A method for early warning of airport runway icing risk by incorporating neural networks

Abstract

This invention relates to the field of aviation equipment technology, specifically to a method for early warning of airport runway icing risk by fusing neural networks, comprising the following steps: S1: Establishing an airport runway icing risk warning model by fusing a temporal convolutional network and a bidirectional gated recurrent unit; S2: Determining the model input quantities and performing normalization processing to eliminate the influence of dimensions; S3: Adding an attention mechanism to the fused model; S4: Establishing a Golden Jackal optimization algorithm model; S5: Improving the Golden Jackal optimization algorithm and using it to optimize the parameters; S6: Predicting the airport runway risk level; This invention, through the analysis of historical meteorological data, enables the model to predict the icing risk level in real time, helping airport managers to take countermeasures in advance and reduce safety hazards caused by runway icing.

Description

Technical Field

[0001] This invention relates to the field of aviation equipment technology, and specifically to a method for early warning of airport runway icing risk by incorporating neural networks. Background Technology

[0002] With the rapid development of the air transport industry, airport operational safety has become a focus of attention. Airport runway safety is particularly critical in winter due to snow, ice, and slush. Wet runways can affect aircraft control and stopping distance. More than 60% of the world's airports are located in snowy regions. Snow accumulation on airport runways reduces aircraft tire grip, threatening airport operational safety during snowfall and hail. Runway icing not only reduces the safety of aircraft takeoffs and landings but can also lead to flight delays, cancellations, and even aviation accidents. Therefore, establishing accurate runway icing risk early warning models is of great significance for ensuring airport operational safety.

[0003] Chinese patent 201910942268.X discloses a method and apparatus for identifying the ice thickness of transmission line conductors. The method uses an imaging device to acquire images, then uses a pre-trained ice target recognition network to identify ice on the line, and finally analyzes the icing area to determine the specific ice thickness. However, it still has many shortcomings. For example, this method requires acquiring icing images, and airport runways are long-term used equipment with low ice thickness; therefore, unclear images can lead to inaccurate prediction models.

[0004] Therefore, there is an urgent need for an airport runway icing risk early warning method that integrates neural networks to solve the problems mentioned above in airport runway de-icing work. Summary of the Invention

[0005] The purpose of this invention is to provide an airport runway icing risk early warning method that integrates neural networks, uses meteorological data for prediction, obtains the risk level, and then carries out de-icing operations.

[0006] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:

[0007] A method for early warning of airport runway icing risk integrating neural networks is proposed. This method utilizes an attention mechanism combined with a temporal convolutional network (TCN) and a bidirectional gated recurrent unit (IGJO) optimized by an improved Golden Jackal Optimization (IGJO) algorithm. First, a TCN-BiGRU model is constructed. TCN captures long-term dependencies through dilated and causal convolutions, while BiGRU simultaneously captures sequential dependencies. Combining the two allows for better handling of complex time-series data. Second, an attention mechanism is introduced to improve the model's sensitivity to key information while reducing reliance on secondary information. Finally, logistic mapping, elite retention strategy, adaptive inertial weights, collaborative learning mechanism, and adaptive adjustment of control parameters are used to improve the GJO, thereby optimizing the model parameters.

[0008] Includes the following steps:

[0009] S1: Establishment of an airport runway icing risk early warning model integrating temporal convolutional networks and bidirectional gated recurrent units;

[0010] S2: Determine the model input quantities and perform normalization processing to eliminate the influence of dimensions;

[0011] S3: Incorporate an attention mechanism into the fusion model;

[0012] S4: Establishment of the Golden Jackal Optimization Algorithm Model;

[0013] S5: Improve the Golden Jackal optimization algorithm and use the Golden Jackal optimization algorithm to perform parameter optimization;

[0014] S6: Airport runway risk level prediction.

[0015] The beneficial effects of this invention are:

[0016] The airport runway icing risk early warning model based on IGJO-TCN-BiGRU-Attention comprehensively addresses the problems of insufficient prediction accuracy, limited feature extraction capabilities, low hyperparameter selection efficiency, lack of dynamic attention capabilities, and insufficient real-time performance and adaptability in existing technologies. By improving the Ingolfo optimization algorithm (IGJO), the model achieves global-local collaborative optimization, effectively avoiding the limitations of manual parameter tuning through a dynamic search strategy, and significantly improving the hyperparameter optimization efficiency and global convergence performance of deep learning models. The Temporal Convolutional Network (TCN), through the combination of causal convolution and extended convolution, realizes long-term and short-term dynamic modeling of multi-dimensional meteorological time series of airport runways, and can accurately capture the multi-scale variation patterns of features such as temperature and humidity. The Bidirectional Gated Recurrent Unit (BiGRU) further enhances the model's ability to capture forward and backward dependencies of time series, while using the gating mechanism to filter irrelevant information, improving the accuracy and robustness of data feature extraction. The Attention mechanism, by dynamically calculating the importance weights of features, focuses on the most critical time steps and feature points for icing risk prediction, enhancing the model's sensitivity to extreme weather conditions and sudden risks.

