize the discriminator, and then the discriminator is fixed to optimize the generator, and the cycle continues until the Nash equilibrium is reached.

6. The method for generating a flight flow allocation scheme based on GAN according to claim 5, characterized in that: The optimization model is constructed using the linear programming method described in steps 1-3 as follows: Step 1-3-1, design the objective function as follows: minZ=ρ1·f1(x1)+ρ2·f2(x2)+ρ3·f3(x3) Wherein, Z is the objective function in the preliminary traffic allocation scheme, and the objective function includes: total delay cost; x1 represents the air holding strategy; x2 represents the rerouting strategy; x3 represents the trailing interval strategy; ρ1, ρ2 and ρ3 are weights, indicating the importance of each strategy; Step 1-3-2, the design constraints are as follows: The first constraint condition is to combine the weather influence factor W of the kth flight k , restrict the flights added to the sector under high load conditions to reduce the risk of airspace congestion: The second constraint is the sector capacity factor L allowed for the kth flight in bad weather. k , constraining the sector capacity allowed in bad weather: The third constraint is to constrain the total number of flights in a sector within a specific time period: Among them, x k represents the strategy of the kth flight, L max Indicates the maximum sector capacity, W max Indicates the maximum value of the impact of severe weather, C max Indicates the maximum sector traffic value.

7. The method for generating a flight flow allocation scheme based on GAN according to claim 6, characterized in that: The iterative training of the generative adversarial network GAN described in step 2 specifically includes: Step 2-1, using the collected current flight traffic data as the input of the generator in the adversarial generative network GAN network, and obtaining a virtual traffic allocation plan through the generator; Step 2-2, converting the preliminary traffic allocation plan into an image and inputting it into the discriminator as real data; wherein, the specific method of converting the preliminary traffic allocation plan into an image is as follows: The values ​​of the six decision variable information R in the preliminary traffic allocation plan are combined in pairs, and then the combined variable values ​​are converted to the range of 0-255, mapped to different color channels of the image, and the mapped image is used to represent the preliminary traffic allocation plan; Step 2-3, simultaneously input the virtual traffic allocation scheme obtained in step 2-1 into the discriminator; Step 2-4, use the discriminator to discriminate the two input solutions, that is, use the FID evaluation index to calculate the similarity of the two solutions. If the similarity is greater than the threshold, return to step 2-1 for cyclic iterative training. If the similarity is less than the threshold, the training is completed.

8. The method for generating a flight flow allocation scheme based on GAN according to claim 7, characterized in that: Step 3 generally describes the conduct of an air flight situation assessment, specifically including: The situation includes: aircraft position, speed, altitude, route, flight schedule, flight delays, weather conditions and air traffic control instructions; The evaluation indicators include: sector flow, sector flow capacity ratio, route traffic saturation, controller load and severe weather impact; Assessment methods include: Step 3-1, constructing a flight traffic network knowledge graph, wherein the knowledge graph includes flights, weather, control factors and the interactive relationships between the factors; Step 3-2, flight traffic network situation prediction, using the long short-term memory network LSTM and the Prophet method-based model to predict sector traffic and calculate the evaluation index; Step 3-3, complete the air flight situation assessment by collecting flight plans and using the flight flow network situation prediction results.

9. The method for generating a flight flow allocation scheme based on GAN according to claim 8, characterized in that: The calculation described in step 4 obtains the second air flight flow allocation plan, and the specific calculation method includes: F=αD+βC+γL+μW+ωE Among them, F represents the total delay cost; D is the flight density in the preset time period or area; C represents the congestion level in the preset time period or area; L is the airspace restriction factor, which reflects the degree of unavailability or restriction of a specific area; W represents the impact factor of weather conditions; E represents the adjustment factor for emergency or priority flights; By adjusting the parameters, the flight density D in the preset time period or area, the congestion level C in the preset time period or area, the airspace restriction factor L, the weather condition influencing factor W and the adjustment factor E of the emergency or priority flight are obtained when the total delay cost F is the lowest; taking them as constraints, combined with the preliminary flow allocation plan, the control strategy is determined, and the second air flight flow allocation plan is obtained, which is described as follows: Flight Status i,t , represents the state of flight i at time t, the state includes: waiting state, take-off preparation state, flight state, descent preparation state and landing state; Description of location information i,t , represents the position of flight i at time t, represented by latitude and longitude, sector or route number; Traffic distribution decision point D i,k , as follows: Where k is the sequence number of the flight decision point.

10. The method for generating a flight flow allocation scheme based on GAN according to claim 9, characterized in that: The selections described in step 5 result in the final flight flow allocation solution, including: Step 5-1, establishing a hierarchical structure showing the logical relationship of each level in the air flight network model; Step 5-2, introduce experts and use fuzzy language to evaluate the importance of factor indicators at each level to obtain the fuzzy discriminant matrix; Step 5-3, normalize the fuzzy discriminant matrix and calculate the weight of each factor index in the flow allocation decision; Step 5-4, combine the weights with the corresponding evaluation values, conduct fuzzy comprehensive evaluation, and determine the optimal decision-making plan.

Images