Large-area flight recovery energy efficiency optimization method integrating deep learning and genetic algorithm

Abstract

The invention provides a large-area flight recovery energy efficiency optimization method fusing deep learning and a genetic algorithm, belongs to the technical field of intelligent scheduling and optimization, and solves the problem of multi-target efficient recovery after large-scale flight delay. According to the technical scheme, the method comprises the following steps: firstly, constructing standardized scheduling input characteristics and economic loss factors by using real flight data; secondly, designing a multi-objective function, and comprehensively considering total delay, economic loss, the number of serious delay times and scheduling fairness; then, generating a scheduling sample based on an evolutionary strategy and constructing a training set; thirdly, training a deep neural network agent model to replace a high-cost evaluation function; and finally, combining an agent model and an elite screening mechanism, and quickly searching an optimal scheduling scheme in evolutionary optimization. The method has the advantages that large-scale flight recovery scheduling tasks are efficiently completed, global efficiency and local fairness are both considered, a command department is helped to rapidly formulate a recovery strategy, and the method has high practical application value.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling and optimization technology, specifically to a method for optimizing energy efficiency in large-scale flight recovery that integrates deep learning and genetic algorithms. Technical Background

[0002] Genetic algorithms (GA) play a crucial role in flight scheduling optimization, delay propagation suppression, resource reallocation, passenger travel assurance, and emergency response strategy development. By simulating the natural selection process, GA automatically evolves near-optimal scheduling schemes, which helps to quickly restore the stability of the flight network after large-scale flight disruptions. However, traditional centralized data integration and global solution methods often suffer from huge computational costs and slow convergence speeds, limiting their practical application in dynamic environments.

[0003] Deep learning (DL), as a powerful representation learning tool, effectively complements the shortcomings of genetic algorithms in large-scale optimization tasks. With the support of deep learning models, algorithms can automatically extract complex high-dimensional feature patterns and model latent patterns in the search space, thereby improving convergence speed and solution diversity.

[0004] Currently, many methods based on deep learning or genetic algorithms have been proposed for complex scheduling optimization, intelligent resource allocation, and dynamic recovery strategy formulation, such as flight schedule fine-tuning, crew scheduling optimization, and passenger itinerary reconstruction, aiming to improve operational efficiency and reduce delay risks. However, most of these methods are independent and have not yet achieved an organic combination of deep learning and genetic algorithms, nor do they have a systematic application in large-scale flight delay recovery scenarios. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method for optimizing the energy efficiency of large-scale flight recovery by integrating deep learning and genetic algorithms. In the model preprocessing stage, standardized scheduling features and corresponding economic loss indicators are generated based on real flight data, providing an accurate and consistent input foundation for subsequent optimization models and improving the scientific rigor and interpretability of flight recovery decisions. Four objective functions are designed to comprehensively optimize delay time, economic loss, and fairness, and the solutions are obtained by integrating deep learning and genetic algorithms, demonstrating strong application value for large-scale flight recovery optimization.

[0006] The inventive concept of this invention is as follows: First, standardized scheduling input features and economic loss factors are constructed using real flight data; then, a multi-objective function is designed to comprehensively consider total delay, economic loss, number of severe delays, and scheduling fairness; next, scheduling samples are generated based on an evolutionary strategy and a training set is constructed; third, a deep neural network surrogate model is trained to replace the high-cost evaluation function; finally, the optimal scheduling scheme is quickly searched in evolutionary optimization by combining the surrogate model with an elite selection mechanism.

[0007] This invention is achieved through the following measures: a method for optimizing energy efficiency in large-scale flight recovery that integrates deep learning and genetic algorithms, comprising the following steps:

[0008] S10: Using real flight data containing planned take-off and landing times, actual operational status, and flight attribute information, filter and clean relevant flights at the target airport, removing canceled and diverted data while retaining take-off and landing related fields; extract attribute information for each flight, such as planned take-off and landing times, flight distance, airline, month, and take-off and arrival airports; calculate a comprehensive priority index based on predefined airline priority, flight distance standardization functions, and peak period identification rules; and construct three types of economic weighting factors: airline losses, airport losses, and passenger losses.

