Flight string node sequence determination method

Abstract

The invention relates to the field of computer technology application, in particular to a flight string node sequence determination method, which comprises the following steps of: inputting a fully connected graph constructed on the basis of a current event to be processed into a starting node selection module, and randomly selecting a plurality of nodes as starting nodes; inputting the fully connected graph into a node embedding feature extraction module to extract node embedding features of all nodes in the fully connected graph; for each starting node, acquiring a node sequence set corresponding to each starting node by using a target node acquisition module, and taking the node sequence set as a candidate node sequence set; and obtaining a total resource demand value corresponding to each candidate node sequence set to obtain a plurality of total resource demand values, and taking the candidate node sequence set corresponding to the minimum one of the plurality of total resource demand values as a target node sequence set of the current to-be-processed event. The method can enable the obtained node sequence to be more accurate and accord with an actual application scene, and is high in applicability.

Description

Technical Field

[0001] The present invention relates to the field of computer technology applications, and in particular to a method for determining a flight string node sequence. Background Art

[0002] In some application scenarios, it is necessary to specify the execution object for the event to be processed so that the event to be processed can be executed smoothly. For example, for the flight plan to be processed, it is necessary to assign a suitable aircraft to each flight so that the flight plan can be implemented as smoothly as possible. There are currently many solutions to the scheduling problem of flight plans. The present invention strives to provide another solution that can minimize the use of resources and prevent inherent flight delays. Summary of the invention

[0003] In view of the above technical problems, the technical solution adopted by the present invention is: An embodiment of the present invention provides a method for determining a flight string node sequence. The method is implemented based on a trained node sequence determination model. The trained node sequence determination model includes a starting node selection module, a node embedding feature extraction module, and a target node acquisition module. The method includes the following steps: S100, constructing a fully connected graph based on the current set of pending events; one pending event in the current set of pending events represents a flight, and the fully connected graph is an undirected graph containing self-connections, and the nodes in the fully connected graph represent pending events, and any two nodes are connected by an undirected edge.

[0004] S200, input the fully connected graph into a starting node selection module to randomly select kc nodes from all nodes in the fully connected graph as starting nodes, and input them into the target node acquisition module; kc>1.

[0005] S300, inputting the fully connected graph into a node embedding feature extraction module to extract node embedding features of all nodes in the fully connected graph, and inputting the extracted node embedding features into the target node acquisition module.

[0006] S400, for each starting node, using the target node acquisition module to acquire the node sequence set corresponding to each starting node as a candidate node sequence set, and obtaining k candidate node sequence sets; wherein each candidate node sequence in the candidate node sequence set is executed by the same execution object.

[0007] S500, obtaining the total resource requirement value corresponding to each candidate node sequence set, obtaining k total resource requirement values, and taking the candidate node sequence set corresponding to the smallest of the k total resource requirement values ​​as the target node sequence set for the current event to be processed.

[0008] The present invention has at least the following beneficial effects: The method for determining a flight string node sequence provided by an embodiment of the present invention is implemented based on a trained node sequence determination model, wherein the trained node sequence determination model includes a starting node selection module, a node embedding feature extraction module and a target node acquisition module. The method includes: constructing a fully connected graph based on a current event to be processed; inputting the fully connected graph into the starting node selection module to randomly select k nodes from all nodes in the fully connected graph as starting nodes, and inputting them into the target node acquisition module; inputting the fully connected graph into the node embedding feature extraction module to extract node embedding features of all nodes in the fully connected graph, and inputting the extracted node embedding features into the target node acquisition module; for each starting node, using the target node acquisition module to obtain a node sequence set corresponding to each starting node as a candidate node sequence set, and obtaining k candidate node sequence sets; obtaining a total resource demand value corresponding to each candidate node sequence set, obtaining k total resource demand values, and using the candidate node sequence set corresponding to the smallest of the k total resource demand values ​​as the target node sequence set of the current event to be processed. The present invention obtains node sequences based on a plurality of randomly generated starting nodes, which enables the agent to repeatedly solve the same problem from different entry points, and to be exposed to various problem-solving techniques, thereby obtaining a more accurate node sequence. In addition, by selecting the node sequence with the smallest total resource demand value as the target node sequence, the obtained node sequence can be more in line with the actual application scenario and has strong applicability.

[0009] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 A flowchart of a method for determining a flight string node sequence provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0013] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0014] It should be noted that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the steps as sequential processes, many of the steps can be implemented in parallel, concurrently or simultaneously. In addition, the order of the steps can be rearranged. The process can be terminated when its operation is completed, but it can also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0015] A flight string node sequence determination method provided by an embodiment of the present invention is implemented based on a trained node sequence determination model, wherein the trained node sequence determination model includes a starting node selection module, a node embedding feature extraction module and a target node acquisition module.