[0017] The entire model architecture optimizes hyperparameters through IGJO, enabling TCN, BiGRU, and Attention to work collaboratively, forming a highly efficient end-to-end prediction system. The model not only improves the accuracy of runway icing risk warnings but also significantly enhances real-time response performance, enabling rapid processing of large-scale data inputs and adaptation to different airport and weather conditions. Through multi-source data fusion and dynamic feature weighting, this solution can detect runway icing risks in advance under complex weather conditions, providing reliable early warning support for airport scheduling and operations, and significantly reducing aviation safety hazards. This technology overcomes the limitations of traditional methods, combining the advantages of optimization algorithms and deep learning, achieving a comprehensive improvement in prediction efficiency and accuracy, and demonstrating high practicality and promotional value, providing an intelligent and innovative solution for airport icing risk management.

[0018] This invention employs a combination of Temporal Convolutional Networks (TCNs) and Bidirectional Gated Recurrent Units (BiGRUs) to provide superior time-series data modeling capabilities. TCNs capture long-term dependencies through dilated convolutions without increasing computational complexity, while causal convolutions ensure that information in the input sequence comes only from the past, with no information leaking into the future. Through residual connections, TCNs overcome the vanishing gradient problem in deep networks, allowing for the operation of deep models. Meanwhile, BiGRUs, through their forward and backward information processing structures, further uncover hidden patterns and potential relationships in the data. This bidirectional information flow mechanism allows for comprehensive and in-depth analysis of time-series data, thereby improving the accuracy and robustness of risk prediction.

[0019] By introducing an attention mechanism, the model can dynamically allocate weights to different input features, thereby enhancing its sensitivity to key meteorological information. The attention mechanism mimics the selective attention process in humans, enabling the model to focus on information closely related to icing risk while suppressing interference from irrelevant information. By calculating the attention distribution among features, the model can identify the features that contribute most to risk prediction and assign them higher weights. This not only improves prediction accuracy but also enhances the model's adaptability to complex and changing meteorological conditions.

[0020] The improved Golden Jackal Optimization Algorithm (IGJO) achieves efficient optimization of model parameters through mechanisms such as Logistic mapping, elitism retention strategy, and adaptive inertia weights. Logistic mapping breaks population clustering during initialization, increasing global search capability and preventing the algorithm from getting trapped in local optima. The elitism retention strategy ensures that the optimal solution found in iterations is not replaced by a worse solution, improving search stability. Adaptive inertia weights dynamically adjust the search intensity at different stages of the algorithm, enhancing global exploration capability in the early stages and strengthening local exploration capability in the later stages, thus achieving faster and more stable convergence.

[0021] By adaptively adjusting control parameters and employing a collaborative learning mechanism, the model exhibits excellent robustness when facing problems of varying complexity. Adaptive adjustment of control parameters allows the algorithm to self-adjust to adapt to environmental changes, achieving a balance between exploration and exploitation. The collaborative learning mechanism avoids premature convergence by integrating information from multiple candidate solutions and leveraging the complementarity between them. This mechanism enables the model to perform more comprehensive optimization from the perspective of multiple solutions, improving its applicability and prediction accuracy in different contexts.

[0022] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a model structure diagram of an embodiment of the present invention;

[0025] Figure 2 This is a diagram of the TCN causal dilated convolution structure according to an embodiment of the present invention;

[0026] Figure 3This is a diagram of the TCN residual module according to an embodiment of the present invention;

[0027] Figure 4 This is a diagram of the GRU network structure according to an embodiment of the present invention;

[0028] Figure 5 This is an optimal fitness curve diagram of an embodiment of the present invention;

[0029] Figure 6 This is a diagram showing the model prediction results of an embodiment of the present invention;

[0030] Figure 7 This is an iterative curve diagram of an embodiment of the present invention. Detailed Implementation

[0031] 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.

[0032] Example 1

[0033] This embodiment describes an airport runway icing risk early warning method integrating neural networks. It employs an attention mechanism combined with an improved Golden Jackal Optimization (IGJO) algorithm-optimized temporal convolutional networks and bidirectional gated recurrent units to create an airport runway icing risk early warning model. First, a TCN-BiGRU model is constructed. TCN captures long-term dependencies through dilated convolutions and causal convolutions, while BiGRU simultaneously captures sequential dependencies. The combination of these two approaches better handles complex time-series data. Second, an attention mechanism is introduced to improve the model's sensitivity to key information while reducing reliance on secondary information. Finally, Logistic mapping, elite retention strategy, adaptive inertial weights, collaborative learning mechanism, and adaptive adjustment of control parameters are used to improve the GJO model parameters, thereby optimizing the model parameters.

