Flight delay time prediction method and system based on dynamic gating MoE

Abstract

The invention relates to a flight delay time prediction method and system based on dynamic gating MoE, and the method comprises the following steps: collecting flight historical operation data, carrying out the data preprocessing, and constructing a prediction data set; performing feature vectorization coding on the prediction data set in the form of a time sequence matrix, and extracting time sequence features of the prediction data set by using a multi-head attention mechanism; inputting the time sequence characteristics of the prediction data set into a dynamic gating MoE, adaptively selecting and combining the outputs of the routing expert sub-model and the shared expert sub-model through a dynamic gating mechanism, and obtaining fusion characteristics; and generating flight delay time prediction results according to the fusion features, wherein the flight delay time prediction results comprise departure delay time prediction results and arrival delay time prediction results. Compared with the prior art, the method has the advantages that the accuracy and stability of departure and arrival delay time prediction are remarkably improved, and the adaptability and expandability of the model to the complex flight operation situation are enhanced.

Description

Technical Field

[0001] This invention relates to the field of flight delay prediction, and in particular to a method and system for predicting flight delay time based on dynamic gating (MoE). Background Technology

[0002] In recent years, the global air transport industry has experienced rapid and sustained growth, with increasing complexity and density of air networks, making flight delays a more prominent issue. Delays not only directly impact passenger travel experience but also increase operating costs for airlines and airports, disrupt resource allocation, and may trigger a chain reaction of delays, significantly affecting the operational efficiency of the entire transportation system. Accurate and timely prediction of flight delays is of significant practical value for passenger planning, airport resource allocation, and airline emergency response.

[0003] Currently, research and engineering practices targeting flight delays mainly employ three types of methods: statistical models, traditional machine learning models, and deep learning models. Statistical methods (such as ARIMA and its variants) are good at characterizing periodic and linear trends but struggle to handle highly nonlinear and complex factors. Traditional machine learning methods (such as decision trees, random forests, and XGBoost) perform well in processing structured features, but their ability to express high-dimensional temporal dependencies and complex interaction patterns remains limited. Deep learning methods (such as LSTM, CNN, and Transformer) have made significant progress in modeling complex temporal and space-time dependencies, but suffer from insufficient interpretability and limited scalability. For example, patent application CN113128769A discloses a deep learning-based intelligent flight delay prediction method. This method uses a single deep learning model (such as CNN or LSTM) to model historical flight operation data to predict flight delays. This method relies on the fixed structure of a single model for feature learning and prediction, making it difficult to adaptively handle the complex features of multi-source, heterogeneous, and nonlinear flight delays. This results in limitations in the model's expressive power, generalization ability, and scalability when dealing with diverse data distributions and complex interaction patterns, making it prone to getting trapped in local optima. These limitations mean that existing technologies still struggle to simultaneously meet engineering requirements such as high accuracy, interpretability, and scalability when facing flight delay prediction problems with multiple data and multi-modal features. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a flight delay time prediction method and system based on dynamic gating (MoE), which significantly improves the accuracy and stability of departure and arrival delay time prediction, and enhances the model's adaptability and scalability to complex flight operation situations.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for predicting flight delay times based on dynamic gated MoE includes the following steps: Collect historical flight operation data and perform data preprocessing to construct a prediction dataset; The prediction dataset is encoded as a time series matrix using feature vectorization, and the temporal features of the prediction dataset are extracted using a multi-head attention mechanism. The temporal features of the prediction dataset are input into the dynamic gating MoE. The outputs of the routing expert sub-model and the shared expert sub-model are adaptively selected and combined through the dynamic gating mechanism to obtain the fused features. The dynamic gating MoE consists of at least one shared expert sub-model and multiple routing expert sub-models. Flight delay time prediction results are generated based on the fusion features. The flight delay time prediction results include departure delay time prediction results and arrival delay time prediction results.