[0016] As shown in Figure 1, the method comprises the following steps: S100, constructing a fully connected graph based on the current set of events to be processed.

[0017] In the embodiment of the present invention, the fully connected graph is an undirected graph with self-connection, and the nodes in the fully connected graph represent pending events, and any two nodes are connected by undirected edges. A pending event in the current pending event set represents a flight that needs to be executed, that is, the entire fully connected graph represents a flight plan. The characteristics of the node are flight characteristics. The flight characteristics include the aircraft types allowed to be operated by the flight, the planned departure time, the planned arrival time, the planned departure airport, and the planned arrival airport.

[0018] S200, input the fully connected graph into a starting node selection module to randomly select k nodes from all nodes in the fully connected graph as starting nodes, and input them into the target node acquisition module; kc>1.

[0019] In the embodiment of the present invention, kc may be an empirical value, for example, kc=[N×f], where N is the number of nodes in the fully connected graph, f is a preset coefficient, and [ ] indicates rounding. In an exemplary embodiment, f=0.1.

[0020] The selection of the starting node in the aircraft scheduling problem is very important. If the selected flight node has a later take-off time, the subsequent flights with earlier take-off times cannot be executed. In previous training methods, the first node is always selected by the network. When there are repeated seemingly correct starting nodes, it will guide the model to tend to a specific starting point, which may lead to biased strategies. The present invention draws on the idea of ​​multiple optimal strategy optimization and forces the network to always generate multiple paths, all of which have different starting points. Conceptually, the exploration of multiple optimal strategy optimization tends to guide the agent to repeatedly solve the same problem from different entry points, exposing it to various problem-solving techniques.

[0021] S300, inputting the fully connected graph into a node embedding feature extraction module to extract node embedding features of all nodes in the fully connected graph, and inputting the extracted node embedding features into the target node acquisition module.

[0022] In the embodiment of the present invention, the purpose of the node embedding feature extraction module is to reflect the status of the flight plan of the current event set to be processed, so that the intelligent agent can understand the characteristics of each flight node.

[0023] In an embodiment of the present invention, the node embedding feature extraction module may include: a node embedding module, M encoding modules connected in sequence, wherein each encoding module includes a multi-head self-attention layer, a first splicing processing module, a feedforward neural network layer and a second splicing processing module connected in sequence, wherein the node embedding module is respectively connected to the multi-head self-attention layer and the first splicing processing of the first encoding module, the first splicing processing module of each encoding module is also connected to the corresponding second splicing processing module, the second splicing processing module of the previous encoding module of two adjacent encoding modules is respectively connected to the multi-head self-attention layer and the first splicing processing module of the latter encoding module, the second splicing processing module of the Mth encoding module is connected to the target node acquisition module, and the first splicing processing module and the second splicing processing module are both used to: splice the received input features to obtain the corresponding splicing results, and perform affine transformation batch normalization processing on the splicing results, M>1.

[0024] In the embodiment of the present invention, M may be an empirical value. In an exemplary embodiment, M=6.

[0025] In the embodiment of the present invention, the node embedding module can be specifically used for: First, convert the three-letter codes of the flight's departure and arrival airports into numbers in alphabetical order, such as: because C is the third in the alphabet, C can be converted to 03, so CAN can be converted to 030114; map the flight's required aircraft model to numbers: such as mapping to 0, mapping to 1; convert the flight's scheduled departure time and scheduled arrival time into timestamps.

[0026] Next, all features are normalized to obtain the normalized features.

[0027] Then, each node is projected into a h In the embodiment of the present invention, d h = 128 dimensions. The initial node embedding of each node is h0=W0×X+b0, where W0 and b0 are learnable parameter matrices and X is the normalized feature of the node.

[0028] In an embodiment of the present invention, the number of attention heads of the multi-head attention layer can be set based on actual needs. In an exemplary embodiment, the number of attention heads can be 8.

[0029] In the embodiment of the present invention, each attention head is used to calculate the attention value of each node. The query vector q of each node is W Q h, key vector k=W K h, value vector v=W V h. h is the node embedding of the node, W Q and W K The dimension size is (d k ×d h ) parameter matrix, W V The dimension size is (d v ×d h ) is a parameter matrix. In an exemplary embodiment, d k =d v = 16. Each attention head is used to perform the following operations: Get the compatibility between node n1 and node n2, that is, the attention score. If node n1 and node n2 are adjacent, the attention score S(u1, u2) of node n1 and node n2 = q n1 k n2 T / (d k ) 1 / 2 , if node n1 and node n2 are not adjacent, the attention scores of node n1 and node n2 are negative infinity. n1 is the query vector of node n1, k n2 is the key vector of node n2. The values ​​of n1 and n2 range from 1 to N, where N is the number of nodes in the fully connected graph.