[0034] Includes the following steps:

[0035] S1: An airport runway icing risk early warning model was established by integrating a temporal convolutional network (TCN) and a bidirectional gated recurrent unit (BGRU). The TCN model extracts features and temporal relationships through one-dimensional causal convolution, residual connections accelerate convergence, and dilated convolution achieves temporal feature extraction. The BiGRU model, as a variant of the recurrent neural network, has nonlinear fitting capabilities and can effectively extract data features.

[0036] S2: Determine the model input quantities and perform normalization to eliminate the influence of units. Features in meteorological data have different units or orders of magnitude, and unnormalized features may have a disproportionate impact on the model results. Normalization ensures that each feature contributes more fairly to the model, without bias towards certain features due to differences in units or orders of magnitude;

[0037] S3: An attention mechanism is added to the fusion model to enhance the proportion of useful information by assigning weights based on probability, thereby avoiding the loss of useful information;

[0038] S4: The Golden Jackal Optimization Algorithm Model is inspired by the hunting behavior of golden jackals (male and female) in nature, and aims to solve function optimization, engineering design and other complex optimization problems.

[0039] S5: Improve the Golden Jackal optimization algorithm and use the Golden Jackal optimization algorithm to optimize parameters. Improve GJO by using Logistic mapping, elite retention strategy, adaptive inertial weight, collaborative learning mechanism and adaptive adjustment of control parameters.

[0040] S6: Airport runway risk level prediction.

[0041] In this embodiment, the TCN (Temporal Convolutional Networks) model used in the airport runway icing risk early warning model established by S1 integrating the temporal convolutional network and the bidirectional gate control recurrent unit is a new type of network model that is improved and optimized based on CNN. It extracts high-dimensional features from the input data through causal convolution, dilated convolution and residual connections to ensure the temporality of the model.

[0042] The formula for causal dilation convolution is as follows:

[0043]

[0044] In the formula: y [t] To output the value of the sequence at time point t; x [t] ω represents the value of the input sequence at time t. [k] is the weight of the kernel at index k; K is the length of the convolution kernel; b is the bias term; d is the dilation rate.

[0045] The TCN model consists of multiple residual modules. Residual connections between these modules not only enable the network to transfer information across layers but also ensure input-output consistency, helping to address gradient fading and making it easier to train deep architectures. Residual modules allow information to skip several layers, promoting gradient flow.

[0046] In this embodiment, the BiGRU (Bidirectional Gated Recurrent Unit) used in the airport runway icing risk early warning model established by fusing the S1 temporal convolutional network and the bidirectional gated recurrent unit is a variant of the recurrent neural network (RNN). BiGRU combines the concepts of GRU (Gated Recurrent Unit) and bidirectional RNN. GRU aims to solve the gradient vanishing problem in traditional RNNs and simplifies the structure of the memory unit. Its formula is as follows:

[0047] z t =σ(W Z *[h t-1 ,x t ])

[0048] r t =σ(W r *[h t-1 ,x t ])

[0049]

[0050] Where: z t Update the gated signal; h t-1 Implicit historical state; x t The input data at time t; W Z W r W q Both are weight matrices; σ is the sigmoid function; r t For reset signal; h t This is a hidden state; These are candidate latent states.

[0051] To fully consider the information patterns arising from changes in the data, a bidirectional GRU structure is adopted. In the forward GRU layer, data is passed to the neural network for training through forward propagation, and correlations in the data are identified. In the backward GRU layer, the output data is backpropagated to uncover inverse correlations. BiGRU can obtain information between data from both directions, improving model performance. The formula is as follows:

[0052]

[0053] In the formula: GRU(·) is a gated loop unit; For the output of the hidden layers in the forward and backward directions; α t β t b represents the weights of the corresponding hidden layers. t For bias.

[0054] In this embodiment, step S2 involves determining the model input quantities and performing normalization to eliminate the influence of dimensions. The formula is as follows:

[0055]

[0056] In the formula, X n The normalized value of the data; x max x min These represent the maximum and minimum values ​​in the data, respectively. To ensure that the output results have the same dimensions as the true values, the output data also needs to be denormalized.

[0057] In this embodiment, step S3: An attention mechanism is added to the fusion model. Attention is a method that allows a model to selectively focus on important parts of the input data. It mimics the characteristics of human attention, allowing the neural network to dynamically assign different weights to different parts of the input, thereby processing information more efficiently. The core of the attention mechanism lies in its ability to improve the model's sensitivity to key information while reducing its dependence on secondary information, thus improving the model's prediction accuracy. The calculation steps of the attention mechanism are as follows:

[0058]

[0059] In the formula, e t The value represents the attention distribution at time t; μ and ω are the weight values; a t This represents the weights that the attention module assigns to the features.