[0006] Furthermore, the historical flight operation data includes flight number, departure point, destination, planned departure time, planned arrival time, actual departure time, actual arrival time, preceding flight delay information, and other factors related to flight operation.

[0007] Furthermore, the specific steps for collecting historical flight operation data, performing data preprocessing, and constructing a prediction dataset include: Data cleaning was performed on historical flight operation data to obtain a valid dataset; The effective dataset is encoded by using label encoding to convert discrete data (excluding time) in the effective dataset into numerical representation, using cyclic encoding to spatially map the time in the effective dataset, and normalizing the continuous data in the effective dataset. Delay tags are obtained based on historical flight operation data; The prediction dataset is obtained by combining the delay labels and the valid dataset after data encoding.

[0008] Furthermore, the delay label includes a category label for earlier arrivals on the same day and a numerical label for the delay of preceding flights. The category label for earlier arrivals on the same day indicates whether there were any delays of flights on the same route before the current flight departed on the same day. The numerical label for the delay of preceding flights indicates the delay time of the previous flight segment performed by the aircraft operating the current flight before arriving at the airport.

[0009] Furthermore, the specific steps of encoding the prediction dataset as a time series matrix using feature vectorization and extracting the temporal features of the prediction dataset using a multi-head attention mechanism include: The prediction dataset is grouped by flight number and sorted by time to obtain time series data; The time series data is mapped to a hidden dimension and then standardized using RMS Norm. The standardized time-series data is encoded with positional codes, and time-series features are extracted using a multi-head attention module. This module generates query vectors, key vectors, and value vectors to calculate attention weights between different time steps within the time-series data, thereby capturing key temporal dependencies. The time-series features are as follows: In the formula, As a time series feature, It is time series data. For standardized time-series data, For multi-head attention modules, For the output of the multi-head attention module, For splicing operations, In the multi-head attention module Output of size To output the projection matrix, For attention mechanisms, For input multi-head attention module number The query vector of the head. For input multi-head attention module number The key vector of the head, For input multi-head attention module number The value vector of the head, For the multi-head attention module The dimensions of the head.

[0010] Further, the temporal features of the predicted dataset are input into the dynamically gated MoE, and the outputs of the routing expert sub-model and the shared expert sub-model are adaptively selected and combined through the dynamic gating mechanism to obtain the fused features. The specific steps include: The temporal features of the prediction dataset are standardized using RMS Norm and then input into the dynamic gated MoE. Based on the standardized temporal characteristics, the score of each routing expert in the routing expert sub-model is obtained through a gating network, and the K routing experts with the highest scores are selected for activation using a Top-K coefficient gating strategy. The scores of the K highest-scoring routing experts are normalized using the Softmax activation function to obtain the activation weight of each routing expert. The temporal features of the predicted dataset are input into the K routing experts with the highest scores to obtain the expression of the temporal features by each routing expert; The output of the routing expert sub-model is obtained based on the expression of temporal features and the corresponding activation weights of each routing expert; The temporal features of the prediction dataset are input into each shared expert in the shared expert sub-model to obtain the output of the shared expert sub-model; The output of the routing expert sub-model and the output of the shared expert sub-model are combined to obtain the output of the dynamic gating MoE; The output of the dynamically gated MoE is residually connected and added to the input to obtain the fused feature.

[0011] Furthermore, the fusion feature is as follows: In the formula, As a feature of fusion, As a time series feature, For the time series characteristics after standardization, For dynamic gating MoE, For the output of the dynamically gated MoE, This is the output of the routing expert sub-model. To share the output of the expert sub-model, The number of routing expert sub-models, Activate the weight for experts, For the first The representation of temporal features by each routing expert sub-model To share the number of expert sub-models, For the first The representation of temporal features by a shared expert sub-model.

