Data processing method, device and medium based on hybrid bidirectional time-domain convolutional network

Abstract

The present invention relates to the field of computer technology applications, and in particular to a data processing method, device and medium based on a hybrid bidirectional time-domain convolutional network. The method comprises: receiving a training request containing an original model and an original data set, wherein the original model is composed of a cascade of a bidirectional time-domain convolutional network module, a bidirectional gated recurrent unit module, a feature weight distribution module and an output prediction module, and the training data is the historical flight data of a specified airport. Secondly, the data set is divided into multiple scales by designing three sliding window mechanisms. Finally, the original model is independently trained based on the sequence data sets of each scale to generate a corresponding target model to capture the flight operation rules at different time granularities. This method effectively enhances the model's ability to model complex time series dependencies in flight data through the combination of multi-scale window division and hybrid network architecture, and is suitable for time series analysis tasks such as flight delay prediction.

Description

Technical Field

[0001] The present invention relates to the field of computer technology applications, and in particular to a data processing method, device and medium based on a hybrid bidirectional time-domain convolutional network. Background Art

[0002] In some application scenarios, such as flight delay prediction, traditional machine learning methods (such as LightGBM and XGBoost) are widely used due to their computational efficiency. However, single models struggle to capture nonlinear correlations between multiple source features. In recent years, deep learning models have become a research hotspot due to their ability to model time series features. Models such as LSTM-RNN and CNN-LSTM have improved prediction accuracy by integrating time series features, while DCNN and CliqueNet have enhanced local pattern recognition. However, these models generally suffer from insufficient spatial feature mining, making it difficult to characterize the impact of airport network topology on delay propagation. The introduction of graph neural networks (GNNs) has promoted progress in spatial feature modeling. Models such as Graph WaveNet and STGCN construct topological graphs using airports as nodes and extract spatial dependencies between airports using graph convolutional networks (GCNs), significantly improving model generalization. However, existing GNN models still have two major limitations: first, they lack in-depth exploration of multi-scale patterns in time series (such as diurnal and weekly seasonality); second, they lack the ability to adaptively model dynamic features across time periods, making it difficult to capture the temporal dynamics of delay evolution in complex scenarios. The inherent flaws of traditional time series models further exacerbate these issues: RNN-like models are prone to vanishing gradients in long sequences and struggle to preserve long-term dependencies. While CNNs can extract local features, their fixed receptive field limits global pattern recognition. This "local-global" feature processing split fundamentally stems from the single time series modeling paradigm's lack of bidirectional dynamic correlation, making it incapable of simultaneously characterizing both short-term fluctuations and long-term trends. Therefore, building predictive models that integrate multi-scale spatiotemporal features and possess bidirectional dynamic modeling capabilities has become a key approach to overcoming existing technological bottlenecks. Summary of the Invention

[0003] In view of the above technical problems, the technical solution adopted by the present invention is:

[0004] According to a first aspect of the present invention, a data processing method based on a hybrid bidirectional time-domain convolutional network is provided, the method comprising the following steps:

[0005] S100, receiving the current model training request information, wherein the model training request information includes the original model to be trained and the original data set used to train the model; wherein the original model includes a bidirectional time-domain convolutional network module, a bidirectional gated recurrent unit module, a feature weight allocation module and an output prediction module connected in sequence, and the original data set includes multiple original data, each original data contains input information and output information, the input information includes basic attributes of the flight, time-related features, space-related features, operating status and environmental information; the output information is the actual arrival time of the flight.

[0006] S200, perform a first preprocessing operation on the input information of the original data to obtain a first preprocessing result and use it as an input feature vector, perform a second preprocessing operation on the output information to obtain a second preprocessing result and use it as a target value; combine the input feature vector and the target value corresponding to each piece of original data into a piece of training data, thereby obtaining a training data set for training the original model.

[0007] S300, based on the first sliding window mechanism, the second sliding window mechanism and the third sliding window mechanism, the training data set is divided into a first training data set, a second training data set and a third training data set; wherein, the sliding step length of the first sliding window mechanism is 1 first time dimension, and the window length is L1 first time dimensions, L1≥1; the sliding step length of the second sliding window mechanism is 1 second time dimension, and the window length is L2 second time dimensions, L2≥1; the sliding step length of the third sliding window mechanism is 1 third time dimension, and the window length is L3 third time dimensions, L3≥1, and the duration of the first time dimension is less than the duration of the second time dimension, and the duration of the second time dimension is less than the duration of the third time dimension.

