Airport visibility prediction method based on deep learning

Abstract

The present invention provides an airport visibility prediction method based on deep learning, including: S1. Acquire the airport's meteorological data, extract the features of each visibility influencing factor, and obtain a feature subset; S2. Based on the airport's minimum takeoff standard and the feature subset, construct an input layer, the input layer is composed of different low visibility intervals; S3. Based on the low visibility interval, perform model training to construct a recurrent neural network to obtain a feature association layer; S4. Based on the feature association layer, fuse the self-attention mechanism and the probabilistic sparse self-attention mechanism respectively, and obtain the self-attention recurrent neural network and the probabilistic sparse self-attention recurrent neural network in turn to complete the construction of the attention layer; S5. Construct a prediction output layer based on the attention layer, the prediction output layer is used to obtain the visibility prediction result by weighted summing the output of the attention layer. The present invention realizes the prediction of low visibility at the airport by integrating the neural network prediction model of the self-attention mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of airport visibility prediction, and in particular to an airport visibility prediction method based on deep learning. Background Art

[0002] Visibility is the maximum horizontal distance that a person with normal vision can see from the weather background. Bad weather is likely to cause reduced visibility at the airport. Reduced visibility can also cause delays in airport flights and flight congestion, causing great difficulties for air traffic control and airport staff, and greatly reducing the efficiency of civil aviation transportation. Therefore, predicting bad weather at the airport, such as thunderstorms, low visibility, turbulence, icing and other weather phenomena, is of great significance to ensure the safe and orderly entry and departure of aircraft.

[0003] In recent years, airport visibility prediction methods include traditional target detection sources, laser instrument detection, and deep learning prediction. Traditional target detection sources and laser instrument detection procedures are complex and labor-intensive. Under specific light conditions, it is difficult to determine the target reference object, and the visibility measurement is not accurate enough. Using deep learning to predict visibility will not be affected by light intensity, nor will it consume too much manpower and material resources. Therefore, using deep learning to predict visibility has become a key technology.

[0004] The existing RNN, LSTM, and GRU face long sequence problems when predicting airport visibility:

[0005] Gradient vanishing and exploding: Although LSTM and GRU have largely alleviated the problem of gradient vanishing or exploding in RNN, this problem may still exist in extremely long sequences, especially for RNN, affecting the model's ability to learn long-term dependencies;

[0006] Computational complexity: When processing long sequences, LSTM and GRU have high computational costs due to their complex gating mechanisms, which may lead to long training times and high demand for computing resources;

[0007] Model interpretability: RNN, LSTM, and GRU models are relatively complex, especially the gating mechanism in LSTM and GRU, which reduces the interpretability of the model and makes it more difficult to understand and debug the model. These shortcomings are mainly caused by the model structure and the essential characteristics of processing sequence data.

[0008] To this end, the present invention attempts a new airport visibility prediction method based on deep learning. Summary of the invention

[0009] The present invention provides an airport visibility prediction method based on deep learning, which can reduce prediction error and improve prediction accuracy.

[0010] This specification discloses an airport visibility prediction method based on deep learning, including:

[0011] S1. Obtain the airport's meteorological data, extract the features of each visibility influencing factor, and obtain a feature subset;

[0012] S2. constructing an input layer based on the airport minimum takeoff standard and the feature subset, wherein the input layer consists of different low visibility intervals;

[0013] S3. Based on the low visibility interval, a model training is performed to construct a recurrent neural network to obtain a feature association layer;

[0014] S4. Based on the feature association layer, the self-attention mechanism and the probabilistic sparse self-attention mechanism are respectively integrated to obtain the self-attention recurrent neural network and the probabilistic sparse self-attention recurrent neural network in turn, thereby completing the construction of the attention layer;

[0015] S5. Construct a prediction output layer based on the attention layer, and the prediction output layer is used to obtain a visibility prediction result by weighted summing the outputs of the attention layer.

[0016] In this specification, the meteorological data includes the airport's historical visibility records, temperature, humidity, wind speed, wind direction, air pressure, cloud cover, precipitation and satellite image data. In S1, the meteorological data is first cleaned and standardized, and missing values ​​are processed by interpolation method, and then the features of various influencing factors of visibility are extracted to obtain the feature subset.