[0009] S20: Construct a multi-objective optimization model with the goals of minimizing total delay time, minimizing total economic loss, minimizing the number of flights delayed by more than 30 minutes, and maximizing scheduling fairness; design nonlinear inequality constraints, allowing a maximum of 3 flights to take off and land in each time slice and ensuring that the total number of takeoffs and landings during special peak periods does not exceed a threshold.

[0010] S30: Based on the initial population generation strategy of genetic algorithm, construct a large batch of random flight time allocation schemes, and call the real objective function and constraint function to evaluate each scheme; construct the evaluation results into sample pairs with flight time slice vectors as input and multiple objective values ​​as output; introduce a preprocessing mechanism of logarithmic transformation and standard deviation normalization;

[0011] S40: Construct a deep neural network structure containing multiple fully connected layers, batch normalization layers, and nonlinear activation functions; adopt a multi-objective regression output layer to simultaneously predict four objective values: delay, loss, number of flights exceeding limits, and fairness; use the Adam optimizer with a learning rate decay scheduling strategy for training, and set appropriate batch size and maximum training epochs.

[0012] S50: During the optimization process, a surrogate model is used to replace the real objective function for rapid evaluation; each generation of optimization executes an elite selection mechanism to perform real evaluation and update on several individuals with the best surrogate evaluation values; through iterative optimization, a set of flight recovery scheduling schemes with optimal comprehensive performance under multi-objective equilibrium is finally output.

[0013] Further, step S10 includes the following steps:

[0014] S11: Filter valid flights from the original flight dataset, removing canceled and diverted flights. The filtering criteria are defined as follows:

[0015] H cancel =0 AND H diverted =0 (1)

[0016] Among them, H cancel H indicates that the flight has been cancelled. diverted This indicates that the flight has made an emergency landing.

[0017] S12: For each valid flight that successfully completes a flight mission, F i Extract its key attribute information to construct attribute vector A. i =[f cd ,f ca ,f dis ,f car ,f mon ,f ori ,f des ], where f cd For the planned takeoff time, f ca For the planned arrival time, f dis f represents the flight distance. car For airline code, f mon For the month of flight, f ori For the departure airport code, f des For destination airport code;

[0018] S13: Predefine the airline priority factor P based on the airline's service level. ci Max-min normalization is used to eliminate the influence of flight distance dimensions, and the flight distance standardization factor P is used. di The calculation formula is:

[0019]

[0020] Where D max and D min D represents the maximum or minimum distance in a flight. i Let P represent the distance to the i-th flight. For the peak period identification factor, based on airport operational patterns, peak periods are identified and assigned a weight of 1, while off-peak periods are assigned a weight of 0.5. The peak period identification factor is defined as P. ti Taking into account the three factors, calculate the comprehensive result.

[0021] Priority indicators:

[0022] P i priority =0.5 + 0.3 × P ci +0.1×P di +0.1×P ti (3)

[0023] Airlines loss coefficient The calculation formula is:

[0024]

[0025] in For airline rating adjustment factors, Airport loss coefficient for the i-th flight in the quarter The calculation formula is:

[0026]

[0027] in This represents the size factor and passenger loss coefficient of the i-th airport. This mainly includes time cost, missed flight losses, and decreased satisfaction. The calculation formula is as follows:

[0028]

[0029] Further, step S20 includes the following steps:

[0030] S21: The first objective is to minimize the total delay time. The objective function for the total delay time is defined as the weighted sum of the delay times of all flights, and the calculation formula is as follows:

[0031]

[0032] Where w i w is the weighting coefficient for the i-th flight. i =P i priority f delay (x) represents the delay time of the i-th flight. This objective function ensures that only flights with delays exceeding 15 minutes are included in the calculation. delay (x)>15) is an indicator function, which is 1 when the condition is true, and 0 otherwise;

[0033] S22: The second objective is to minimize the total economic loss. The objective function for total economic loss considers the losses from airlines, airports, and passengers, and the calculation formula is as follows:

[0034]

[0035] Where C delay C is the conversion factor for delay loss. delay= 1 / 10000;

[0036] S23: The third objective is to minimize the number of flights delayed by more than 30 minutes. The objective function for the number of severely delayed flights is defined as follows:

[0037]

[0038] Where I(f) delay (x) > 30) is an indicator function, with a value of 1 when the delay exceeds 30 minutes;