[0008] S400: Based on each training data set, the original model is trained to obtain a trained original model as a target model corresponding to the training data set.

[0009] According to a second aspect of the present invention, an electronic device is provided, comprising a processor and a memory; the processor is configured to execute the steps of the method according to the first aspect of the present invention by calling a program or instruction stored in the memory.

[0010] According to a third aspect of the present invention, there is provided a computer-readable storage medium storing a program or instructions, wherein the program or instructions enable a computer to execute the steps of the method according to the first aspect of the present invention.

[0011] The present invention has at least the following beneficial effects:

[0012] The data processing method based on a hybrid bidirectional time-domain convolutional network (BTCN) provided by the present invention uses a dynamic sliding window to segment airport delay time series into time windows of varying sizes, such as daily, weekly, and monthly. This method effectively captures delay patterns at different time scales. The processed time series is then fed into a bidirectional time-domain convolutional network and a bidirectional gated recurrent unit network. This bidirectional structure enables the model to simultaneously capture both forward and backward dynamic characteristics and dependencies in the time series, enhancing its ability to capture periodic features and sensitivity to temporal dynamics. Finally, a multi-head attention mechanism module is used to weight hidden states across different time dimensions, improving the accuracy of flight delay predictions.

[0013] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present invention, nor is it intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0015] Figure 1 A flowchart of a data processing method based on a hybrid bidirectional time-domain convolutional network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0018] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of the steps can be performed in parallel, concurrently, or simultaneously. In addition, the order of the steps can be rearranged. A process can be terminated when its operation is completed, but can also have additional steps not included in the accompanying drawings. A process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0019] An embodiment of the present invention provides a data processing method based on a hybrid bidirectional time-domain convolutional network. The method aims to capture long-term and short-term trends through different time windows by combining a hybrid bidirectional time-domain convolutional network model with a dynamic sliding window mechanism, so that the model can more comprehensively learn the dynamic characteristics of the past and future in the time series, thereby improving the time series modeling capabilities for flight delay prediction.

[0020] The embodiment of the present invention provides a data processing method based on a hybrid bidirectional time-domain convolutional network, such as Figure 1 As shown, the method includes the following steps:

[0021] S100, receiving current model training request information, where the model training request information includes an original model to be trained and an original data set for training the model.

[0022] The original model includes a bidirectional time-domain convolutional network module, a bidirectional gated recurrent unit module, a feature weight allocation module and an output prediction module, which are connected in sequence.

[0023] In an embodiment of the present invention, a bidirectional time-domain convolutional network module is composed of a forward residual module, a data flipping layer, a backward residual module and a feature splicing module. The forward residual module and the backward residual module have the same structure, and each residual module includes a one-dimensional convolution layer, a first convolution block, a second convolution block and a residual splicing unit. The first convolution block and the second convolution block have the same structure, and each convolution block includes a dilated causal convolution layer, a weight normalization layer and a batch normalization layer, an activation function layer and a dropout layer connected in sequence, wherein the data flipping layer is connected to the dilated convolution layer of the first convolution block of the backward residual module, and the dropout layer of the first convolution block is connected to the dilated causal convolution layer of the second convolution block, that is, the output of the dropout layer of the first convolution block serves as the input of the dilated causal convolution layer of the second convolution block, the one-dimensional convolution layer of each residual module and the dropout layer of the second convolution block are connected to the corresponding residual splicing unit, and the residual splicing unit of each residual module is connected to the feature splicing module.

[0024] In this embodiment of the present invention, the input data is the one-dimensional convolution layer, the extended causal convolution layer of the forward residual module, and the data flipping layer. The one-dimensional convolution layer is used to perform a one-dimensional convolution operation on the input data, and the data flipping layer is used to flip the input data. The activation function layer uses the ReLU activation function for processing. The residual splicing unit is used to splice the outputs of the one-dimensional convolution layer and the dropout layer of the second convolution block, and the feature splicing module is used to splice the outputs of the forward residual block and the backward residual block.