[0017] In this specification, the different low visibility intervals include a first low visibility interval and a second low visibility interval, and the influencing factor features in the feature subset are classified into the first low visibility interval or the second low visibility interval based on the minimum take-off standard of the airport.

[0018] In this specification, based on the influencing factor characteristics of the first low visibility interval and the second low visibility interval, model training is performed to obtain a first recurrent neural network and a second recurrent neural network.

[0019] In this specification, based on the first recurrent neural network, the self-attention mechanism is integrated to obtain a self-attention recurrent neural network; based on the second recurrent neural network, the probabilistic sparse self-attention mechanism is integrated to obtain a probabilistic sparse self-attention recurrent neural network.

[0020] In this specification, the corresponding first fully connected layer and second fully connected layer are respectively constructed based on the self-attention recurrent neural network and the probabilistic sparse self-attention recurrent neural network, the first fully connected layer and the second fully connected layer constitute the prediction output layer, and the output of the prediction output layer is the weighted sum of the outputs of the first fully connected layer and the second fully connected layer.

[0021] In this specification, the process of constructing a recurrent neural network includes:

[0022] First, an RNN network is constructed, and two control gates are added inside the RNN network to construct an LSTM network with long short-term memory; a forget gate is added on the basis of the LSTM network to construct a GRU network;

[0023] Use the forget gate to discard irrelevant information in the historical state, and then use the output of the sigmoid layer to decide whether to retain the historical state information of the previous step; if the output is 1, it means it is fully retained, otherwise it means clearing all the historical state information before this:

[0024] ;

[0025] The input gate learns new information:

[0026] ;

[0027] Updates of the input gate and forget gate to the state unit:

[0028] ;

[0029] ;

[0030] Finally, the output gate determines the output information:

[0031] ;

[0032] .

[0033] In this specification, the outputs of each control gate of the GRU network are:

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] Among them, the S-type function Control the information flow input into the network within (0, 1); hyperbolic tangent function As a nonlinear activation function of a neural network, it can transform the input into the output within (-1, 1). , , , Represent the short-term and long-term states of the previous time series and the current time series respectively, Indicates input, represents dot product, , , , , , are the gating parameter matrices, , , is the state bias of the control.

[0039] In this specification, weight functions are introduced into the self-attention recurrent neural network and the probabilistic sparse self-attention recurrent neural network respectively, so that the weights are related to the input values ​​to prevent the weights in the prediction output layer from being fixed; the weight functions are as follows:

[0040] ;

[0041] ;

[0042] represents the input vector; Indicates bias; Represents the unnormalized attention score, where a higher score indicates a better match between the input and the target; and Both represent weight matrices; is the probability distribution of attention; the weighted sum of the input variables obtained according to the probability of attention distribution is used to calculate the output vector:

[0043] .

[0044] In this specification, based on the self-attention recurrent neural network, KL divergence is used to distinguish the activation distribution and the inert distribution, and the first The approximate probability step-by-step evaluation function of query vectors and corresponding key vectors: ;

[0045] ;

[0046] For the matrix No. OK, For the matrix No. Transpose rows, For the matrix The transposed matrix of is the sample size, , is the number of longs, is the key vector dimension, is the dimension of the input features, is the sparse matrix, which is the same as the matrix have the same number of rows and columns, after sparse indivual Select the first few with larger weights for combination. .

[0047] This manual can at least achieve the following beneficial effects:

[0048] By integrating the neural network prediction model with the self-attention mechanism, we can predict the low visibility of the airport at different time periods in the future. This will help the airport to select the corresponding network model according to the different future prediction step sizes to be output when low visibility weather occurs again, thus achieving more accurate predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use 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 ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 It is a flow chart of the airport visibility prediction method based on deep learning involved in the present invention.

[0051] Figure 2 Schematic diagram of the network structure of the RNN involved in the present invention.

[0052] Figure 3 Schematic diagram of the network structure of the LSTM involved in the present invention.

[0053] Figure 4 Schematic diagram of the network structure of the GRU involved in the present invention.