[0012] Furthermore, the routing experts employ a Transformer Encoder structure, and the specific steps for obtaining each routing expert's representation of the temporal features by inputting the temporal features of the prediction dataset into the K highest-scoring routing experts include: The temporal features are position-encoded to obtain a position-encoded vector. The position-encoded vector is then added element-wise to the temporal features to obtain a sequence representation containing positional information. The multi-head attention layer captures the dependencies between different time steps within the sequence, and obtains the output of the multi-head attention layer. The output of the multi-head attention layer is coupled with the sequence representation containing positional information through residual connection and normalization to obtain the normalized result; The normalization result is input into a fully connected feedforward neural network to enhance the features at each time step, and the output of the fully connected feedforward neural network is obtained. The output and normalized result of the fully connected feedforward neural network are subjected to residual connection and normalization processing to complete the processing of a single Transformer Encoder structure; After processing by an N-layer stacked Transformer Encoder structure, each routing expert's expression of the timing characteristics is obtained.

[0013] Furthermore, based on the predicted departure delay time and arrival delay time, the predictive capability of the dynamic gating system (MoE) is evaluated from three aspects: the magnitude of the prediction error, the error distribution, and the accuracy of the delay classification.

[0014] According to another aspect of the present invention, a flight delay time prediction system based on dynamic gating (MoE) is provided, comprising: The prediction dataset building module is used to collect historical flight operation data, perform data preprocessing, and build a prediction dataset. The temporal feature extraction module is used to encode the prediction dataset in the form of a time series matrix using feature vectorization, and to extract the temporal features of the prediction dataset using a multi-head attention mechanism. The fusion feature acquisition module is used to input the temporal features of the prediction dataset into the dynamic gating MoE, and adaptively select and combine the outputs of the routing expert sub-model and the shared expert sub-model through the dynamic gating mechanism to obtain the fusion features. The dynamic gating MoE consists of at least one shared expert sub-model and multiple routing expert sub-models. The flight delay time prediction module is used to generate flight delay time prediction results based on the fusion features. The flight delay time prediction results include departure delay time prediction results and arrival delay time prediction results.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces a multi-head attention mechanism-based temporal feature extraction and dynamic gating (MoE). First, it adaptively extracts multi-dimensional temporal features from flight operation data with weights. Then, it uses a Top-K gating network to dynamically activate the most relevant routing expert sub-models to obtain the output of the routing experts. This output is then weighted and fused with the output of the shared experts. This solves the problem that traditional prediction methods, due to their fixed structure, are not sensitive enough to capturing chain reactions and multi-factor coupling effects of flight delays in complex operating environments, resulting in insufficient prediction accuracy. This significantly improves the accuracy and stability of departure and arrival delay time prediction.

[0016] 2. This invention constructs a MoE architecture with multi-expert division of labor and dynamic routing mechanism. The gating network calculates the scores of each expert in real time based on the input features and implements load balancing constraints. When new influencing factors need to be included, only the number of corresponding expert networks needs to be increased without reconstructing the entire model. This solves the problem that existing deep learning models have fixed model architectures under variable operating conditions and are difficult to flexibly adjust according to different prediction scenarios. It significantly enhances the model's adaptability to complex flight operation situations and the system's scalability.

[0017] 3. This invention introduces a correlation analysis mechanism between expert activation weights and input features, and uses Pearson correlation coefficient to generate an expert-feature correlation heatmap, dynamically revealing the sensitivity and division of labor logic of different experts to key influencing factors of flight delays. This solves the problems of opaque decision-making process and low reliability of results in delay prediction of traditional deep learning models, and significantly improves the interpretability and decision support capability of prediction models in air scheduling applications. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a flight delay time prediction method based on dynamic gated MoE proposed in this invention. Figure 2 This is a model structure diagram of a dynamically gated MoE; Figure 3 A heatmap showing the correlation between input features and expert scores; Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0020] The following English abbreviations are involved: Mixture of Experts (MoE) Multi-Head Attention (MHA) Root Mean Square Normalization (RMS Normalization) Example 1 This embodiment provides a method for predicting flight delay times based on dynamic gating (MoE), such as... Figure 1 As shown, it includes the following steps: S1. Collect historical flight operation data and perform data preprocessing to construct a prediction dataset.