[0025] The forward residual module processes the input data in the forward direction, while the subsequent residual module processes the input data in the reverse direction. Data is fed into the residual module for data diversion in two main pathways. One pathway first passes through a dilated causal convolutional layer, followed by a weight normalization layer and a batch normalization layer. Weight normalization scales and rotates the weights to separate the norm and direction of the weight vector, making it easier to control model complexity. Batch normalization normalizes the input of each mini-batch of data, ensuring that the input of each neuron has the same mean and variance, accelerating model convergence. Finally, an activation function layer and a dropout layer randomly drop features or neurons to reduce the risk of overfitting the model to the training data, preventing the network from becoming overly dependent on specific neurons during training and improving model generalization. The other pathway performs a 1×1 convolution operation. The outputs of the two pathways are summed to produce the output of a single residual module. The outputs of the forward and backward residual modules are combined to form the final model result. The depth of the bidirectional temporal convolutional network module can be flexibly controlled by adjusting the number of stacked residual blocks.

[0026] In an embodiment of the present invention, the residual connection of the bidirectional time-domain convolutional network module provides a direct path from the input to the deep layer, allowing the gradient to bypass some intermediate layers and be directly transmitted back to the shallow layer, thereby reducing the attenuation of the gradient during the backpropagation process. Even if the network layer is very deep, the input information can be effectively transmitted to the output layer, thereby avoiding the gradient vanishing problem.

[0027] In this embodiment of the present invention, the bidirectional GRU module combines the advantages of both bidirectional models and gating mechanisms. It consists of two directional GRU networks: a forward GRU network and a backward GRU network. The forward GRU network processes time series data from front to back, while the backward GRU network processes time series data from back to front. This bidirectional structure can simultaneously capture information from both the past and the future, thereby more comprehensively modeling the temporal relationships in time series data.

[0028] Those skilled in the art should understand that the workflow of the bidirectional gated recurrent unit module belongs to the scope of the prior art. To avoid redundancy, the present invention omits its detailed description.

[0029] In an embodiment of the present invention, the feature weight assignment module is a multi-head attention mechanism module. The feature weight assignment module is used to enable the model to selectively focus on different parts of the input sequence, assigning different weights to each part of the input sequence, thereby highlighting information that is more critical to the task. The multi-head attention mechanism is an extension of self-attention. It allows the model to "view" input data from multiple perspectives. By introducing multiple parallel attention heads, each head independently calculates attention, and the results are spliced ​​and linearly transformed, thereby capturing more fine-grained information and complex semantic relationships.

[0030] In an embodiment of the present invention, the number of attention heads can be set based on actual needs. In an exemplary embodiment, the number of attention heads can be 4. The key value of the attention mechanism is a number that can be divided by the number of attention heads.

[0031] The workflow of the feature weight assignment module is:

[0032] (1) Apply three different linear transformations (fully connected layers) to the input vector to generate query, key, and value vectors respectively. Split the linearly transformed query, key, and value vectors into multiple parts, each of which represents an attention head.

[0033] Those skilled in the art will appreciate that existing methods may be used to generate query, key, and value vectors.

[0034] (2) For each attention head, the scaled dot product attention mechanism is used to calculate the attention output and obtain the scaled similarity matrix.

[0035] Those skilled in the art will appreciate that any method that uses a scaled dot product attention mechanism to calculate attention output and obtain a scaled similarity matrix falls within the scope of protection of the present invention.

[0036] (3) Perform Softmax normalization on the scaled similarity matrix to obtain the attention weight of each element.

[0037] (4) Use the attention weight corresponding to each attention head to perform weighted summation on the corresponding value vector to obtain the final output vector of the attention head.

[0038] (5) The final output vectors of all attention heads are concatenated to obtain the output result of the feature weight allocation module.

[0039] In an embodiment of the present invention, the fully connected layer is used to map multiple important features obtained by the feature weight distribution module into a one-dimensional space to obtain the final output result of the network, that is, the arrival delay time prediction value corresponding to each flight data.

[0040] In this embodiment of the present invention, the raw data set includes multiple pieces of raw data, each of which contains input information and output information. The input information includes basic flight attributes, time-related characteristics, space-related characteristics, operating status, and environmental information; the output information is the actual arrival time of the flight. The raw data set can specifically be flight data for a specified airport within a set time period, for example, flight data for a specified airport within a year.