[0054] Figure 5 It is a schematic diagram of the prediction network flow of the airport visibility prediction method based on deep learning involved in the present invention. DETAILED DESCRIPTION

[0055] In the following, only some exemplary embodiments are briefly described. As those skilled in the art will appreciate, the described embodiments may be modified in various ways without departing from the spirit or scope of the embodiments of the present invention. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.

[0056] The disclosure below provides many different embodiments or examples to implement different structures of the embodiments of the present invention. In order to simplify the disclosure of the embodiments of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the embodiments of the present invention. In addition, the embodiments of the present invention can repeat reference numbers and / or reference letters in different examples, and this repetition is for the purpose of simplification and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed.

[0057] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0058] like Figure 1 As shown, the embodiment of this specification discloses a method for predicting airport visibility based on deep learning, including:

[0059] S1. Obtain the airport's meteorological data, extract the features of each visibility influencing factor, and obtain a feature subset;

[0060] S2. constructing an input layer based on the airport minimum takeoff standard and the feature subset, wherein the input layer consists of different low visibility intervals;

[0061] S3. Based on the low visibility interval, a model training is performed to construct a recurrent neural network to obtain a feature association layer;

[0062] S4. Based on the feature association layer, the self-attention mechanism and the probabilistic sparse self-attention mechanism are respectively integrated to obtain the self-attention recurrent neural network and the probabilistic sparse self-attention recurrent neural network in turn, thereby completing the construction of the attention layer;

[0063] S5. Construct a prediction output layer based on the attention layer, and the prediction output layer is used to obtain a visibility prediction result by weighted summing the outputs of the attention layer.

[0064] In some embodiments of the present specification, the meteorological data includes historical airport visibility records, temperature, humidity, wind speed, wind direction, air pressure, cloud cover, precipitation and satellite image data. In S1, the meteorological data is first cleaned and standardized, and missing values ​​are processed by interpolation method, and then the features of various influencing factors of visibility are extracted to obtain the feature subset.

[0065] In some embodiments of the present specification, the different low visibility intervals include a first low visibility interval and a second low visibility interval, and the influencing factor features in the feature subset are classified into the first low visibility interval or the second low visibility interval based on the airport minimum take-off standard.

[0066] In some embodiments of the present specification, based on the influencing factor characteristics of the first low visibility interval and the second low visibility interval, model training is performed to obtain a first recurrent neural network and a second recurrent neural network, respectively.

[0067] In some embodiments of the present specification, based on the first recurrent neural network, a self-attention mechanism is integrated to obtain a self-attention recurrent neural network; based on the second recurrent neural network, a probabilistic sparse self-attention mechanism is used to obtain a probabilistic sparse self-attention recurrent neural network.

[0068] In some embodiments of the present specification, a corresponding first fully connected layer and a second fully connected layer are respectively constructed based on the self-attention recurrent neural network and the probabilistic sparse self-attention recurrent neural network, and the first fully connected layer and the second fully connected layer constitute the prediction output layer, and the output of the prediction output layer is the weighted sum of the outputs of the first fully connected layer and the second fully connected layer.

[0069] In some embodiments of this specification, the process of constructing a recurrent neural network includes:

[0070] First, an RNN network is constructed, and two control gates are added inside the RNN network to construct an LSTM network with long short-term memory; a forget gate is added on the basis of the LSTM network to construct a GRU network;

[0071] Use the forget gate to discard irrelevant information in the historical state, and then use the output of the sigmoid layer to decide whether to retain the historical state information of the previous step; if the output is 1, it means it is fully retained, otherwise it means clearing all the historical state information before this:

[0072] ;

[0073] The input gate learns new information:

[0074] ;

[0075] Updates of the input gate and forget gate to the state unit:

[0076] ;

[0077] ;

[0078] Finally, the output gate determines the output information:

[0079] ;

[0080] .

[0081] In some embodiments of this specification, the outputs of each control gate of the GRU network are respectively:

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] Among them, the S-type function Control the information flow input into the network within (0, 1); hyperbolic tangent function As a nonlinear activation function of a neural network, it can transform the input into the output within (-1, 1). , , , Represent the short-term and long-term states of the previous time series and the current time series respectively, Indicates input, represents dot product, , , , , , are the gating parameter matrices, , , is the state bias of the control.