[0039] S24: The fourth objective is to maximize scheduling fairness, which is measured by minimizing the standard deviation of delay time. The formula is as follows:

[0040]

[0041] Where μ is the average delay time, calculated using the following formula:

[0042]

[0043] S25: The first constraint is to constrain the capacity of each time slice to ensure that the airport's processing capacity is not exceeded. The formula for calculating the constraint is:

[0044]

[0045] Where I(x) i =j) is an indicator function, with a value of 1 when flight i is assigned to time slice j, R capslot =3, indicating that the capacity limit for each time slot is 3 aircraft. The second constraint is the total capacity constraint for peak periods, ensuring the normal operation of flights during peak hours. Its calculation formula is:

[0046]

[0047] in, and These are the start and end time slices for a specific period, C. special This refers to the capacity threshold for specific time periods.

[0048] Further, step S30 includes the following steps:

[0049] S31: The population representation of the genetic algorithm is a set of flight time allocation schemes, where each individual represents a complete scheduling scheme. The required number of samples is determined based on the problem complexity, and the calculation formula is as follows:

[0050] N num =max(1000,10×N×log(N)) (14)

[0051] Where N represents the number of flights;

[0052] S32: For each individual P i The actual objective function is called for evaluation, and the calculation formula is as follows:

[0053] F i f =[f1(P i ),f2(P i ),f3(P i ),f4(P i )] T (15)

[0054] The degree of constraint violation is determined by the constraint vector F. i g The calculation formula is as follows:

[0055] F i g =[g1(P i ),g2(P i )] T (16)

[0056] S33: Convert the time slice allocation for each individual into a feature vector X i The input feature vector is constructed, and the evaluation result is used to construct the target vector Y. i The output target vector is represented by the given vector, and the final training sample set is:

[0057] N data ={(X i ,Y i | i = 1, 2, ..., N num} (17)

[0058] S34: Perform a logarithmic transformation on the target vector, the calculation formula is as follows:

[0059] Y i log =[log(1+f1(P)] i )),log(1+f2(P i )),log(1+f3(P i )),log(1+f4(P i ))] T (18)

[0060] Calculate the mean and standard deviation after logarithmic transformation, and then perform Z-score standardization on them.

[0061] Further, step S40 includes the following steps:

[0062] S41: Construct four fully connected layers. The mathematical expression for the first fully connected layer is:

[0063] z1 = W1 × x norm +b1 (19)

[0064] Where W1 is the weight matrix, b1 is the bias vector, z1 is the linear transformation output, a batch normalization layer is constructed, batch statistical calculation and standardization transformation are performed, and finally the learnable parameter transformation is performed. The nonlinear activation function is the ReLU activation function.

[0065] S42: Multi-objective regression predicts four target values ​​simultaneously. The calculation formula is as follows:

[0066] Predict the total delay time, total economic loss, number of flights delayed by more than 30 minutes, and scheduling fairness indicators.

[0067] Further, step S50 includes the following steps:

[0068] S51: During the optimization process, a surrogate model is used to quickly evaluate the true objective function. The surrogate model is used to predict and then destandardizes to obtain the true objective value. The calculation formula is as follows:

[0069]

[0070] Where F net (x) represents the trained deep neural network proxy model;

[0071] S52: Elite selection is based on multi-objective comprehensive evaluation, forming a weight vector w, where the weights correspond to delay time, economic loss, number of severely delayed flights, and fairness, respectively. The true objective function is used to evaluate elite individuals, and the true evaluation formula is:

[0072] F real (P elite )=[f1(P elite ),f2(P elite ),f3(P elite ),f4(P elite )] T (twenty two)

[0073] Each objective function is calculated based on actual flight data, and through iterative optimization, a flight recovery scheduling scheme with optimal comprehensive performance under multi-objective equilibrium is finally output.

[0074] The technical solution of the present invention has the following advantages compared with the prior art:

[0075] 1. This invention presents a method for optimizing the energy efficiency of large-scale flight recovery by integrating deep learning and genetic algorithms, proposing a deep learning surrogate strategy. Sample pairs are generated using real flight data, and data quality is optimized through preprocessing. A surrogate model replaces the real evaluation, updating only for elite individuals, reducing costs while maintaining accuracy. This strategy provides a clear direction for the search, avoids random exploration, accelerates optimization convergence, and reduces computation time, meeting the timeliness requirements of airport scheduling.