[0041] In this embodiment of the present invention, the basic attributes include flight number, aircraft tail number, origin airport, destination airport, etc. Time-related features include estimated takeoff time, takeoff delay time, aircraft taxi-out time, wheels-off time, wheels-on time, aircraft taxi-in time, estimated arrival time, previous flight delay time, estimated flight time, actual flight time, and airborne flight time. Spatial-related features include controlled flight distance, air traffic volume, and airport congestion level. Environmental information includes meteorological parameters for the origin and destination airports.

[0042] The air traffic flow in each flight data is equal to the sum of the number of inbound flights and the number of outbound flights at the destination airport corresponding to the flight data within the time period to which the actual arrival time of the flight data belongs.

[0043] In this embodiment of the present invention, flight data is grouped into hourly time intervals, and the number of departing and landing flights within each time window is counted to represent the actual air traffic flow within that time interval. For example, for a flight whose actual arrival time is 10:30, since 10:30 falls within the time period from 10:00 to 11:00, the air traffic flow corresponding to this flight data is equal to the sum of the number of arriving and departing flights at the destination airport corresponding to this flight data during this time period.

[0044] In an embodiment of the present invention, the airport congestion level in each flight data item is equal to the sum of the average taxi-in time and the average taxi-out time of the destination airport corresponding to the flight data item during the time period to which the actual arrival time of the flight data item belongs. In an embodiment of the present invention, the flight data item is grouped according to hourly time intervals, and the average taxi-in time and the average taxi-out time of the airport item within each time window are counted to represent the airport congestion level within the time period. For example, for a certain flight data item, its actual arrival time is 10:30. Since 10:30 belongs to the time period from 10:00 to 11:00, the airport congestion level corresponding to the flight data item is equal to the sum of the average taxi-in time and the average taxi-out time of the destination airport corresponding to the flight data item during the time period from 10:00 to 11:00. The average taxi-in time of the airport item is equal to the average taxi-in time of all flights in the corresponding time period, and the average taxi-in time of the airport item is equal to the average taxi-out time of all flights in the corresponding time period.

[0045] In this embodiment of the present invention, meteorological parameters may include temperature, humidity, visibility, wind speed, precipitation, and weather phenomena. In this embodiment of the present invention, the original dataset can be derived from the 2023 annual flight on-time performance data for Atlanta International Airport (ATL) compiled by the U.S. Bureau of Transportation Statistics (BTS). Data from the top ten airports with the most frequent connections to ATL are selected. Corresponding weather data from the National Oceanic and Atmospheric Administration (NOAA), covering characteristics such as dry-bulb temperature and wind speed, is aligned with flight data by airport number and time to construct a flight delay prediction dataset incorporating meteorological information. Missing values ​​are imputed using the mean of the preceding and following three days, and are deleted when no reference data is available. Outlier data is directly discarded. A total of 139,232 valid data items are obtained after cleaning.

[0046] S200, perform a first preprocessing operation on the input information of the original data to obtain a first preprocessing result and use it as an input feature vector, perform a second preprocessing operation on the output information to obtain a second preprocessing result and use it as a target value; combine the input feature vector and the target value corresponding to each piece of original data into a piece of training data, thereby obtaining a training data set for training the original model.

[0047] In the embodiment of the present invention, the first preprocessing operation includes operations such as data cleaning, code conversion, and normalization. The second preprocessing operation includes operations such as feature coding and numerical conversion.

[0048] S300 : Divide the original data set into a first training data set, a second training data set, and a third training data set based on a first sliding window mechanism, a second sliding window mechanism, and a third sliding window mechanism, respectively.

[0049] In this embodiment of the present invention, the sliding step size of the first sliding window mechanism is one first time dimension, and the window length is L1 first time dimensions, where L1 ≥ 1. The sliding step size of the second sliding window mechanism is one second time dimension, and the window length is L2 second time dimensions, where L2 ≥ 1. The sliding step size of the third sliding window mechanism is one third time dimension, and the window length is L3 third time dimensions, where L3 ≥ 1. In addition, the duration of the first time dimension is less than the duration of the second time dimension, and the duration of the second time dimension is less than the duration of the third time dimension.