[0021] Historical flight operation data includes flight number, departure point, destination, scheduled departure time, scheduled arrival time, actual departure time, actual arrival time, preceding flight delay information, and other factors related to flight operation.

[0022] The system retrieves historical airport data, including date information, flight number, aircraft type, origin and destination, scheduled departure time, scheduled arrival time, actual departure time, actual arrival time, number of passengers and baggage, etc., to obtain flight data.

[0023] The specific steps for collecting historical flight operation data, performing data preprocessing, and constructing a prediction dataset include: The historical flight operation data was cleaned to obtain a valid dataset. Missing values ​​and values ​​with fewer than 100 flight numbers were removed to construct reliable time-series features. Ultimately, 289 sets of flight numbers, totaling 50,364 flight records, were retained as the valid dataset.

[0024] The valid dataset is encoded using label encoding. Discrete data, excluding time (e.g., flight number, departure point, destination, and aircraft type), is converted into numerical representations. For time features in the dataset, including month, day, hour, and minute, cyclic encoding is used to map these features to a two-dimensional space of sine and cosine functions. For continuous data, normalization is used to map its feature values ​​to... Within the range Delay tags are obtained based on historical flight operation data. By analyzing historical delay data, delay tags at the route level, airport level, and preceding flight level are generated. After correlation analysis of flight delay time and delay tags, only the category tag for earlier arrivals on the same day and the numerical tag for preceding flight delays are retained. The category tag for earlier arrivals on the same day indicates whether there were any delays of flights on the same route before the current flight departed on the same day. The numerical tag for preceding flight delays indicates the delay time of the previous flight segment performed by the aircraft operating the current flight before arriving at the airport.

[0025] S2. Encode the prediction dataset as a time series matrix using feature vectorization, and extract the temporal features of the prediction dataset using a multi-head attention mechanism.

[0026] The specific steps for encoding the prediction dataset as a time series matrix using feature vectorization and extracting the temporal features of the prediction dataset using a multi-head attention mechanism include: The prediction dataset is grouped by flight number and sorted by time, and then time series data is formed with a historical time step of 14 and a prediction time step of 1. After a fully connected layer, the input dimension of the time series data is mapped to the hidden dimension of the model to obtain... Standardization is performed using RMS Norm. Standardized time series data After adding positional encoding, the input passes through a multi-head attention module. The multi-head attention module calculates the attention weights between different time steps within the time series data by generating query vectors, key vectors, and value vectors, in order to capture key temporal dependencies in the time series data.

[0027] MHA divides the attention mechanism into k parallel heads, each processing a different subspace representation of the input. In the multi-head attention module... The output for each head is: The final output is obtained by concatenating the representations of all the headers and performing a linear transformation: The output MHA of the multi-head attention module ( By performing a residual connection and adding it to the input, gradient vanishing can be avoided, thus obtaining the temporal features. The formula is as follows: In the formula, As a time series feature, As a hidden dimension of time series data, For standardized time-series data, For multi-head attention modules, For the output of the multi-head attention module, For splicing operations, In the multi-head attention module Output of size To output the projection matrix, For attention mechanisms, For input multi-head attention module number The query vector of the head. For input multi-head attention module number The key vector of the head, For input multi-head attention module number The value vector of the head, For the multi-head attention module The dimensions of the head.

[0028] S3. Input the temporal features of the prediction dataset into the dynamic gating MoE. The outputs of the routing expert sub-model and the shared expert sub-model are adaptively selected and combined through the dynamic gating mechanism to obtain the fused features.