[0030] Get the attention weights W of nodes n1 and n2 n1n2 =e S（u1，u2） / ∑ N n3=1 e S（u1，u3） , S(u1, u3) is the attention score of node u1 and node u3, the value of u3 ranges from 1 to N, and e is a natural number.

[0031] Get the attention value SP of node n1 n1 =∑ N n2=1 W n1n2 ×v n2 ;v n2 is the value vector of node n2.

[0032] In the embodiment of the present invention, the dimension size is (d v ×d h ) parameter matrix W m Project the attention values ​​of each node obtained by all attention heads back to d h dimensional vector. That is, the multi-head attention value SM obtained by node n1 through the multi-head attention layer n1 =∑ M1 n4=1 W m ×SP n1n4 , n4 ranges from 1 to M1. M1 is the number of attention heads.

[0033] In the embodiment of the present invention, the specific working principle of the splicing module may be the prior art. h The affine parameter matrix of .

[0034] In the embodiment of the present invention, the feedforward neural network layer can use the dimension d f =512 hidden sublayer and ReLu activation function, and the specific calculation process may be the prior art.

[0035] S400, for each starting node, use the target node acquisition module to obtain the node sequence set corresponding to each starting node as the candidate node sequence set, obtain k candidate node sequence sets and send them to the target node sequence acquisition module; wherein each candidate node sequence in the candidate node sequence set is executed by the same execution object, and the execution object is the aircraft.

[0036] In an embodiment of the present invention, the target node acquisition module is used to parse the relationship between flight nodes through the flight node features extracted by the node embedding feature extraction module, and compile and output the flight string sequence, which may include a candidate node selection unit, a multi-head attention layer, a single-head attention layer and a target node selection unit. The candidate node selection unit is used to screen the flight nodes, and the attention layer is used to calculate the probability of the flight nodes and select the next node of the flight sequence. At the same time, after the node selection, the decoder needs to update the context embedding to ensure that the new graph structure information is captured in time.

[0037] Furthermore, S400 specifically includes: S401, set the starting node counter r=1.

[0038] S402, if r≤k, set the execution object counter g=1, and execute S403; if r>k, obtain k candidate node sequence sets, and execute S500.

[0039] S403, if the current node set is not empty, execute S404, otherwise, obtain the candidate node sequence set corresponding to the r-th starting node, set r=r+1, and execute S402; the initial value of the current node set is the node set formed by all nodes in the fully connected graph.

[0040] S404, using the candidate node selection unit to select nodes that meet the preset constraints between the currently selected nodes corresponding to the g-th execution object from the current node set, if the nodes that meet the preset constraints between the currently selected nodes corresponding to the g-th execution object are obtained, the nodes that meet the preset constraints between the currently selected nodes corresponding to the g-th execution object are used as the current candidate nodes of the g-th execution object, and S405 is executed; if the nodes that meet the preset constraints between the currently selected nodes corresponding to the g-th execution object are not obtained, S409 is executed; wherein, the initial value of g is 1, and the initial value of the currently selected node corresponding to the first execution object is the r-th starting node.

[0041] In the embodiment of the present invention, the preset constraint condition satisfies the following conditions: If M j =1, indicating that the jth node in the current node set is a candidate node for the currently selected node. j = 0, indicating that the jth node in the current node set is not a candidate node for the currently selected node, where M j is the total constraint value corresponding to the jth node, M j =C j ⊙C j E ⊙C jL , C j is the transit time constraint value corresponding to the jth node, C j E is the model constraint value corresponding to the jth node, C j E is the airport connection constraint value corresponding to the jth node, C j The following conditions are met: If the slack time between the currently selected node and the jth node is greater than or equal to 0, C j =1, otherwise, C j =0; C j E The following conditions are met: If the aircraft model corresponding to the currently selected node is applicable to the jth node, C j E =1, otherwise, C j E =0; C j L The following conditions are met: If the departure airport corresponding to the jth node is the airport corresponding to the currently selected node, C j L =1, otherwise, C j L =0; ⊙ represents the XOR operation, j ranges from 1 to Q, and Q is the number of nodes in the current node set.