[0050] In an embodiment of the present invention, the first time dimension is day, the second time dimension is week, and the third time dimension is month, that is, the duration of the first time dimension is one day, the duration of the second time dimension is one week, and the duration of the third time dimension is one month.

[0051] In the embodiment of the present invention, different time window mechanisms are used to capture the long-term and short-term trends of sequence data, so that the model can more comprehensively learn the dynamic characteristics of the past and future in the time series, thereby improving the time series modeling capabilities of flight delay prediction.

[0052] S400: Based on each training data set, the original model is trained to obtain a trained original model as the target model corresponding to the training data set, that is, three target models are obtained. In this embodiment of the present invention, during the training of the original model, the hyperparameters that need to be updated include the learning rate, the number of neurons in the bidirectional gated recurrent unit, the number of attention mechanism keys in the feature weight allocation module, and the L2 regularization coefficient.

[0053] Furthermore, the hyperparameters are updated using a crested porcupine optimization algorithm, which is a novel meta-heuristic algorithm that uses a cyclic population reduction technique to accelerate convergence and maintain population diversity.

[0054] The crested porcupine optimization algorithm first specifies the population size and number of evolutions, defining the adjustment range of the original model's hyperparameters, including the model learning rate, the number of neurons in the bidirectionally gated recurrent unit, the number of attention mechanism keys, and the L2 regularization coefficient. Among these are individual hyperparameters of the population. Next, by simulating the behavior of crested porcupines, the algorithm continuously evolves using different defense strategies. Each time, the fitness of the new population individuals is calculated to determine whether to update the current optimal solution and the global optimal solution. When the number of population evolutions reaches the maximum, the algorithm returns the global optimal solution and the algorithm convergence curve. The global optimal solution contains the optimal values ​​of each parameter defined in the model.

[0055] In this embodiment of the present invention, the hyperparameter adjustment range can be set based on actual needs. In one exemplary embodiment, the model learning rate can be adjusted from 0.0001 to 0.01, the number of neurons in the bidirectional gated recurrent unit can be adjusted from 50 to 100, the number of attention mechanism keys can be adjusted from 30 to 50, and the L2 regularization coefficient can be adjusted from 0.0001 to 0.001. The population size and number of evolutions can be experimental values. In one adaptive embodiment, the population size is 5 and the number of evolutions is 10.

[0056] In an exemplary embodiment of the present invention, the fitness function used by the crested porcupine optimization algorithm is a mean square error loss function.

[0057] In another exemplary embodiment of the present invention, the fitness function Fitness used in the crested porcupine optimization algorithm satisfies the following conditions:

[0058] Fitness=w1×(-F1)+w2×F2+w3(-F3).

[0059] F1 is the model prediction accuracy indicator. F2 is the ratio of actual positive examples to all true positive examples among the top K predicted positive examples in the model prediction results. F3 is the model complexity indicator, with w1 being the first coefficient, w2 being the second coefficient, and w3 being the third coefficient.

[0060] In the embodiment of the present invention, F1 is the squared absolute error. F2 is used to characterize the recall ability of the model for high-priority samples and is the K recall rate. For example, if there are 15 seriously delayed flights, and the model predicts that 10 of the top 20 most likely delayed flights actually have serious delays, then F2 = 10 / 15 ≈ 0.667. F3 can be equal to the sum of the model parameters and 10 6 ratio.

[0061] In an embodiment of the present invention, w1+w2+w3=1, w1, w2 and w3 can be set based on actual needs. For example, the optimal weight combination can be found through grid search or Bayesian optimization using a validation set. In an illustrative embodiment, w1=w2=0.4, w3=0.2.

[0062] Experiments show that the crested porcupine optimization algorithm using this fitness function can improve the flight delay prediction model: the K-recall rate for samples with delays greater than 2 hours is increased to 75% (compared to 60% for traditional methods); the number of hyperparameters is reduced by 30% (from 5 million to 3.5 million), and the inference speed is increased by 25%.

[0063] Furthermore, S400 may specifically include:

[0064] S401, train the current model based on the sample data of the current batch of each training data set to obtain the trained current model; wherein the initial state of the current model is the original model, and the sample data of the current batch is the sequence data within a time window in the corresponding sequence data set.