[0060] In this embodiment, the S4: Golden Jackal optimization algorithm model establishment mainly consists of four steps: (1) population initialization; (2) finding prey; (3) surrounding and continuing until the prey stops moving; (4) attacking the prey;

[0061] In this embodiment, the population initialization uses GJO, a population-based optimization algorithm, where the initial solution Y_0 is randomly and uniformly distributed across the search space.

[0062] Y0 = Y min +rand(Y max -Y min )

[0063] In the formula: Y max Ymin The upper and lower limits of the solution are defined; rand is a random variable in the range [0, 1].

[0064] Initialize the prey matrix Prey:

[0065]

[0066] In the formula, represents the j-th dimension of the i-th prey; n represents the number of prey; and d represents the solution variable.

[0067] During the optimization process, a fitness function is applied to estimate the fitness of each prey, resulting in the fitness matrix F of all prey. OA :

[0068]

[0069] In the formula: f is the fitness function; the optimal fitness corresponds to the position coordinates of the male jackal, and the second-best fitness corresponds to the position coordinates of the female jackal.

[0070] In this embodiment, the prey-hunting process, in nature, involves golden jackals waiting and searching for prey, with the male taking the lead in attacking first, followed by the female. The corresponding calculation model is as follows:

[0071] Y1(t)=Y M (t)-E·|Y M (t)-R L ·Prey(t)|

[0072] Y2(t)=Y FM (t)-E·|Y FM (t)-R L ·Prey(t)|

[0073] In the formula: t is the current iteration number; E is the escape energy of the prey; R L Y is a random vector based on the Levy distribution; Prey(t) is the position vector of the prey; M (t) and Y FM Y1(t) and Y2(t) represent the positions of the male and female jackals, respectively; Y1(t) and Y2(t) are the positions of the prey corresponding to the male and female jackals in the new iteration.

[0074] The formula for calculating the escape energy E of the prey is as follows:

[0075]

[0076] In the formula: E1 is the decreasing prey energy; E0 is the initial state of prey energy; T is the maximum number of iterations; c1 is a constant of 1.5, and E1 decreases linearly from 1.5 to 0 throughout the iteration process; r represents any value between [0,1].

[0077] A random vector R based on the Levy distribution L The representation is as follows:

[0078] R L =0.05*LF(y)

[0079]

[0080] In the formula, μ and ν represent random values ​​in [0,1]; β = 1.5; Γ(·) is the gamma function.

[0081] Update the position of the golden jackal using the following formula:

[0082]

[0083] In this embodiment, the process of surrounding and attacking the prey involves the prey's escape energy rapidly decreasing when attacked by the golden jackal. The jackal then surrounds the prey detected in the previous stage, begins its attack, and consumes it. The predation model is as follows:

[0084] Y1(t)=Y M (t)-E·|R L ·Y M (t)-Prey(t)|

[0085] Y2(t)=Y FM (t)-E·|R L ·Y FM (t)-Prey(t)|

[0086] In the formula: t is the current iteration number; E is the escape energy of the prey; R L Y is a random vector based on the Levy distribution; Prey(t) is the position vector of the prey; M (t) and Y FM Y1(t) and Y2(t) represent the positions of the male and female jackals, respectively; Y1(t) and Y2(t) are the positions of the prey corresponding to the male and female jackals in the new iteration.

[0087] In this embodiment, the Logistic mapping in S5 initializes the prey. GJO uses a random prey initialization method; if the initial solutions cluster, it will affect the optimization accuracy and increase the number of iterations. Using a chaotic mapping to update the initial position can effectively solve this problem. This invention uses the Logistic mapping to generate initial solutions distributed throughout the search space through the nonlinear dynamic characteristics of chaotic systems, thereby increasing the diversity of the population, improving the algorithm's global search capability, and helping to avoid the initial population getting trapped in local optima. The formula is as follows:

[0088] x n+1 =4x n(1-x n )

[0089] In the formula: x n x is the value of the chaotic variable in the current iteration step; n+1 This is the value of the chaotic variable for the next iteration.

[0090] In this embodiment, the elite retention strategy in S5 is used because GJO is prone to getting trapped in local optima, has a slow convergence speed, and poor stability. The elite retention strategy

[15] ensures that the currently found optimal solution is not lost in each iteration. This is achieved by retaining the best solution in the iteration and preventing it from being covered by a new non-optimal solution in subsequent iterations. This strategy can prevent the loss of the optimal solution due to randomness, thereby improving the stability and reliability of the algorithm. The implementation method is that after each iteration, if the fitness of the new solution is not as good as the fitness of the current optimal solution, the current optimal solution is retained. Its expression is:

[0091] if

[0092] then

[0093] In the formula: and These are the optimal solutions for the current generation and the previous generation, respectively. and These are the adaptability ratings for the current generation and the previous generation, respectively.