[0087] In some embodiments of the present specification, weight functions are introduced into the self-attention recurrent neural network and the probabilistic sparse self-attention recurrent neural network respectively, so that the weights are related to the input values ​​to prevent the weights in the prediction output layer from being fixed; the weight functions are as follows:

[0088] ;

[0089] ;

[0090] represents the input vector; Indicates bias; Represents the unnormalized attention score, where a higher score indicates a better match between the input and the target; and Both represent weight matrices; is the probability distribution of attention; the weighted sum of the input variables obtained according to the attention distribution probability is used to calculate the output vector: .

[0091] In some embodiments of this specification, based on the self-attention recurrent neural network, KL divergence is used to distinguish the activation distribution and the inert distribution, and the first The approximate probability step-by-step evaluation function of query vectors and corresponding key vectors: ;

[0092] ;

[0093] For the matrix No. OK, For the matrix No. Transpose rows, For the matrix The transposed matrix of is the sample size, , is the number of longs, is the key vector dimension, is the dimension of the input features, is the sparse matrix, which is the same as the matrix have the same number of rows and columns, after sparse indivual Select the first few with larger weights for combination. .

[0094] The technical concept of the present invention is as follows:

[0095] 1. Algorithm principles and construction of RNN, LSTM and GRU models

[0096] The introduction of RNN and LSTM strengthens the internal relationship between input sequences. Recurrent neural networks strengthen historical memory of sequences by constructing hidden layers between sequences, so that the nodes after the sequence contain valid information from the previous section in the next calculation. Therefore, recurrent neural networks play an important role in capturing internal connections of information, speech recognition, and machine translation. The structure of RNN is shown in the attached figure. Figure 2 As shown, U‌ represents the weight matrix from the input layer to the hidden layer, ‌V‌ represents the weight matrix from the hidden layer to the output layer, ‌W‌ represents the weight matrix from the hidden layer to itself, and ‌O‌ represents the value of the output layer.

[0097] In addition, with the development of deep learning, two control gates were added to the RNN to construct a recurrent neural network with long short-term memory (LSTM), which optimized the problem of gradient vanishing in the RNN network. Compared with the input and output of the network, the newly added forget gate, as the most important unit gate in the LSTM, makes the LSTM structure more robust after the activation functions of each gate. The network structure diagram is shown in the attached figure. Figure 3 shown.

[0098] First, the irrelevant information in the historical state is discarded by the forget gate, and then the output of the sigmoid layer is used to decide whether to retain the historical state information of the previous step; if the output is 1, it means that it is completely retained, otherwise, it means that all historical state information before this is cleared:

[0099] ; (1)

[0100] The input gate learns new information:

[0101] ; (2)

[0102] Updates of the input gate and forget gate to the state unit:

[0103] ; (3)

[0104] ; (4)

[0105] Finally, the output gate determines the output information:

[0106] ; (5)

[0107] ; (6)

[0108] First, the forget gate discards information that is irrelevant to the historical state, retains useful information, and receives the current input information. And the hidden state information of the previous moment , then concatenate the two pieces of information and multiply them by the weight , add bias Finally, the output of the sigmoid layer determines whether to retain the historical state information of the previous step; if the output is 1, it means that it is completely retained, otherwise, it means that all historical state information before this is cleared. Secondly, after the forget gate is the input gate responsible for updating the cell state, which transmits the current input information and the hidden state information of the previous moment respectively. To the Sigmoid function and the Tanh function, and then multiply the output values ​​of the two functions. The purpose is to use the Sigmoid function to make a selection of the output values ​​of the Tanh function. Finally, the cell state updated by the addition information processed by the forget gate and the input gate is output. , the output gate takes the current input information And the hidden state information of the previous moment Passed to the Sigmoid function, the current cell state is passed to the Tanh function, as shown below, the output values ​​of the Sigmoid function and the Tanh function are multiplied to obtain the new hidden state information.