[0042] In airlines, aircraft scheduling is to arrange a series of flights for each aircraft based on the flight schedule determined by market demand, under the conditions of time and space connection, i.e., flight string. Slack time refers to the extra time reserved in the actual operation of a flight to cope with possible delays and unexpected situations. The slack time between two flights is the difference between the planned stopover time difference of the two flights and the minimum stopover time. If there is not enough slack time between two flights, the delay of the preceding flight will affect the subsequent flight. Assuming that the minimum stopover time is 30 minutes, the historical average delay time of flight Fa is 50 minutes, which is greater than the slack time of flight Fa and flight Fc, which is 30 minutes. This will increase the probability of delays in the subsequent flight sequence connected by flight Fa and flight Fc. The historical average delay time of flight Fb is 5 minutes, which is less than the slack time of flight Fb and flight Fd, which is 30 minutes. It can be considered that the slack time of Fb and Fd can buffer the delay of the preceding flight Fb. In a flight sequence with good robustness (Fb will connect with flight Fd, and flight Fb will connect with flight Fc), the robustness of the flight plan can be improved by optimizing the flight sequence.

[0043] S405, based on the average value of the node embedding features of all nodes in the fully connected graph, the node embedding features of the currently selected node corresponding to the g-th execution object and the number of currently used execution objects, the multi-head attention layer is used to obtain the contextual attention features corresponding to the fully connected graph as the current contextual attention features corresponding to the g-th execution object.

[0044] S406, based on the current context attention feature corresponding to the g-th execution object, using the single-head attention layer to obtain the node weight of each node in the fully connected graph as the current node weight corresponding to the g-th execution object.

[0045] S407, based on the current node weight corresponding to the g-th execution object, use the target node selection unit to obtain the target node from the current candidate nodes of the g-th execution object as the current target node corresponding to the g-th execution object; wherein the current target node corresponding to the g-th execution object is the node corresponding to the maximum node weight of the current node weight of the g-th execution object.

[0046] S408, take the current target node corresponding to the g-th execution object as the currently selected node corresponding to the g-th execution object, and add it to the currently selected node set corresponding to the g-th execution object, and delete the current target node corresponding to the g-th execution object from the current node set, and execute S403; the initial value of the currently selected node set corresponding to the g-th execution object is empty.

[0047] S409, obtain the candidate node sequence corresponding to the g-th execution object based on the currently selected node set corresponding to the g-th execution object, set g=g+1, and randomly select a node from the current node set as the initial value of the currently selected node corresponding to the g+1-th execution object, and execute S403.

[0048] Further, in S405, using the multi-head attention layer to obtain the contextual attention features corresponding to the fully connected graph specifically includes: S4051, obtaining the current splicing feature FP; wherein FP satisfies the following conditions: if time step t=1, FP=[F avg , z, P t ], if t>1, FP=[F avg , F t , P t ],F avg is the average value of the node embedding features of all nodes, z is the set input placeholder, P t is the number of currently used execution objects, F t Embeds a feature for the node of the currently selected node.

[0049] S4052, using each attention head of the multi-head attention layer to obtain the attention score vector between the current context embedding feature and the node embedding features of all nodes in the fully connected graph; wherein the query vector q corresponding to the current context embedding feature c The following conditions are met: c =W Q ×FP,W Q is the query weight matrix of the attention head, and the key vector k corresponding to the a-th node in the fully connected graph a =W K ×h a , the value vector v corresponding to the a-th node a =W V ×h a , W K is the key weight matrix of the attention head, h a is the node embedding feature of the a-th node, W V is the value weight matrix of the attention head, where if the a-th node is not a candidate node, the attention score between the current context embedding feature and the node embedding feature of the a-th node is set to negative infinity; if the a-th node is not a candidate node, the attention score between the current context embedding feature and the node embedding feature of the a-th node is set to S(c, a)=q c k a T / (d k ) 1 / 2 , d k is the dimension of the key vector, a ranges from 1 to N, where N is the number of nodes in the fully connected graph.

[0050] S4053, based on the current context embedding feature obtained by each attention head and the attention score vector between all nodes in the fully connected graph, obtain the attention value Av=∑ N a=1 (e S（c，a） / ∑ N b1=1 e S（c ，b1） )×v a , S(c, b1) is the attention score between the current context embedding feature and the node embedding feature of the b1th node in the fully connected graph, b1 ranges from 1 to N, and e is a natural number.

[0051] S4064, based on the attention value of the current context embedded feature obtained by all attention heads, obtain the output result of the multi-head attention as the context attention feature.

[0052] The specific method of obtaining contextual attention features can refer to the method of obtaining the multi-head attention value mentioned above.