[0094] In this embodiment, the adaptive inertia weight in S5 is an important concept in Particle Swarm Optimization (PSO) and plays a crucial role in the algorithm's development and exploration phases. The balance between convergence speed and exploration capability in GJO is insufficient; a fixed inertia weight may cause the algorithm to converge too quickly or too slowly in certain situations. By dynamically adjusting the weight ω, the algorithm initially increases exploration capability to cover a larger search space, while later increasing development capability to refine the search for local optima. This mechanism improves the algorithm's global search capability and convergence speed. In PSO, the inertia weight determines the degree of influence of the current velocity on the new velocity. Its formula is as follows:

[0095]

[0096] In the formula: ω is the inertia weight; ω max ω is the maximum value of the initial inertia weight; min Iter represents the minimum value of the final inertia weight. max `iter` represents the maximum number of iterations for the algorithm; `iter` represents the current iteration number.

[0097] In the early stages of the algorithm, the inertia weight ω is relatively large, resulting in higher particle velocities and enabling the exploration of a larger search space, thus enhancing global search capabilities. In the later stages, the inertia weight ω gradually decreases, reducing particle velocities and focusing more on local searches, thereby enhancing local search capabilities. The inertia weight ω decreases linearly from ω max Smooth transition to ω min This enables a smooth transition from global search to local search.

[0098] In this embodiment, the collaborative learning mechanism in S5 learns and updates by integrating information from multiple candidate solutions. GJO lacks a global perspective and has limited ability to solve complex optimization problems. In the improved GJO, a new solution is formed by averaging the positions of the male and female jackals and the current solution. This approach increases the diversity of solutions, prevents premature convergence, and utilizes complementary information from multiple solutions to improve algorithm performance.

[0099]

[0100] In the formula: To search for the location of individuals after the update step; M i The new position for the optimal jackal; F i For the new position of the second-best jackal; This is the new position following the position of the elite individual.

[0101] The new position is obtained by averaging the positions of male jackals, female jackals, and elite individuals. This comprehensive consideration of information from the best, second-best, and elite individuals allows the algorithm to achieve a better balance between global and local search, effectively improving the algorithm's robustness and adaptability.

[0102] In this embodiment, the adaptive adjustment of control parameters in S5 addresses the difficulty of adapting to different problems with fixed GJO parameters (such as population size, number of iterations, mutation rate, etc.). The adaptive adjustment of control parameters can adaptively adjust certain parameters in the algorithm, enabling the algorithm to better balance global and local searches at different iteration stages, influencing the location update and search behavior of the search agent, and improving the convergence speed of the algorithm. The formula is as follows:

[0103]

[0104] In the formula: E1 is the control parameter for the current iteration; E max `iter` represents the initial maximum value of the control parameter; `iter` represents the minimum value of the control parameter. max is the maximum number of iterations for the algorithm; l is the current iteration number.

[0105] In the early stages of the algorithm, a larger control parameter encourages the search agent to explore extensively, enhancing global search capabilities. In the later stages, the control parameter gradually decreases, encouraging the search agent to perform refined local searches, enhancing local search capabilities. As the number of iterations increases, the control parameter E1 gradually decreases from E... max linearly decrease to E min Achieve a smooth transition from global search to local search.

[0106] In this embodiment, the data used in the S6 airport runway risk level prediction is a combination of airport de-icing records from a certain region and meteorological data for the corresponding time period. The sampling period is from January 7, 2023 to January 30, 2023; and from November 8, 2023 to February 4, 2024. The sampling interval is 1 hour, and a total of 2034 sets of data are selected from 6:00 to 23:00 each day. The airport runway icing risk level is divided into four levels: No icing risk (0), in which case no de-icing preparation is required and there is no risk of icing on the airport runway; Low icing risk (1), there will be no significant icing conditions in the airport and surrounding areas, or slight icing can be dealt with quickly; Medium icing risk (2), slight icing will occur around the airport, which can be dealt with quickly; High icing risk (3), icing occurs on the airport runway, taxiway, apron and aircraft surface, requiring enhanced monitoring and possible de-icing; Very high icing risk (4), poses a significant threat to airport operations, including but not limited to runway icing, slippery taxiway, icing of aircraft fuselage and wings, etc., requiring frequent de-icing operations.

[0107] Example 2

[0108] Reference Figure 1 This embodiment provides an airport runway icing risk early warning model that integrates neural networks. The specific steps are as follows:

[0109] 1) The collected meteorological data are reordered to reduce errors caused by specific sequences in the training data. Since the TCN-BIGRU prediction model used in this invention requires time-series data, the dataset needs to be reconstructed when defining variables. Each iteration starts from the first sample and continues until sufficient historical data is available for prediction. Then, 90% of the data is selected as the training set and 10% as the test set. Data preprocessing is completed to prevent the TCN loss function from failing to converge.