[0029] The Dynamically Gated (MoE) consists of at least one shared expert sub-model and multiple routing expert sub-models, and its structure is as follows: Figure 2 As shown. The specific steps for inputting the temporal features of the prediction dataset into the dynamically gated MoE, and adaptively selecting and combining the outputs of the routing expert sub-model and the shared expert sub-model through the dynamic gating mechanism to obtain the fused features include: The temporal features of the prediction dataset are standardized using RMS Norm and then input into the dynamic gated MoE. The score of each routing expert in the routing expert sub-model is obtained through a gating network based on the standardized time-series characteristics. The Top-K coefficient gating strategy is used to select the K highest-scoring routing experts for activation. The scores of the K highest-scoring routing experts are normalized using the Softmax activation function to obtain the expert activation weights for each expert. ; The temporal features of the prediction dataset are input into the K highest-scoring routing experts to obtain each routing expert's expression of the temporal features; The output of the routing expert sub-model is obtained based on each routing expert's representation of the temporal features and the corresponding activation weights. The formula is as follows: In the formula, This is the output of the routing expert sub-model. The number of routing expert sub-models, Activate the weight for experts, For the first The representation of temporal characteristics by a routing expert sub-model.

[0030] The temporal features of the prediction dataset are input into each shared expert in the shared expert sub-model to obtain the output of the shared expert sub-model. The formula is as follows: In the formula, To share the output of the expert sub-model, To share the number of expert sub-models, For the first The representation of temporal features by a shared expert sub-model.

[0031] The outputs of the routing expert sub-model and the shared expert sub-model are combined to obtain the output of the dynamic gating MoE. The formula is as follows: In the formula, For the time series characteristics after standardization, For dynamic gating MoE, For the output of the dynamically gated MoE, This is the output of the routing expert sub-model. To share the output of the expert sub-model.

[0032] The output of the dynamically gated MoE is residually connected and added to the input to obtain the fused features. The formula is as follows: In the formula, As a feature of fusion, As a time series feature, For the time series characteristics after standardization, For dynamic gating MoE, This is the output of the dynamically gated MoE.

[0033] To address the issue of unbalanced expert load, a load balancing strategy without auxiliary loss is employed. When selecting the K highest-scoring routing experts, a bias is added to each expert. ,get The formula is as follows: The bias is dynamically adjusted based on the expert's load, using the following formula: In the formula, For the bias update rate, This represents the average number of times all experts are selected under ideal conditions. Let r be the number of times the r-th routing expert is selected. The sign function for load deviation.

[0034] The routing experts employ a Transformer Encoder structure, inputting the temporal features of the prediction dataset into the K highest-scoring routing experts respectively. The specific steps to obtain each routing expert's representation of the temporal features include: The temporal features are position-encoded to obtain a position-encoded vector. The position-encoded vector is then added element-wise to the temporal features to obtain a sequence representation containing positional information. The multi-head attention layer captures the dependencies between different time steps within the sequence, and obtains the output of the multi-head attention layer. The output of the multi-head attention layer is residually connected and normalized with the sequence representation containing positional information to obtain the normalized result; The normalized result is input into a fully connected feedforward neural network to enhance the features at each time step, and the output of the fully connected feedforward neural network is obtained. The output and normalized result of the fully connected feedforward neural network are subjected to residual connection and normalization processing to complete the processing of a single Transformer Encoder structure; After processing by an N-layer stacked Transformer Encoder structure, each routing expert's expression of the timing characteristics is obtained.

[0035] S4. Generate flight delay time prediction results based on the fusion features. The flight delay time prediction results include departure delay time prediction results and arrival delay time prediction results.

[0036] The hidden dimension is mapped to the prediction dimension through a fully connected layer in the output layer, and the Tanh activation function is used to generate the prediction results for future prediction time steps. The flight delay time prediction results include departure delay time prediction results and arrival delay time prediction results.

[0037] Evaluation metrics were calculated separately for departures and arrivals: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Delay Classification Accuracy. The predictive ability of the dynamically gated MoE was evaluated from three aspects: the magnitude of the prediction error, the error distribution, and the accuracy of delay classification. MAE reflects the average deviation between the predicted and actual values, measured in minutes; a smaller value indicates higher prediction accuracy. RMSE is more sensitive to larger errors; a larger value indicates the presence of significant outlier predictions. Accuracy reflects the model's ability to determine the delay status.