[0053] Further, in S406, the node weight W of the ath node in the fully connected graph is a The following conditions must be met: W a =(e C（a） ) / ∑ N b1=1 e C（b1） ; C(a) is the attention score between the current context embedding feature corresponding to the single-head attention layer and the node embedding feature of the a-th node, where C(a) satisfies the following conditions: if the a-th node is a candidate node, C(a)=G•tanh(q s k sa T / (d k ) 1 / 2 ); where • represents the dot product, G is the preset value, tanh() is the hyperbolic tangent function, q s is the current context embedding feature corresponding to the single-head attention layer, q s =W s q A c , W s q is the query weight matrix of the single-head attention layer, A c is the contextual attention feature, k sa is the key vector corresponding to the a-th node obtained by the single-head attention layer, k sa =W s k h a , W s k is the key weight matrix of the single-headed attention layer. If the a-th node is not a candidate node, C(a) is equal to negative infinity; C(b1) is the attention score between the current context embedding feature corresponding to the single-headed attention layer and the node embedding feature of the a-th node.

[0054] In the embodiment of the present invention, the output range of C(a) tanh and G is converted to the interval [-G, G], and a larger value allows the model to learn more complex feature representations. In an exemplary embodiment, G=10. S500, obtain the total resource requirement value corresponding to each candidate node sequence set, obtain k total resource requirement values, and use the candidate node sequence set corresponding to the smallest of the k total resource requirement values ​​as the target node sequence set of the current event to be processed.

[0055] In the embodiment of the present invention, the total resource demand value D corresponding to the rth starting node r The following conditions must be met: D r =∑ f（r）i=1 ∑ z（r） u=1 RC iu × iu +∑ z（r） u=1 P u ×T u ×I u ; Among them, RC iu The resources required for the i-th execution object to execute the pending event corresponding to the u-th node are specifically the running cost, i ranges from 1 to f(r), f(r) is the number of execution objects corresponding to the r-th starting node, u ranges from 1 to z(r), z(r) is the number of nodes in the candidate node sequence set corresponding to the r-th starting node; x iu is a variable. If the i-th execution object executes the pending event corresponding to the u-th node, x iu =1, otherwise, x iu =0;P u The delay coefficient corresponding to the u-th node is iu =1,P u =Q u-front +(1-Q u-front )×P u-front , Q u-front is the delay coefficient determined based on the relaxation time between the predecessor node of the jth node and the jth node, P u-front is the delay coefficient of the predecessor node of the jth node, T u is the historical average delay time corresponding to the u-th node, I u is the resource value required for the delay of the u-th node, i.e., the delay cost.

[0056] It is known to those skilled in the art that if x iu = 0, indicating that the i-th execution object has not executed the pending event corresponding to the u-th node, so P u =0.

[0057] Furthermore, in an embodiment of the present invention, the trained node sequence determination model is obtained by the following steps: S10, constructing an initial node sequence determination model, that is, initializing the parameters of the node sequence determination model.

[0058] S20, dividing the sample data set into B batches of training data sets, wherein the sample data set consists of a plurality of events to be processed within a set time period, namely, flight plans.

[0059] In the embodiment of the present invention, B may be an empirical value.

[0060] S30, based on the current batch of training data sets, construct a corresponding fully connected graph, and input it to the current node sequence determination model to obtain a target node sequence set corresponding to the current batch of training data sets; the initial value of the current node sequence determination model is the initial node sequence determination model.

[0061] The specific process of obtaining the target node sequence set corresponding to the current batch of training data sets may refer to the contents defined in the aforementioned S200 to S500.

[0062] S40, obtaining the gradient of the total resource demand value corresponding to the candidate node sequence set corresponding to the current batch of training data sets as the current gradient, and updating the parameters of the current node sequence determination model based on the current gradient; wherein the current gradient ▽J satisfies the following conditions: ▽J≈(1 / kc)∑ kc v=1 (D v -D avg )▽log(W v1 ×……×W vb2 ×……×W vz（v） ), ▽ represents the gradient, D v is the total resource demand value corresponding to the vth starting node corresponding to the current batch of training data sets, D avg is the average value of the total resource demand values ​​corresponding to the k starting nodes corresponding to the current batch of training data sets. The value of v ranges from 1 to k. W vb2 is the node weight of the b2th node in the candidate node sequence set corresponding to the vth starting node, b2 ranges from 1 to z(v), and z(v) is the number of nodes in the candidate node sequence set corresponding to the vth starting node.

[0063] S50, determine whether the current node sequence determination model meets the preset model end condition. If so, determine the current node sequence as the trained node sequence determination model. Otherwise, use the next batch of training data sets as the current batch of training data sets and execute S40.