[0076] 2. This invention comprehensively considers four types of optimization objectives. By optimizing total delay time, economic losses, the number of severe delays, and scheduling fairness, it ensures that the generated flight recovery plan is closer to actual scheduling needs. These four objectives cover flight operations, economy, passenger experience, and fairness, avoiding the limitations of traditional solutions. The generated recovery plan better meets the actual needs of airports, airlines, and passengers, and has strong feasibility for implementation.

[0077] 3. In the preprocessing stage, this invention extracts multi-dimensional features such as planned take-off and landing times, flight distances, and airline information. Combined with predefined rules and weighting factors, this method can be adapted to flight recovery tasks under different airports and operating conditions. It also has interpretability and scalability, providing a stable and flexible model foundation for intelligent scheduling systems and adapting to flight recovery tasks in different scenarios. Attached Figure Description

[0078] The technical solution and its beneficial effects of the present invention will become apparent from the following detailed description of specific embodiments in conjunction with the accompanying drawings.

[0079] Figure 1 The diagram shown is a flowchart of a large-area flight recovery energy efficiency optimization method that integrates deep learning and genetic algorithms according to an embodiment of the present invention.

[0080] Figure 2 The diagram shown is a model of a large-area flight recovery energy efficiency optimization method that integrates deep learning and genetic algorithms according to an embodiment of the present invention. Detailed Implementation

[0081] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0082] Example 1

[0083] This embodiment provides a method for optimizing energy efficiency in large-scale flight recovery by integrating deep learning and genetic algorithms, such as... Figures 1-2 As shown, it includes the following steps:

[0084] S10: Using real flight data containing planned take-off and landing times, actual operational status, and flight attribute information, filter and clean relevant flights at the target airport, removing canceled and diverted data while retaining take-off and landing related fields; extract attribute information for each flight, such as planned take-off and landing times, flight distance, airline, month, and take-off and arrival airports; calculate a comprehensive priority index based on predefined airline priority, flight distance standardization functions, and peak period identification rules; and construct three types of economic weighting factors: airline losses, airport losses, and passenger losses.

[0085] S20: Construct a multi-objective optimization model with the goals of minimizing total delay time, minimizing total economic loss, minimizing the number of flights delayed by more than 30 minutes, and maximizing scheduling fairness; design nonlinear inequality constraints, allowing a maximum of 3 flights to take off and land in each time slice and ensuring that the total number of takeoffs and landings during special peak periods does not exceed a threshold.

[0086] S30: Based on the initial population generation strategy of genetic algorithm, construct a large batch of random flight time allocation schemes, and call the real objective function and constraint function to evaluate each scheme; construct the evaluation results into sample pairs with flight time slice vectors as input and multiple objective values ​​as output; introduce a preprocessing mechanism of logarithmic transformation and standard deviation normalization;

[0087] S40: Construct a deep neural network structure containing multiple fully connected layers, batch normalization layers, and nonlinear activation functions; adopt a multi-objective regression output layer to simultaneously predict four objective values: delay, loss, number of flights exceeding limits, and fairness; use the Adam optimizer with a learning rate decay scheduling strategy for training, and set appropriate batch size and maximum training epochs.

[0088] S50: During the optimization process, a surrogate model is used to replace the real objective function for rapid evaluation; each generation of optimization executes an elite selection mechanism to perform real evaluation and update on several individuals with the best surrogate evaluation values; through iterative optimization, a set of flight recovery scheduling schemes with optimal comprehensive performance under multi-objective equilibrium is finally output.

[0089] Specifically, step S10 includes the following steps:

[0090] S11: Filter valid flights from the original flight dataset, removing canceled and diverted flights. The filtering criteria are defined as follows:

[0091] H cancel =0 AND H diverted =0 (23)

[0092] Among them, H cancel H indicates the flight is cancelled. divertedThis indicates that the flight has made an emergency landing.