[0038] The formula for calculating the mean absolute error of the departure delay time prediction results is as follows: The formula for calculating the root mean square error of the departure delay time prediction results is as follows: The accuracy of the departure delay prediction results for delay classification is calculated using a 15-minute delay threshold (delays greater than 15 minutes are considered delays). The formula is as follows: In the formula, The mean absolute error of the predicted departure delay time is... The root mean square error of the predicted departure delay time is given. To ensure the accuracy of departure delay classification, The number of samples in the test set. For the predicted first Departure delay time for each sample This corresponds to the actual departure delay time. To predict departure delay status, This indicates a genuine departure delay. This is an indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise.

[0039] The formula for calculating the mean absolute error of the arrival delay time prediction results is as follows: The formula for calculating the root mean square error of the arrival delay time prediction results is as follows: The accuracy of the arrival delay prediction results for delay classification is calculated using a 15-minute delay threshold (delays greater than 15 minutes are considered delays). The formula is as follows: In the formula, The mean absolute error of the predicted arrival delay time is... The root mean square error of the predicted arrival delay time is... To ensure the accuracy of arrival delay classification, The number of samples in the test set. For the predicted first Arrival delay time for each sample This corresponds to the actual arrival delay time. To predict arrival delays, This indicates a genuine arrival delay. This is an indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise.

[0040] To quantify the relationship between input features and expert networks, the expert activation weights output by the gating network can be recorded when each sample is input to the test set. The Pearson correlation coefficient between each input feature (after processing the sequence length) and the scores of different experts can be calculated. A heatmap of the correlation between input features and expert scores is shown below. Figure 3As shown, different expert networks in this invention exhibit varying sensitivities to input features, with some networks focusing on learning specific feature patterns. Intuitively, this means that the scores of different expert networks for the same input feature show different correlations. For example, Expert 1 shows a strong correlation with the mean and variance of departure delays over historical time steps. When the variance of historical flight delays is large, Expert 1 scores higher, indicating that it may focus on capturing and predicting specific patterns of delay propagation or deterioration. Expert 2's score shows a strong negative correlation with the number of days (day of the month), indicating that it may be more inclined to capture short-term cyclical changes within a month. Expert 3's score shows a significant positive correlation with the year and month, indicating that it may pay more attention to flight patterns in different years or months, tending to capture long-term trends or seasonal changes in delays.

[0041] In this embodiment, the model's hyperparameters are set as follows: batch_size is set to 256, learning rate is set to 0.00001, and cosine annealing is used to dynamically adjust the learning rate; the hidden layer dimension is 256, the number of heads in the multi-head attention module is 8, the number of shared experts in the dynamically gated MoE is 1, the number of routing experts is 4, the number of multi-head attention heads in the TransformerEncoder structure of the routing experts is 8, and the number of layers is 4; the loss function is weighted mean squared error, calculated as follows: In the formula, For loss function, The weight for departure delay losses is used to adjust the optimal balance between the two objectives.

[0042] The specific evaluation metrics for flight delay prediction results are as follows: MAE for departure delay prediction is 15.32, RMSE is 32.00, and Accuracy is 67.45%; MAE for arrival delay prediction is 17.71, RMSE is 34.47, and Accuracy is 90.25%. This indicates that there are some extreme error points in departure delay prediction, and these outliers may originate from unforeseen factors during the departure phase (such as temporary scheduling changes, sudden weather changes, etc.). The delay classification accuracy of arrival delay prediction is significantly improved compared to departure delay prediction, indicating that the error distribution of arrival delays is more concentrated.