[0109] Then, in order to enable the constructed network to continuously learn the time series and abandon some invalid information in the previous section, a forget gate was added on the basis of LSTM to construct the GRU network model. This model also merges long-term states and short-term states. Compared with the LSTM network, the reduction in parameters makes the network converge faster during training. In some cases, GRU will perform better than LSTM. Figure 4 For the GRU network structure, the outputs of each control gate are:

[0110] ; (7)

[0111] ; (8)

[0112] ; (9)

[0113] ; (10)

[0114] Among them, the S-type function Control the information flow input into the network within (0, 1); hyperbolic tangent function As a nonlinear activation function of a neural network, it can transform the input into the output within (-1, 1). , , , Represent the short-term and long-term states of the previous time series and the current time series respectively, Indicates input, represents dot product, , , , , , are the gating parameter matrices, , , is the state bias of the control.

[0115] 2. Self-attention mechanism and probabilistic sparse self-attention mechanism

[0116] The self-attention mechanism improves the traditional attention mechanism and adaptively weights the current input information to discover the dependency characteristics within the sequence. Compared with the fully connected neural network layer, the self-attention mechanism is much less dependent on external information, so it can efficiently solve the problem of long sequence prediction and compile the generated feature information of different connection weights into a sequence of fixed size. The key to this mechanism is to introduce a weight function into the network so that the weight is related to the input value to prevent the problem of fixed weights in the fully connected layer. The weight function is as follows:

[0117] ; (11)

[0118] ; (12)

[0119] Where: represents the input vector; Indicates bias; Represents the unnormalized attention score, where a higher score indicates a better match between the input and the target; and Both represent weight matrices; is the probability distribution of attention. Finally, the weighted sum of the input variables obtained according to the attention distribution probability is calculated to calculate the output vector:

[0120] ; (13)

[0121] However, the time and space computational complexity of the self-attention mechanism is high, and the attention weight distribution corresponding to different query values ​​is not focused. If you want to reduce the complexity, you only need to calculate the self-attention that has a greater impact on the result. Studies have found that the probability distribution of the self-attention mechanism has potential sparsity. Without affecting the performance of the model, a probabilistic strategy with selection count can be designed. KL divergence is used to distinguish between activation distribution and inert distribution, and the first The approximate probability step-by-step evaluation function of query vectors and corresponding key vectors:

[0122] ; (14)

[0123] ; (15)

[0124] For the matrix No. OK, For the matrix No. Transpose the rows, For the matrix The transposed matrix of is the sample size, , is the number of longs, is the key vector dimension, is the dimension of the input features, is the sparse matrix, which is the same as the matrix have the same number of rows and columns, after sparse indivual Select the first few with larger weights for combination. .

[0125] 3. Implementation of prediction model after integrating attention mechanism

[0126] The process of RNN, LSTM and GRU prediction network based on the fusion of self-attention mechanism is shown in the attached figure. Figure 5 As shown in the figure. A-RNN, A-LSTM, A-GRU after integrating the self-attention mechanism and P-RNN, P-LSTM, P-GRU after integrating the probabilistic sparse self-attention mechanism are based on the message data sample set, and the input layer is the processed two-interval sample set; the feature association layer is composed of the traditional recurrent neural network, which predicts the airport visibility at future times by mining the temporal correlation between low visibility and other meteorological factor samples at the corresponding time; the attention layer (Attention) improves the accuracy of the prediction model by strengthening the relevant attention time series features between low visibility and the features between low visibility and other influencing factors; finally, the prediction output layer is a fully connected layer, which realizes the visibility prediction for the next 1-3 hours by weighted summing the hidden states output by the attention layer.

[0127] 4. Model performance evaluation indicators

[0128] In order to objectively evaluate the performance of the models used in this paper, mean absolute error (MAE), root mean square error (RMSE) and symmetric mean absolute percentage error (SMAPE) are selected as evaluation indicators for the above 9 models. MAE reflects the actual situation of the prediction output error; RMSE is more sensitive to values ​​with large errors and can evaluate its prediction accuracy; SMAPE is used to calculate the relative error of the model. This indicator can avoid the problem of formula failure caused by the true value being too small. The smaller the values ​​of the above three indicators, the better the prediction effect of the model. The calculation formulas of the three indicators are as follows:

[0129] ; (16)

[0130] ; (17)

[0131] ; (18)

[0132] Where: n is the total sample size of the test set; is the visibility forecast value; is the true value of visibility.