[0093] S12: For each valid flight that successfully completes a flight mission, F i Extract its key attribute information to construct attribute vector A. i =[f cd ,f ca ,f dis ,f car ,f mon ,f ori ,f des ], where f cd For the planned takeoff time, f ca For the planned arrival time, f dis f represents the flight distance. car For airline code, f mon For the month of flight, f ori For the departure airport code, f des For destination airport code;

[0094] S13: Predefine the airline priority factor P based on the airline's service level. ci Max-min normalization is used to eliminate the influence of flight distance dimensions, and the flight distance standardization factor P is used. di The calculation formula is:

[0095]

[0096] Where D max and D min D represents the maximum or minimum distance in a flight. i Let P represent the distance to the i-th flight. For the peak period identification factor, based on airport operational patterns, peak periods are identified and assigned a weight of 1, while off-peak periods are assigned a weight of 0.5. The peak period identification factor is defined as P. ti Taking into account the three factors, calculate the comprehensive result.

[0097] Priority indicators:

[0098] P i priority =0.5 + 0.3 × P ci +0.1×P di +0.1×P ti (25)

[0099] Airlines loss coefficient The calculation formula is:

[0100]

[0101] in For airline rating adjustment factors, Airport loss coefficient for the i-th flight in the quarter The calculation formula is:

[0102]

[0103] in This represents the size factor and passenger loss coefficient of the i-th airport. This mainly includes time cost, missed flight losses, and decreased satisfaction. The calculation formula is as follows:

[0104]

[0105] Specifically, step S20 includes the following steps:

[0106] S21: The first objective is to minimize the total delay time. The objective function for the total delay time is defined as the weighted sum of the delay times of all flights, and the calculation formula is as follows:

[0107]

[0108] Where w i w is the weighting coefficient for the i-th flight. i =P i priority f delay (x) represents the delay time of the i-th flight. This objective function ensures that only flights with delays exceeding 15 minutes are included in the calculation. delay (x)>15) is an indicator function, which is 1 when the condition is true, and 0 otherwise;

[0109] S22: The second objective is to minimize the total economic loss. The objective function for total economic loss considers the losses from airlines, airports, and passengers, and the calculation formula is as follows:

[0110]

[0111] Where C delay C is the conversion factor for delay loss. delay = 1 / 10000;

[0112] S23: The third objective is to minimize the number of flights delayed by more than 30 minutes. The objective function for the number of severely delayed flights is defined as follows:

[0113]

[0114] Where I(f) delay (x) > 30) is an indicator function, with a value of 1 when the delay exceeds 30 minutes;

[0115] S24: The fourth objective is to maximize scheduling fairness, which is measured by minimizing the standard deviation of delay time. The formula is as follows:

[0116]

[0117] Where μ is the average delay time, calculated using the following formula:

[0118]

[0119] S25: The first constraint is to constrain the capacity of each time slice to ensure that the airport's processing capacity is not exceeded. The formula for calculating the constraint is:

[0120]

[0121] Where I(x) i =j) is an indicator function, with a value of 1 when flight i is assigned to time slice j, R capslot =3, indicating that the capacity limit for each time slot is 3 aircraft. The second constraint is the total capacity constraint for peak periods, ensuring the normal operation of flights during peak hours. Its calculation formula is:

[0122]

[0123] in, and These are the start and end time slices for a specific period, C. special This refers to the capacity threshold for specific time periods.

[0124] Specifically, step S30 includes the following steps:

[0125] S31: The population representation of the genetic algorithm is a set of flight time allocation schemes, where each individual represents a complete scheduling scheme. The required number of samples is determined based on the problem complexity, and the calculation formula is as follows:

[0126] N num =max(1000,10×N×log(N)) (36)

[0127] Where N represents the number of flights;

[0128] S32: For each individual P i The actual objective function is called for evaluation, and the calculation formula is as follows:

[0129] F i f =[f1(P i ),f2(P i ),f3(P i ),f4(P i)] T (37)

[0130] The degree of constraint violation is determined by the constraint vector F. i g The calculation formula is as follows:

[0131] F i g =[g1(P i ),g2(P i )] T (38)

[0132] S33: Convert the time slice allocation for each individual into a feature vector X i The input feature vector is constructed, and the evaluation result is used to construct the target vector Y. i The output target vector is represented by the given vector, and the final training sample set is:

[0133] N data ={(X i ,Y i | i = 1, 2, ..., N num} (39)

[0134] S34: Perform a logarithmic transformation on the target vector, the calculation formula is as follows:

[0135] Y i log =[log(1+f1(P)] i )),log(1+f2(P i )),log(1+f3(P i )),log(1+f4(P i ))] T (40)

[0136] Calculate the mean and standard deviation after logarithmic transformation, and then perform Z-score standardization on them.