[0064] In an embodiment of the present invention, the preset model end condition can be set based on actual needs, for example, the number of iterations reaches a preset number or the total resource demand value is at a minimum value.

[0065] In summary, the flight string node sequence determination method provided by the embodiment of the present invention has at least the following advantages: 1. Improve the accuracy of node sequence Traditional methods usually only start from a fixed starting point to find the optimal solution, which may cause the algorithm to fall into a local optimum. However, the present invention can increase the diversity of the search space and help find the global optimal solution by randomly selecting multiple starting points and generating candidate node sequence sets based on these starting nodes. The present invention enables the intelligent agent to repeatedly solve the same problem from different entry points, thereby being exposed to various problem-solving techniques. This method of multiple optimal strategy optimization significantly improves the accuracy of the final obtained node sequence.

[0066] 2. Enhance the applicability and robustness of the model The present invention selects the node sequence with the smallest total resource requirement as the target node sequence, which not only considers the direct resources required to execute the event, but also the potential costs caused by delays, thereby ensuring the feasibility and economy of the generated node sequence in actual operation. This strategy makes the obtained node sequence more in line with the actual application scenario and enhances the applicability and robustness of the model.

[0067] 3. Optimize flight schedules and reduce costs The present invention generates a more reasonable flight sequence by comprehensively considering multiple factors such as the time and space connection between flights, aircraft type constraints, airport connections, etc. This optimized flight plan reduces the number of unnecessary aircraft and aircraft type assignment costs, while reducing the potential delay costs, thereby reducing the total cost.

[0068] 4. Reduce flight delays When generating a node sequence, the present invention fully considers factors such as the slack time between flights and the historical average delay time, and reduces the delay propagation between flights by optimizing the arrangement of flight strings, thereby reducing the total delay time.

[0069] 5. Improve solution efficiency The present invention adopts a method based on deep learning, and can quickly generate an optimized solution after a given input by training a node sequence to determine the model. Compared with traditional iterative algorithms, the present invention avoids the process of re-iteration when solving new instances, thereby improving the solution efficiency.

[0070] An embodiment of the present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the method described in the embodiment of the present invention.

[0071] The embodiment of the present invention further provides a computer-readable storage medium storing computer-executable instructions, wherein the computer instructions are used to execute the method described in the embodiment of the present invention.

[0072] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.

[0073] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for determining a flight string node sequence, characterized in that: The method is implemented based on a trained node sequence determination model, wherein the trained node sequence determination model includes a starting node selection module, a node embedding feature extraction module, and a target node acquisition module; the method includes the following steps: S100, constructing a fully connected graph based on the current set of pending events; one pending event in the current set of pending events represents one flight, and the fully connected graph is an undirected graph containing self-connections, and nodes in the fully connected graph represent pending events, and any two nodes are connected by an undirected edge; S200, inputting the fully connected graph into a starting node selection module, randomly selecting kc nodes from all nodes in the fully connected graph as starting nodes, and inputting them into the target node acquisition module; kc>1; S300, inputting the fully connected graph into a node embedding feature extraction module to extract node embedding features of all nodes in the fully connected graph, and inputting the extracted node embedding features into the target node acquisition module; S400, for each starting node, using the target node acquisition module to acquire a node sequence set corresponding to each starting node as a candidate node sequence set, and obtaining k candidate node sequence sets; wherein each candidate node sequence in the candidate node sequence set is executed by the same execution object; S500, obtaining the total resource requirement value corresponding to each candidate node sequence set, obtaining k total resource requirement values, and taking the candidate node sequence set corresponding to the smallest of the k total resource requirement values ​​as the target node sequence set for the current event to be processed.