[0133] Based on pytorch, the traditional neural network and the neural network model after fusion of attention were built, with a total of 9 models. The internal hyperparameters of the network were set according to the characteristics of the model, and the model training speed and prediction result accuracy were adjusted. The hyperparameters of the 9 models were finally determined as shown in Table 1. The training was carried out in intervals in the low visibility training samples. After predicting the visibility for the next 1-3 hours, the predictions were made on the test sets respectively, and the output results were inversely normalized to obtain the predicted values ​​of visibility.

[0134]

[0135] 5. Evaluation index analysis

[0136] The comparison results of MAE, RMSE and SMAPE of the 9 models are shown in Table 2. It can be found that when predicting the low visibility interval, whether it is the fusion of self-attention mechanism or probabilistic sparse self-attention mechanism, the prediction error is smaller and the accuracy is higher than that of the traditional recurrent neural network.

[0137]

[0138] Note: The best experimental result is displayed in bold black font, and the result with the largest error is displayed in red font with shading.

[0139] 6. Visual analysis of prediction results

[0140] Through the analysis of the results of each evaluation index in Table 2, the optimal model, the worst model and the real message value are selected for visualization analysis. It can be seen from the table that when predicting the future 1h, the difference between the predicted value and the real value of A-GRU with self-attention mechanism is very small, and it has good prediction performance. In addition, when predicting 2h and 3h, P-RNN with probabilistic sparse self-attention mechanism and A-RNN with self-attention mechanism are also selected with high prediction accuracy. Similarly, P-GRU, P-RNN, and A-GRU are respectively used as the optimal prediction models for the future 1-3h in the second interval, which have significantly improved the prediction ability compared with the traditional neural network without the introduction of attention mechanism.

[0141] In summary, the purpose of this application is to optimize the existing deep learning prediction model of visibility. This model significantly improves the prediction accuracy under low visibility conditions by integrating the self-attention mechanism into the traditional RNN, LSTM, and GRU networks. The specific innovation points are summarized as follows:

[0142] 1. Innovative design of dynamic self-attention fusion layer: This application proposes a new dynamic self-attention mechanism layer, which can automatically assign weights according to the importance of features at different time steps of the input sequence. On the basic architecture of RNN, LSTM, and GRU, this mechanism is integrated inside the recurrent unit or as an additional layer to ensure that the model can focus more on historical meteorological features that are critical to visibility prediction when processing sequence data, especially in low visibility conditions, significantly enhancing the model's ability to capture key information.

[0143] 2. Multi-scale feature extraction and fusion strategy: In order to further improve the prediction accuracy, multi-scale feature extraction technology is adopted. By combining short-term local features with long-term global trends, the self-attention mechanism is used to flexibly allocate attention weights at different time scales, thereby achieving effective recognition of complex meteorological patterns, especially in subtle changes in low visibility conditions.

[0144] 3. Adaptive learning rate adjustment mechanism: In view of the particularity of the visibility prediction task, the present invention introduces the Adam optimizer algorithm to dynamically adjust the learning rate of the network during the training process, especially after integrating the self-attention mechanism, which effectively accelerates the model convergence speed while avoiding overfitting, ensuring that the model can still maintain stable and reliable prediction performance in the face of extreme low visibility conditions.

[0145] 4. Lightweight and efficient optimization: Taking into account the resource limitations in actual deployment, this application also optimizes the model structure, reduces the number of model parameters, and maintains the effectiveness of the self-attention mechanism, thereby achieving high prediction accuracy while reducing computational complexity and memory usage, making it more suitable for real-time airport visibility prediction systems.

[0146] In summary, this application provides a more accurate, efficient and adaptable solution for airport low visibility prediction by integrating the attention mechanism into the RNN, LSTM and GRU networks and combining multi-scale feature extraction and optimization strategies, which is expected to play an important role in aviation safety management and operations.

[0147] The above-described embodiments are used to illustrate the present invention, not to limit the present invention, so changes in the exemplified values ​​or replacement of equivalent elements should still fall within the scope of the present invention.