[0110] 2) Set initialization parameters, defining initial parameters such as maximum number of training iterations, gradient threshold, and initial learning rate.

[0111] 3) The TCN layer processes the input sequence and captures local spatiotemporal features; then the BiGR layer captures the long-term dependencies and contextual information of the sequence in two directions; the attention mechanism further focuses on key information on the output of BiGRU, and enhances the model's attention to important features through weighted summation, thereby improving the model's accuracy.

[0112] 3) The IBJO optimization algorithm optimizes the learning rate, the number of BIGRU neurons, the key value of the attention mechanism, and the regularization parameter. The optimal value of each training is fed into TCN-BIGRU-Attention, and the MAE before and after optimization is compared to find the best fitness. After obtaining the optimal parameters, they are fed into the model, and the results are inversely normalized before output.

[0113] [Temporal Convolutional Networks (TCN) model] is a novel network model based on and optimized from CNNs. It extracts high-dimensional features from input data through causal convolution, dilated convolution, and residual connections, ensuring the model's temporal stability. (The causal dilated convolution structure is referenced...) Figure 2 .

[0114] The TCN model consists of multiple residual modules. Each residual module is residually connected, which not only enables the network to transfer information across layers but also ensures input-output consistency, helps solve gradient fading problems, and makes it easier to train deep architectures. Residual modules allow information to skip several layers, promoting gradient flow. The TCN residual module references... Figure 3 .

[0115] GRU aims to solve the vanishing gradient problem in traditional RNNs and simplifies the structure of memory cells. (See reference...) Figure 4 .

[0116] Where: x t h represents the input data at time t. t-1 The hidden state output represents the historical time t-1. σ and tanh represent the sigmoid function and the hyperbolic tangent function, respectively. In order to simplify the internal structure and improve the computational efficiency, the GRU network structure integrates the input gate, forget gate and output gate structure in LSTM into an update gate (2) and a reset gate (3). The forgetting and selective memory of the neural network can be realized by using only an update gate, which greatly reduces the number of parameters.

[0117] Benchmark test functions were used to compare IGJO, GJO, MOF (Moth to a Flame) and GEO (Golden Eagle) algorithms. Single-peaked test functions f1, f2, and f3 were used to test the convergence progress and speed, while multi-peaked test functions f4, f5, and f6 were used to test the global search capability. The algorithm population was 20, the number of iterations was 200, the optimal solution was 0, and each algorithm was run independently 30 times. The best fitness was compared. Results are referenced... Figure 5 See Table 1.

[0118] Table 1:

[0119]

[0120]

[0121] The data used in this embodiment is a combination of airport de-icing records and meteorological data for the corresponding time periods in a certain region. The sampling period was from January 7, 2023 to January 30, 2023; and from November 8, 2023 to February 4, 2024. The sampling interval was 1 hour, and a total of 2034 sets of data were selected from 6:00 to 23:00 each day. The airport runway icing risk level is divided into four levels: No icing risk (0), in which case no de-icing preparation is required and there is no risk of icing on the airport runway; Low icing risk (1), there will be no significant icing conditions in the airport and surrounding areas, or slight icing can be dealt with quickly; Medium icing risk (2), slight icing may occur around the airport, which can be dealt with quickly; High icing risk (3), icing occurs on the airport runway, taxiway, apron, and aircraft surface, requiring enhanced monitoring and possible de-icing; Very high icing risk (4), poses a significant threat to airport operations, including but not limited to runway icing, slippery taxiways, and icing on the aircraft fuselage and wings, requiring frequent de-icing operations. Meteorological data used include temperature, humidity, air pressure, precipitation, and ground wind speed. Some data are referenced in Table 2.

[0122] Table 2:

[0123]

[0124] To measure prediction error across multiple scales, this embodiment uses four evaluation metrics: Mean Squared Error (MSE); Root Mean Squared Error (RMSE); Mean Absolute Error (MAE); and Coefficient of Determination (R-Square, R0). 2 The expressions for each indicator are as follows:

[0125]

[0126]

[0127] In the formula, To predict the risk level; y i The actual risk level; is the average risk level; n is the number of samples.

[0128] The smaller the MSE, RMSE, and MAE values, the closer the R² result is to 1, indicating higher prediction accuracy of the model.