2. The method according to claim 1, characterized in that The target node acquisition module includes a candidate node selection unit, a multi-head attention layer, a single-head attention layer and a target node selection unit; S400 specifically includes: S401, set the starting node counter r=1; S402, if r≤k, set the execution object counter g=1, and execute S403; if r>k, obtain k candidate node sequence sets, and execute S500; S403, if the current node set is not empty, execute S404, otherwise, obtain the candidate node sequence set corresponding to the r-th starting node, set r=r+1, and execute S402; the initial value of the current node set is the node set formed by all nodes in the fully connected graph; S404, using the candidate node selection unit to select nodes satisfying preset constraints between currently selected nodes corresponding to the g-th execution object from the current node set, if nodes satisfying preset constraints between currently selected nodes corresponding to the g-th execution object are obtained, the obtained nodes satisfying preset constraints between currently selected nodes corresponding to the g-th execution object are used as current candidate nodes of the g-th execution object, and S405 is executed; if nodes satisfying preset constraints between currently selected nodes corresponding to the g-th execution object are not obtained, S409 is executed; wherein, the initial value of g is 1, and the initial value of the currently selected node corresponding to the first execution object is the r-th starting node; S405, based on the average value of the node embedding features of all nodes in the fully connected graph, the node embedding features of the currently selected node corresponding to the g-th execution object, and the number of currently used execution objects, using the multi-head attention layer to obtain the contextual attention feature corresponding to the fully connected graph as the current contextual attention feature corresponding to the g-th execution object; S406, based on the current context attention feature corresponding to the g-th execution object, using the single-head attention layer to obtain the node weight of each node in the fully connected graph as the current node weight corresponding to the g-th execution object; S407, based on the current node weight corresponding to the g-th execution object, using the target node selection unit to obtain a target node from the current candidate nodes of the g-th execution object as the current target node corresponding to the g-th execution object; wherein the current target node corresponding to the g-th execution object is a node corresponding to the maximum node weight among the current node weights of the g-th execution object; S408, taking the current target node corresponding to the g-th execution object as the currently selected node corresponding to the g-th execution object, and adding it to the currently selected node set corresponding to the g-th execution object, and deleting the current target node corresponding to the g-th execution object from the current node set, and executing S403; the initial value of the currently selected node set corresponding to the g-th execution object is empty; S409, obtain the candidate node sequence corresponding to the g-th execution object based on the currently selected node set corresponding to the g-th execution object, set g=g+1, and randomly select a node from the current node set as the initial value of the currently selected node corresponding to the g+1-th execution object, and execute S403.

3. The method according to claim 2, characterized in that The preset constraints meet the following conditions: If M j =1, indicating that the jth node in the current node set is a candidate node for the currently selected node. j = 0, indicating that the jth node in the current node set is not a candidate node for the currently selected node, where M j is the total constraint value corresponding to the jth node, M j =C j ⊙C j E ⊙C j L , C j is the transit time constraint value corresponding to the jth node, C j E is the model constraint value corresponding to the jth node, C j E is the airport connection constraint value corresponding to the jth node, C j The following conditions are met: If the slack time between the currently selected node and the jth node is greater than or equal to 0, C j =1, otherwise, C j =0; C j E The following conditions are met: If the aircraft model corresponding to the currently selected node is applicable to the jth node, C j E =1, otherwise, C j E =0; C j L The following conditions are met: If the departure airport corresponding to the jth node is the airport corresponding to the currently selected node, C j L =1, otherwise, C j L =0; ⊙ represents the XOR operation, j ranges from 1 to Q, and Q is the number of nodes in the current node set.

4. The method according to claim 3, characterized in that In S405, using the multi-head attention layer to obtain the contextual attention features corresponding to the fully connected graph specifically includes: S4051, obtaining the current splicing feature FP; wherein FP satisfies the following conditions: if time step t=1, FP=[F avg , z, P t ], if t>1, FP=[F avg , F t , P t ], F avg is the average value of the node embedding features of all nodes, z is the set input placeholder, P t is the number of currently used execution objects, F t Embed features for nodes of currently selected node; S4052, using each attention head of the multi-head attention layer to obtain an attention score vector between the current context embedding feature and the node embedding features of all nodes in the fully connected graph; S4053, based on the current context embedding feature obtained by each attention head and the attention score vector between all nodes in the fully connected graph, obtain the attention value Av=∑ N a=1 (e S（c，a） / ∑ N b1=1 e S（c，b1） )×v a , S(c, b1) is the attention score between the current context embedding feature and the node embedding feature of the b1th node in the fully connected graph, b1 ranges from 1 to N, and e is a natural number; S4054, based on the attention value of the current context embedded feature obtained by all attention heads, obtain the output result of the multi-head attention as the context attention feature.

5. The method according to claim 4, characterized in that The query vector q corresponding to the current context embedding feature c The following conditions are met: c =W Q ×FP,W Q is the query weight matrix of the attention head, and the key vector k corresponding to the a-th node in the fully connected graph a =W K ×h a , the value vector v corresponding to the a-th node a =W V ×h a , W K is the key weight matrix of the attention head, h a is the node embedding feature of the a-th node, W V is the value weight matrix of the attention head, where if the a-th node is not a candidate node, the attention score between the current context embedding feature and the node embedding feature of the a-th node is set to negative infinity; if the a-th node is not a candidate node, the attention score between the current context embedding feature and the node embedding feature of the a-th node is set to S(c, a)=q c k a T / (d k ) 1 / 2 , d k is the dimension of the key vector, a ranges from 1 to N, where N is the number of nodes in the fully connected graph.