Claims

1. The airport visibility prediction method based on deep learning is characterized by: include: S1. Obtain the airport's meteorological data, extract the features of each visibility influencing factor, and obtain a feature subset; S2. constructing an input layer based on the airport minimum takeoff standard and the feature subset, wherein the input layer consists of different low visibility intervals; S3. Based on the low visibility interval, construct a recurrent neural network and perform model training to obtain a feature association layer; S4. Based on the feature association layer, the self-attention mechanism and the probabilistic sparse self-attention mechanism are respectively integrated to obtain the self-attention recurrent neural network and the probabilistic sparse self-attention recurrent neural network in turn, thereby completing the construction of the attention layer; S5. constructing a prediction output layer based on the attention layer, wherein the prediction output layer is used to obtain a visibility prediction result by weighted summing the outputs of the attention layer; The meteorological data includes the airport's historical visibility records, temperature, humidity, wind speed, wind direction, air pressure, cloud cover, precipitation, and satellite image data. In S1, the meteorological data is first cleaned and standardized, and missing values ​​are processed by an interpolation method, and then the features of various visibility influencing factors are extracted to obtain the feature subset; The different low visibility intervals include a first low visibility interval and a second low visibility interval, and the influencing factor features in the feature subset are classified into the first low visibility interval or the second low visibility interval based on the minimum take-off standard of the airport; Based on the influencing factor characteristics of the first low visibility interval and the second low visibility interval, model training is performed to obtain a first recurrent neural network and a second recurrent neural network respectively; Based on the first recurrent neural network, a self-attention mechanism is integrated to obtain a self-attention recurrent neural network; based on the second recurrent neural network, a probabilistic sparse self-attention mechanism is integrated to obtain a probabilistic sparse self-attention recurrent neural network; Based on the self-attention recurrent neural network and the probabilistic sparse self-attention recurrent neural network, a corresponding first fully connected layer and a second fully connected layer are respectively constructed, the first fully connected layer and the second fully connected layer constitute the prediction output layer, and the output of the prediction output layer is the weighted sum of the outputs of the first fully connected layer and the second fully connected layer.

2. The airport visibility prediction method based on deep learning according to claim 1, characterized in that: The process of constructing a recurrent neural network includes: First, an RNN network is constructed, and two control gates are added inside the RNN network to construct an LSTM network with long short-term memory; a forget gate is added on the basis of the LSTM network to construct a GRU network; Use the forget gate to discard irrelevant information in the historical state, and then use the output of the sigmoid layer to decide whether to retain the historical state information of the previous step; if the output is 1, it means it is fully retained, otherwise it means clearing all the historical state information before this: ； The input gate learns new information: ； Updates of the input gate and forget gate to the state unit: ； ； Finally, the output gate determines the output information: ； ； Represents the short-term state of the current time series, , Represent the long-term status of the previous time series and the current time series respectively, Indicates input, represents dot product, , , are the gating parameter matrices, , , is the state bias of the control.

3. The airport visibility prediction method based on deep learning according to claim 2 is characterized in that: The outputs of each control gate of the GRU network are: ； ； ； ； Among them, the S-type function Control the information flow input into the network within (0, 1); hyperbolic tangent function As a nonlinear activation function of a neural network, it can transform the input into the output within (-1, 1). Represents the short-term state of the previous time series, Indicates input, represents dot product, , , are the gating parameter matrices respectively.

4. The airport visibility prediction method based on deep learning according to claim 1, characterized in that: A weight function is introduced into the self-attention recurrent neural network and the probabilistic sparse self-attention recurrent neural network respectively, so that the weight is related to the input value to prevent the weight in the prediction output layer from being fixed; the weight function is as follows: ； ； represents the input vector; Indicates bias; Represents the unnormalized attention score, where a higher score indicates a better match between the input and the target; and Both represent weight matrices; is the probability distribution of attention; the weighted sum of the input variables obtained according to the probability of attention distribution is used to calculate the output vector: 。 5. The airport visibility prediction method based on deep learning according to claim 1, characterized in that: Based on the self-attention recurrent neural network, KL divergence is used to distinguish the activation distribution and the inert distribution, and the first The approximate probability step-by-step evaluation function of query vectors and corresponding key vectors: ; ； For the matrix No. OK, For the matrix No. Transpose rows, For the matrix The transposed matrix of is the sample size, , is the number of longs, is the key vector dimension, is the dimension of the input features, is the sparse matrix, which is the same as the matrix Have the same number of rows and columns.

Images