[0129] Table 3:

[0130]

[0131] This invention uses Matlab 2023A software to build an algorithm model and conducts experimental simulation analysis. To better verify the effectiveness of the model, multiple models were used for prediction, and the comparison results are as follows: Figure 6 As shown in Table 3, the individual TCN, GRU, and BiGRU models do not possess good predictive capabilities, with the obtained data being relatively low. However, the TCN-GRU and TCN-BiGRU models with added attention mechanisms exhibit better predictive performance than the TCN-GRU and TCN-BiGRU models. The proposed IGJO-TCN-BiGRU-Attention model, compared to the GJO-TCN-BiGRU-Attention model and the TCN-BiGRU-Attention model, has a higher coefficient of determination than the selected comparison methods, while its mean squared error, root mean square error, mean absolute error, and symmetric absolute percentage error are the lowest compared to the selected methods. The proposed method's R-value compared to TCN-BiGRU-Attention... 2 The efficiency increased by 10.0399%, while MSE, RMS E, and MAE decreased by 0.0126182, 0.055331, and 0.004609, respectively. This is compared to the R of GJO-TC N-BiGRU-Attention. 2 It rose by 5.5485%, while MSE, RMSE, and MAE decreased by 0.0103852, 0.047221, and 0.036935, respectively.

[0132] The method of this invention compares the R of GJO-TCN-BiGRU-Attention. 2 The accuracy increased by 5.5485%, while MSE, RMSE, and MAE decreased by 0.0103852, 0.047221, and 0.036935, respectively. This demonstrates that the method proposed in this invention has higher accuracy in predicting airport runway risk warnings.

[0133] In summary, this invention uses temperature, humidity, air pressure, precipitation, and surface wind speed as basic inputs to construct an airport runway icing risk early warning model. It employs Logistic mapping, an elite retention strategy, adaptive inertial weights, a collaborative learning mechanism, and adaptive adjustment of control parameters to improve the Jinhao optimization algorithm. The learning rate, the number of neurons in the bidirectional gated recurrent unit, the attention mechanism key value, and the regularization parameters are optimized to ultimately achieve airport runway risk early warning. The goal of this invention is to develop an efficient and accurate airport runway icing risk early warning system, providing a scientific basis for airport operation management and improving the safety and reliability of winter operations. Through analysis of historical meteorological data, the model can predict icing risk levels in real time, helping airport managers take proactive measures to reduce safety hazards caused by runway icing.

[0134] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for early warning of airport runway icing risk integrating neural networks, characterized in that, Includes the following steps: S1: Establish an airport runway icing risk early warning model by fusing temporal convolutional networks and bidirectional gated recurrent units; S2: Determine the input quantities for the airport runway icing risk early warning model and normalize them to eliminate the influence of dimensions; use temperature, humidity, air pressure, precipitation, and surface wind speed as the basic input quantities; S3: An attention mechanism is added to the airport runway icing risk early warning model; S4: Establishment of the Golden Jackal Optimization Algorithm Model; S5: Improve the Golden Jackal optimization algorithm model and use the improved Golden Jackal optimization algorithm model to optimize the learning rate, the number of neurons in the bidirectional gate control loop unit, the attention mechanism key value, and the regularization parameter in the airport runway icing risk early warning model; S6: Conduct airport runway risk level prediction; Step S5 includes: The Logistic mapping is used to generate initial solutions throughout the search space by leveraging the nonlinear dynamics of the chaotic system. The formula is as follows: x n+1 =4x n (1-x n ) In the above formula: x n x is the value of the chaotic variable in the current iteration step. n+1 The value of the chaotic variable for the next iteration step; The elite retention strategy preserves the best solution in each iteration and prevents it from being overwritten by new, suboptimal solutions in subsequent iterations. The implementation involves retaining the current best solution after each iteration if the fitness of a new solution is lower than that of the current best solution. The expression is: if then In the formula: and These are the optimal solutions for the current generation and the previous generation, respectively. and These are the adaptability of the current generation and the previous generation, respectively; The algorithm dynamically adjusts the weights ω to increase exploration capabilities in the early stages to cover a larger search space, and then increases development capabilities in the later stages to refine the search for local optima; the formula is as follows: In the formula: ω is the inertia weight; ω max ω is the maximum value of the initial inertia weight; min Iter represents the minimum value of the final inertia weight. max `iter` represents the maximum number of iterations for the algorithm; `iter` represents the current iteration number. The inertia weight ω decreases linearly from ω max Smooth transition to ω min This enables a smooth transition from global search to local search. A new solution is formed by averaging the positions of the male and female jackals and the current solution; information from multiple candidate solutions is integrated and updated through a collaborative learning mechanism. In the formula: To search for the location of individuals after the update step; M i The new position for the optimal jackal; F i For the new position of the second-best jackal; This is the new position following the position of the elite individual; The new position is obtained by averaging the positions of male jackals, female jackals, and elite individuals, taking into account information from the best, second-best, and elite individuals. The adaptive adjustment control parameters adjust the parameters in the algorithm adaptively, and the formula is as follows: In the formula: E1 is the control parameter for the current iteration; E max This represents the initial maximum value of the control parameter; E min Iter is the minimum value of the control parameter. max is the maximum number of iterations for the algorithm; l is the current iteration number.