[0137] Specifically, step S40 includes the following steps:

[0138] S41: Construct four fully connected layers. The mathematical expression for the first fully connected layer is:

[0139] z1 = W1 × x norm +b1 (41)

[0140] Where W1 is the weight matrix, b1 is the bias vector, z1 is the linear transformation output, a batch normalization layer is constructed, batch statistical calculation and standardization transformation are performed, and finally the learnable parameter transformation is performed. The nonlinear activation function is the ReLU activation function.

[0141] S42: Multi-objective regression predicts four target values ​​simultaneously. The calculation formula is as follows:

[0142] Predict the total delay time, total economic loss, number of flights delayed by more than 30 minutes, and scheduling fairness indicators.

[0143] Specifically, step S50 includes the following steps:

[0144] S51: During the optimization process, a surrogate model is used to quickly evaluate the true objective function. The surrogate model is used to predict and then destandardizes to obtain the true objective value. The calculation formula is as follows:

[0145]

[0146] Where F net (x) represents the trained deep neural network proxy model;

[0147] S52: Elite selection is based on multi-objective comprehensive evaluation, forming a weight vector w, where the weights correspond to delay time, economic loss, number of severely delayed flights, and fairness, respectively. The true objective function is used to evaluate elite individuals, and the true evaluation formula is:

[0148] F real (P elite )=[f1(P elite ),f2(P elite ),f3(P elite ),f4(P elite )] T (44)

[0149] Each objective function is calculated based on actual flight data, and through iterative optimization, a flight recovery scheduling scheme with optimal comprehensive performance under multi-objective equilibrium is finally output.

[0150] Example 2

[0151] Referring to Example 1, this example uses the parameters and results calculated in Example 1 to compare with the traditional genetic algorithm, demonstrating the superiority of this example. In the specific comparison, we used different models and compared them across different metrics. The final results show that this example is preferred over the comparison method.

[0152] 1. Comparison Algorithm

[0153] GA model: A stochastic search and optimization method based on natural selection and genetic mechanisms. It continuously evolves to find the optimal or near-optimal solution in the search space by simulating the processes of selection, crossover, and mutation in biological evolution.

[0154] 2. Configuration Details

[0155] This embodiment presents a multi-objective flight recovery energy efficiency optimization method based on a deep learning agent model. The method first loads real historical flight data from the flightdata.csv file and preprocesses the target airport ATL (Airline Time Limit), including filtering valid flights, converting planned and actual times to minutes, and calculating flight priorities and economic loss coefficients for airlines, airports, and passengers. The optimization model discretizes flight takeoff and landing times with a 5-minute time slice granularity, and sets a regular capacity constraint of up to 3 flights per 5 minutes and a limit of up to 33 flights during the special period from 00:00 to 01:00. The core objective of the optimization is to minimize a four-dimensional objective function consisting of total delay time, total economic loss, the number of flights delayed by more than 30 minutes, and delay time fairness. To efficiently solve this complex optimization problem, this embodiment employs a deep neural network. The structure progresses from the input layer to a 128-node fully connected layer, then to BatchNorm, ReLU, a 64-node fully connected layer, then to BatchNorm, ReLU, a 32-node fully connected layer, ReLU, and finally a 4-node output layer as a surrogate model. The Adam optimizer is used for 50 training epochs with an initial learning rate of 0.001. The training data is generated by randomly sampling the solution space 100 times and incorporating a constraint violation penalty mechanism. The objective values ​​are logarithmically standardized to improve training stability. Subsequently, a genetic algorithm with a population size of 30 and 30 generations is integrated for optimization search. This algorithm includes elite retention (3 individuals per generation), tournament selection, single-point crossover, and random mutation (probability 0.1, amplitude over 3 time slices). The trained surrogate model is used to quickly evaluate population fitness, with precise evaluation only performed on elite individuals to ensure solution quality and search efficiency. Finally, the optimization results output the optimal values ​​of the four objective functions.