[0043] Example 2 This embodiment provides a flight delay time prediction system based on dynamic gating (MoE), including: The prediction dataset building module is used to collect historical flight operation data, perform data preprocessing, and build a prediction dataset. The temporal feature extraction module is used to encode the prediction dataset as a time series matrix using feature vectorization, and to extract the temporal features of the prediction dataset using a multi-head attention mechanism. The fusion feature acquisition module is used to input the temporal features of the prediction dataset into the dynamic gating MoE. Through the dynamic gating mechanism, the outputs of the routing expert sub-model and the shared expert sub-model are adaptively selected and combined to obtain the fusion features. The dynamic gating MoE consists of at least one shared expert sub-model and multiple routing expert sub-models. The flight delay prediction module is used to generate flight delay prediction results based on fusion features. The flight delay prediction results include departure delay prediction results and arrival delay prediction results.

[0044] The rest is the same as in Example 1.

[0045] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for predicting flight delay time based on dynamic gating (MoE), characterized in that, Includes the following steps: Collect historical flight operation data and perform data preprocessing to construct a prediction dataset; The prediction dataset is encoded as a time series matrix using feature vectorization, and the temporal features of the prediction dataset are extracted using a multi-head attention mechanism. The temporal features of the prediction dataset are input into the dynamic gating MoE. The outputs of the routing expert sub-model and the shared expert sub-model are adaptively selected and combined through the dynamic gating mechanism to obtain the fused features. The dynamic gating MoE consists of at least one shared expert sub-model and multiple routing expert sub-models. Flight delay time prediction results are generated based on the fusion features. The flight delay time prediction results include departure delay time prediction results and arrival delay time prediction results.

2. The flight delay time prediction method based on dynamic gating MoE according to claim 1, characterized in that, The historical flight operation data includes flight number, departure point, destination, planned departure time, planned arrival time, actual departure time, actual arrival time, preceding flight delay information, and other factors related to flight operation.

3. The flight delay time prediction method based on dynamic gating MoE according to claim 1, characterized in that, The specific steps for collecting historical flight operation data, performing data preprocessing, and constructing a prediction dataset include: Data cleaning was performed on historical flight operation data to obtain a valid dataset; The effective dataset is encoded by using label encoding to convert discrete data (excluding time) in the effective dataset into numerical representation, using cyclic encoding to spatially map the time in the effective dataset, and normalizing the continuous data in the effective dataset. Delay tags are obtained based on historical flight operation data; The prediction dataset is obtained by combining the delay labels and the valid dataset after data encoding.

4. The flight delay time prediction method based on dynamic gating MoE according to claim 3, characterized in that, The delay labels include a category label for earlier arrivals on the same day and a numerical label for the delay of preceding flights. The category label for earlier arrivals on the same day indicates whether there were any delays of flights on the same route before the current flight departed on the same day. The numerical label for the delay of preceding flights indicates the delay time of the previous flight segment performed by the aircraft operating the current flight before arriving at the airport.

5. The flight delay time prediction method based on dynamic gating MoE according to claim 1, characterized in that, The specific steps for encoding the prediction dataset as a time series matrix using feature vectorization and extracting the temporal features of the prediction dataset using a multi-head attention mechanism include: The prediction dataset is grouped by flight number and sorted by time to obtain time series data; The time series data is mapped to a hidden dimension and then standardized using RMS Norm. The standardized time-series data is encoded with positional codes, and time-series features are extracted using a multi-head attention module. This module generates query vectors, key vectors, and value vectors to calculate attention weights between different time steps within the time-series data, thereby capturing key temporal dependencies. The time-series features are as follows: In the formula, As a time series feature, It is time series data. For standardized time-series data, For multi-head attention modules, For the output of the multi-head attention module, For splicing operations, In the multi-head attention module Output of size To output the projection matrix, For attention mechanisms, For input multi-head attention module number The query vector of the head. For input multi-head attention module number The key vector of the head, For input multi-head attention module number The value vector of the head, For the multi-head attention module The dimensions of the head.