2. The airport runway icing risk early warning method fused with neural networks as described in claim 1, characterized in that: Step S1 specifically includes: Temporal Convolutional Neural Network (TCN) extracts high-dimensional features from input data through causal convolution, dilated convolution, and residual connections, ensuring the temporal nature of the model. The formula for causal dilation convolution is as follows: In the formula: y [t] To output the value of the sequence at time point t; x [t-d·k] ω represents the value of the input sequence at time t. [k] is the weight of the kernel at index k; K is the length of the convolution kernel; b is the bias term; d is the dilation rate; The TCN model consists of multiple residual modules, and each residual module is connected by residuals. The BiGRU (Bidirectional Gated Recurrent Unit) combines the features of a GRU and a bidirectional RNN, and its formula is as follows: z t =σ(W Z *[h t-1 ,x t ]) r t =σ(W r *[h t-1 ,x t ]) Where: z t Update the gated signal; h t-1 Implicit historical state; x t The input data at time t; W Z W r W q Both are weight matrices; σ is the sigmoid function; r t For reset signal; h t This is a hidden state; Candidate hidden states; A bidirectional gated recurrent unit (GRU) structure is employed. In the forward GRU layer, data is passed to the neural network for training through forward propagation, and correlations between data are sought. The input layer consists of 6 neurons, and the output layer consists of 1 neuron. In the backward GRU layer, the output data is backpropagated to uncover inverse correlations between data. BiGRU obtains information between data from both forward and backward directions, improving model performance. The formula is as follows: In the formula: GRU(·) is a gated loop unit; These are the outputs of the hidden layers for the forward and backward views, respectively; α t β t These are the respective weights of the hidden layers; b t For bias.

3. The airport runway icing risk early warning method fused with neural networks as described in claim 1, characterized in that: Step S2 specifically includes: determining the model input quantities and performing normalization processing to eliminate the influence of dimensions; the formula is as follows: In the formula, X n The normalized value of the data; x max x min These are the maximum and minimum values ​​in the data, respectively. To ensure that the output results have the same dimensions as the true values, the output data also needs to be denormalized.

4. The airport runway icing risk early warning method fused with neural networks as described in claim 1, characterized in that: Step S4 specifically includes the following sub-steps: S4.1: Population Initialization: The Golden Jackal Optimization Algorithm is a population-based optimization algorithm where the initial solution Y0 is randomly and uniformly distributed across the search space. Y0=Y min +rand(Y max -AND min ) In the formula: Y max Y min , respectively, are the upper and lower limits of the solution; rand is a random variable in the range [0, 1]; Initialize the prey matrix Prey: In the above formula Let d0 be the d0th dimension of the n0th prey; n0 is the number of prey; d0 is the dimension of the variable being solved. During the optimization process, a fitness function is applied to estimate the fitness of each prey, resulting in the fitness matrix F of all prey. OA : In the formula: f is the fitness function; the optimal fitness corresponds to the position coordinates of the male jackal, and the second-best fitness corresponds to the position coordinates of the female jackal; S4.2: Finding Prey The corresponding calculation model is as follows: Y1(t Y )=Y M (t Y )-E·|Y M (t Y )-R L ·Prey(t Y )| Y2(t Y )=Y FM (t Y )-E·|Y FM (t Y )-R L ·Prey(t Y )| In the above formula: t Y R is the current iteration number; E is the escape energy of the prey; L Prey(t) is a random vector based on the Levy distribution. Y Y is the position vector of the prey; M (t Y ) and Y FM (t Y Y1(t) represents the position of the male and female jackals, respectively; Y ) and Y2(t Y () represent the locations of the prey for the male and female jackals in the new iteration; The formula for calculating the escape energy E of the prey is as follows: In the above formula: E1 is the decreasing prey energy; E0 is the initial state of the prey energy; T is the maximum number of iterations; c1 is a constant of 1.5, and E1 decreases linearly from 1.5 to 0 throughout the iteration process; r represents any value between [0,1]. S4.3: Encircle until the prey stops moving: Random vector R based on Levy distribution L The representation is as follows: R L =0.05*LF(y) In the above formula, μ and ν represent random values ​​in [0,1], β = 1.5, and Γ(·) is the gamma function. Update the position of the golden jackal using the following formula: S4.4: Attacking Prey: Its predation model is as follows: Y1(t Y )=Y M (t Y )-E·|Y M (t Y )-R L ·Prey(t Y )| Y2(t Y )=Y FM (t Y )-E·|Y FM (t Y )-R L ·Prey(t Y )| In the above formula: t Y R is the current iteration number; E is the escape energy of the prey; L Prey(t) is a random vector based on the Levy distribution. Y Y is the position vector of the prey; M (t Y ) and Y FM (t Y Y1(t) represents the position of the male and female jackals, respectively; Y ) and Y2(t Y () represent the positions of the prey for the male and female jackals in the new iteration.

Images