[0156] 3. Comparison Results

[0157] As shown in Table 1 below, compared with the traditional genetic algorithm (GA), the deep learning-based augmentation optimization method provided in this embodiment achieves a significant overall performance improvement. Experimental results show that the total delay time of this method is 50496 minutes, a 12.0% reduction compared to GA's 57368 minutes, and fairness is improved by 48.7%. Significant optimizations are achieved in both key indicators, while maintaining the same level of severe delay control capability, ensuring that the number of flights exceeding 30 minutes is 465. Although the economic loss indicator increases slightly, this method achieves a better overall balance in multi-objective trade-offs, with the optimal comprehensive fitness value decreasing from 59944.42 to 53612.48, resulting in an overall performance improvement of 10.6%. This demonstrates that this embodiment, guided by a deep learning agent model, can obtain a fairer, more efficient, and comprehensively better flight recovery solution, effectively addressing the shortcomings of traditional methods in solution space search capability and multi-objective balance.

[0158] Table 1

[0159]

[0160] The above description is merely an exemplary embodiment of the present invention and does not limit the scope of patent protection of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A large-area flight recovery energy efficiency optimization method fusing deep learning and genetic algorithm, characterized in that, The method comprises the following steps: S10: filtering and cleaning the target airport related flights by using real flight data containing scheduled take-off and landing time, actual operation, and flight attribute information, removing canceled and alternate landing data, and retaining take-off and landing related fields; extracting the scheduled take-off and landing time, flight distance, airline, month, and take-off and landing airport attribute information of each flight; calculating a comprehensive priority index based on predefined airline priority, flight distance standardization function, and peak period identification rules; and constructing airline loss, airport loss, and passenger loss three types of economic weight factors; S20: constructing a multi-objective optimization model with the objectives of minimizing total delay time, minimizing total economic loss, minimizing the number of flights delayed for more than 30 minutes, and maximizing scheduling fairness; designing nonlinear inequality constraints, allowing a maximum of 3 flights to take off and land in each time slice and a total number of take-offs and landings in special peak periods not exceeding a threshold; S30: constructing a large number of random flight time allocation schemes based on the initial population generation strategy of genetic algorithm, and evaluating each scheme by calling real objective function and constraint function; constructing sample pairs with flight time slice vector as input and multi-objective values as output; introducing a preprocessing mechanism of logarithmic transformation and standard deviation normalization; S40: constructing a deep neural network structure containing multiple fully connected layers, batch normalization layers, and nonlinear activation functions; using a multi-objective regression output layer to simultaneously predict delay, loss, number of flights exceeding the limit, and fairness; using Adam optimizer with learning rate decay scheduling strategy for training, and setting appropriate batch size and maximum training rounds; S50: using a surrogate model to replace the real objective function for rapid evaluation during optimization; performing elite selection mechanism in each generation of optimization, and performing real evaluation and updating on individuals with optimal proxy evaluation values; and finally outputting a set of flight recovery scheduling schemes with optimal comprehensive performance under multi-objective balance through iterative optimization.

2. The method of claim 1, wherein The step S10 comprises the following steps: S11: selecting valid flights from the original flight data set, removing canceled and alternate landing flights, and defining the selection condition as: wherein H cancel represents a flight cancellation status, H diverted represents a flight diversion status; S12: For each successful flight mission performing effective flight F i , extract its key attribute information, constitute attribute vector A i = cd , ca , dis , car , mon , ori , des ], Where f cd is the planned take-off time, f ca is the planned arrival time, f dis is the flight distance, f car is the airline code, f mon is the flight month, f ori is the departure airport code, f des is the destination airport code; S13: Predefine the airline priority factor P according to the service level of the airline ci The flight distance standardization factor P is calculated by using the maximum-minimum value normalization to eliminate the influence of the flight distance dimension. di The calculation formula is as follows: where D max and D min denote the maximum or minimum distance in the flight, D i denotes the distance of the i-th flight, for the peak period identification factor, according to the airport operation rules, the peak period is identified and given a weight of 1, and a non-peak period is given a weight of 0.5, the peak period identification factor is defined as P ti , and the comprehensive priority index is calculated by comprehensively considering the three factors: Airline loss factor The formula is: wherein is the airline class adjustment factor, is the seasonal, airport loss factor for the ith flight The calculation formula is: wherein represents the scale factor of the ith airport, the passenger loss coefficient mainly includes time cost, loss of missing flight and decline of satisfaction, and the calculation formula is:

3. The method of claim 1, wherein The step S20 comprises the following steps: S21: the first objective is to minimize the total delay time, and the total delay time objective function is defined as the weighted sum of all flight delay times, and the calculation formula is: where w i is the weight coefficient of the i-th flight, f delay (x) is the delay time of the i-th flight, the objective function ensures that only flights with a delay time exceeding 15 minutes are counted, I(f delay (x) > 15) is an indicator function that is 1 when the condition is true and 0 otherwise. S22: the second objective is to minimize the total economic loss, and the total economic loss objective function considers the losses of airlines, airports, and passengers, and the calculation formula is: where C delay is the delay loss conversion factor, C delay = 1 / 10000; S23: the third objective is to minimize the number of flights delayed for more than 30 minutes, and the number of severely delayed flights objective function is defined as: where I(f delay (x) > 30) is an indicator function that takes the value 1 when the delay time exceeds 30 minutes. S24: the fourth objective is to maximize scheduling fairness, and the scheduling fairness is measured by minimizing the standard deviation of delay time, and the calculation formula is: wherein μ is the average delay time, and the calculation formula is: S25: the first constraint condition is to constrain the capacity of each time slice to ensure that the airport handling capacity is not exceeded, and the calculation formula of the constraint condition is: where I(x i = j) is an indicator function that takes the value 1 when flight i is assigned to time slice j, R capslot = 3, indicating that the capacity limit for each time slice is 3 flights, and the second constraint is a total capacity constraint for the special peak hours, which ensures the normal operation of flights during peak hours, and its calculation formula is: wherein, and are the start and end time slices of the special time period, respectively, C special is the capacity threshold of the special time period.

4. The method of claim 1, wherein The step S30 comprises the following steps: S31: The population of the genetic algorithm represents a set of flight time allocation schemes, each individual represents a complete scheduling scheme, the required sample size is determined according to the complexity of the problem, and the calculation formula is: N num = max(1000, 10 x N x log(N)) (14); where N is the number of flights; S32: For each individual P i , call the true objective function for evaluation, the formula is: The constraint violation degree is calculated by the constraint vector is represented by the following formula: S33: convert the time slice allocation of each individual into a feature vector X i To construct the input feature vector, the evaluation result is constructed as the target vector Y i To represent the output target vector, the final training sample set is obtained: N data = {(X i ,Y i )| i = 1,2,...,N num} (17); S34: Logarithmic transformation is performed on the target vector, and the calculation formula is: The mean and standard deviation of the log-transformed values are computed and Z-score normalized.

5. The method of claim 1, wherein The step S40 includes the following steps: S41: Four fully connected layers are constructed, and the mathematical expression of the first fully connected layer is: z1 = W1 x x norm + b1 (19); wherein W1 is a weight matrix, b1 is a bias vector, z1 is a linear transformation output, a batch normalization layer is constructed, batch statistics are calculated and normalized changes are made, and finally a learnable parameter transformation is made, and a ReLU activation function is used as a nonlinear activation function; S42: Multi-objective regression simultaneously predicts four target values, and the calculation formula is: Total delay time, total economic loss, number of flights delayed more than 30 minutes and dispatch fairness index are predicted respectively.

6. The method of claim 1, wherein The step S50 includes the following steps: S51: In the optimization process, the surrogate model replaces the real objective function for rapid evaluation, the real target value is obtained by using the surrogate model prediction and inverse standardization, and the calculation formula is: wherein F net (x) is a trained deep neural network proxy model; S52: Elite selection is performed based on multi-objective comprehensive evaluation, a weight vector w is formed, the weights correspond to delay time, economic loss, number of severely delayed flights and fairness, and the real objective function of the elite individual is evaluated, and the real evaluation formula is: F real (P elite ) = [f1(P elite ), f2(P elite ), f3(P elite ), f4(P elite )] T (22) Wherein, each target function is calculated according to the actual flight data, and through iterative optimization, a set of flight recovery scheduling schemes with optimal comprehensive performance under multi-objective balance is finally output.

Images