6. The method according to claim 4, characterized in that In S406, the node weight W of the ath node in the fully connected graph a The following conditions must be met: W a =(e C（a） ) / ∑ N b1=1 e C（b1） ; C(a) is the attention score between the current context embedding feature corresponding to the single-head attention layer and the node embedding feature of the a-th node, where C(a) satisfies the following conditions: if the a-th node is a candidate node, C(a)=G•tanh(q s k sa T / (d k ) 1 / 2 ); where • represents the dot product, G is the preset value, tanh() is the hyperbolic tangent function, q s is the current context embedding feature corresponding to the single-head attention layer, q s =W s q A c , W s q is the query weight matrix of the single-head attention layer, A c is the contextual attention feature, k sa is the key vector corresponding to the a-th node obtained by the single-head attention layer, k sa =W s k h a , W s k is the key weight matrix of the single-headed attention layer. If the a-th node is not a candidate node, C(a) is equal to negative infinity; C(b1) is the attention score between the current context embedding feature corresponding to the single-headed attention layer and the node embedding feature of the a-th node.

7. The method according to claim 1, characterized in that The total resource demand value D corresponding to the rth starting node r The following conditions must be met: D r =∑ f（r） i=1 ∑ z（r） u=1 RC iu ×x iu +∑ z（r） u=1 P u ×T u ×I u ; Among them, RC iu The resources required for the i-th execution object to execute the pending event corresponding to the u-th node, i ranges from 1 to f(r), f(r) is the number of execution objects corresponding to the r-th starting node, u ranges from 1 to z(r), z(r) is the number of nodes in the candidate node sequence set corresponding to the r-th starting node; x iu is a variable. If the i-th execution object executes the pending event corresponding to the u-th node, x iu =1, otherwise, x iu =0;P u The delay coefficient corresponding to the u-th node is iu =1,P u =Q u-front +(1-Q u-front )×P u-front , Q u-front is the delay coefficient determined based on the relaxation time between the predecessor node of the jth node and the jth node, P u-front is the delay coefficient of the predecessor node of the jth node, T u is the historical average delay time corresponding to the u-th node, I u is the resource value required for the delay of the u-th node.

8. The method according to claim 1, characterized in that The node embedding feature extraction module includes: a node embedding module, M encoding modules connected in sequence, wherein each encoding module includes a multi-head self-attention layer, a first splicing processing module, a feedforward neural network layer and a second splicing processing module connected in sequence, wherein the node embedding module is respectively connected to the multi-head self-attention layer and the first splicing processing of the first encoding module, the first splicing processing module of each encoding module is also connected to the corresponding second splicing processing module, the second splicing processing module of the previous encoding module of two adjacent encoding modules is respectively connected to the multi-head self-attention layer and the first splicing processing module of the next encoding module, the second splicing processing module of the Mth encoding module is connected to the target node acquisition module, and the first splicing processing module and the second splicing processing module are both used to: splice the received input features to obtain the corresponding splicing results, and perform affine transformation batch normalization processing on the splicing results, M>1.

9. The method according to claim 6, characterized in that The trained node sequence determination model is obtained by the following steps: S10, constructing an initial node sequence determination model; S20, dividing the sample data set into B batches of training data sets, wherein the sample data set consists of a plurality of events to be processed within a set time period; S30, based on the current batch of training data sets, construct a corresponding fully connected graph, and input it to the current node sequence determination model to obtain a target node sequence set corresponding to the current batch of training data sets; the initial value of the current node sequence determination model is the initial node sequence determination model; S40, obtaining the gradient of the total resource demand value corresponding to the candidate node sequence set corresponding to the current batch of training data sets as the current gradient, and updating the parameters of the current node sequence determination model based on the current gradient; S50, determine whether the current node sequence determination model meets the preset model end condition. If so, determine the current node sequence as the trained node sequence determination model. Otherwise, use the next batch of training data sets as the current batch of training data sets and execute S40.

10. The method according to claim 9, characterized in that in, The current gradient ▽J satisfies the following conditions: ▽J≈(1 / kc)∑ kc v=1 (D v -D avg )▽log(W v1 ×……×W vb2 ×……×W vz（v） ), ▽ represents the gradient, D v is the total resource demand value corresponding to the vth starting node corresponding to the current batch of training data sets, D avg is the average value of the total resource demand values ​​corresponding to the k starting nodes corresponding to the current batch of training data sets. The value of v ranges from 1 to k. W vb2 is the node weight of the b2th node in the candidate node sequence set corresponding to the vth starting node, b2 ranges from 1 to z(v), and z(v) is the number of nodes in the candidate node sequence set corresponding to the vth starting node.

Images