6. The flight delay time prediction method based on dynamic gating MoE according to claim 1, characterized in that, The specific steps for inputting the temporal features of the predicted dataset into the dynamically gated MoE, and adaptively selecting and combining the outputs of the routing expert sub-model and the shared expert sub-model through the dynamic gating mechanism to obtain the fused features include: The temporal features of the prediction dataset are standardized using RMS Norm and then input into the dynamic gated MoE. Based on the standardized temporal characteristics, the score of each routing expert in the routing expert sub-model is obtained through a gating network, and the K routing experts with the highest scores are selected for activation using a Top-K coefficient gating strategy. The scores of the K highest-scoring routing experts are normalized using the Softmax activation function to obtain the activation weight of each routing expert. The temporal features of the predicted dataset are input into the K routing experts with the highest scores to obtain the expression of the temporal features by each routing expert; The output of the routing expert sub-model is obtained based on the expression of temporal features and the corresponding activation weights of each routing expert; The temporal features of the prediction dataset are input into each shared expert in the shared expert sub-model to obtain the output of the shared expert sub-model; The output of the routing expert sub-model and the output of the shared expert sub-model are combined to obtain the output of the dynamic gating MoE; The output of the dynamically gated MoE is residually connected and added to the input to obtain the fused feature.

7. The flight delay time prediction method based on dynamic gating MoE according to claim 6, characterized in that, The fusion feature is: In the formula, As a feature of fusion, As a time series feature, For the time series characteristics after standardization, For dynamic gating MoE, For the output of the dynamically gated MoE, This is the output of the routing expert sub-model. To share the output of the expert sub-model, The number of routing expert sub-models, Activate the weight for experts, For the first The representation of temporal features by each routing expert sub-model To share the number of expert sub-models, For the first The representation of temporal features by a shared expert sub-model.

8. The flight delay time prediction method based on dynamic gating MoE according to claim 6, characterized in that, The routing experts employ a Transformer Encoder structure. The specific steps for obtaining each routing expert's representation of the temporal features by inputting the temporal features of the prediction dataset into the K highest-scoring routing experts include: The temporal features are position-encoded to obtain a position-encoded vector. The position-encoded vector is then added element-wise to the temporal features to obtain a sequence representation containing positional information. The multi-head attention layer captures the dependencies between different time steps within the sequence, and obtains the output of the multi-head attention layer. The output of the multi-head attention layer is coupled with the sequence representation containing positional information through residual connection and normalization to obtain the normalized result; The normalization result is input into a fully connected feedforward neural network to enhance the features at each time step, and the output of the fully connected feedforward neural network is obtained. The output and normalized result of the fully connected feedforward neural network are subjected to residual connection and normalization processing to complete the processing of a single Transformer Encoder structure; After processing by an N-layer stacked Transformer Encoder structure, each routing expert's expression of the timing characteristics is obtained.

9. The flight delay time prediction method based on dynamic gating MoE according to claim 1, characterized in that, The predictive capability of the Dynamic Gated Entity (MoE) is evaluated based on the prediction results of departure and arrival delays, respectively, from three aspects: the magnitude of the prediction error, the error distribution, and the accuracy of the delay classification.

10. A flight delay time prediction system based on dynamic gating (MoE), characterized in that, include: The prediction dataset building module is used to collect historical flight operation data, perform data preprocessing, and build a prediction dataset. The temporal feature extraction module is used to encode the prediction dataset in the form of a time series matrix using feature vectorization, and to extract the temporal features of the prediction dataset using a multi-head attention mechanism. The fusion feature acquisition module is used to input the temporal features of the prediction dataset into the dynamic gating MoE, and adaptively select and combine the outputs of the routing expert sub-model and the shared expert sub-model through the dynamic gating mechanism to obtain the fusion features. The dynamic gating MoE consists of at least one shared expert sub-model and multiple routing expert sub-models. The flight delay time prediction module is used to generate flight delay time prediction results based on the fusion features. The flight delay time prediction results include departure delay time prediction results and arrival delay time prediction results.

Images