[0065] In the embodiment of the present invention, the original model includes m initialized sub-models, each of which has a randomized initialized hyperparameter combination. m is the population size.

[0066] Among them, S401 specifically includes:

[0067] S4011, set the evolution counter c=1;

[0068] S4012, if c≤c0, execute S4013, if c>c0, execute S4017; c0 is the evolution number threshold.

[0069] S4013, based on the fitness of each individual in the current population, the current population is updated using the evolutionary strategy of the crested porcupine optimization algorithm. The updated population constitutes a new generation population, and based on the new generation population, new m sub-models are obtained as the current m sub-models; the training times counter d=1 is initialized, and S4014 is executed.

[0070] In the embodiment of the present invention, the current population is composed of the hyperparameter combinations corresponding to the current m sub-models, and each sub-model corresponds to a set of hyperparameter combinations. The initial state of the current m sub-models is m initialized sub-models.

[0071] In the embodiment of the present invention, based on the fitness of each individual in the current population, the evolutionary strategy of the crested porcupine optimization algorithm is used to update the current population, which may specifically include:

[0072] (1) Calculate the fitness of each individual in the current population based on the aforementioned fitness function.

[0073] (2) Individuals are sorted in descending order according to their fitness and divided into an elite layer, an exploration layer, and an exploitation layer. The elite layer is the individuals corresponding to the top 20% of the fitness after sorting, the exploration layer is the individuals corresponding to the middle 50% of the fitness after sorting, and the exploitation layer is the individuals corresponding to the bottom 30% of the fitness after sorting.

[0074] (3) The improved Lévy flight model is used to perform global exploration on the individuals in the exploration layer, where the improved Lévy flight model satisfies the following conditions:

[0075] Xnew=X best -α×|X best -β×X_i|+γ×Lévy(λ)×(X rand -X i ).

[0076] Among them, X new is the updated individual, X best is the individual position with the best fitness in the current population, X i is the individual position currently being updated, that is, the current hyperparameter combination of the i-th sub-model, where i ranges from 1 to m, and X rand is a reference individual position randomly selected from the current population. α is the scaling factor used to control X best The degree of influence on the update direction, a larger α makes the search more biased towards the global optimal solution, a smaller α retains more individual characteristics, and the value range is [0, 2]. β is the weight coefficient used to adjust X i and X bestThe distance contribution between the two is in the range [0, 1]. γ is the adaptive step size factor, which is used to dynamically adjust the search step size. It gradually decreases with each iteration, allowing the algorithm to transition from global exploration to local exploitation. Its value range is [0, 1]. λ is the exponential parameter of the Lévy distribution, which determines the heavy-tailed nature of the step size distribution. A larger λ produces more small steps, while a smaller λ produces more large steps. Its value range is [1.0, 2.0].

[0077] (4) A bidirectional search model is used to perform local fine-tuning on individuals in the development layer.

[0078] Among them, the bidirectional search model meets the following conditions:

[0079] If rand < 0.5, X new ={X prey -δ1×|ε1×X prey -X i |, if rand ≥ 0.5, X new ={X prey +δ2×|ε2×X prey -X i |.

[0080] Among them, rand is a random number between 0 and 1, which is used to determine the search direction. prey is the currently selected prey position, X best or X rand . δ1 is the forward search step size scaling factor, which controls the speed of approaching the prey. The value range is [0, 1]. A larger δ1 leads to a large step size search, and a smaller δ1 leads to a fine search. δ2 is the reverse search step size scaling factor, which controls the exploration amplitude away from the prey. The value range is [0, 1]. Usually, δ2 is set to < δ1 to ensure that the main search direction is close to the optimal solution. ε1 is the forward search direction weight, which is used to adjust the relative importance of the prey position and the current position. The value range is [0, 1]. A larger ε1 makes the search more biased towards the prey position, and a smaller ε1 retains more individual characteristics. ε2 is the reverse search direction weight, which is similar to ε1 but used for reverse search. The value range is [0, 1]. Usually, ε2> ε1 is set to enhance the exploratory nature of the reverse search.

[0081] (5) Apply Gaussian perturbation to elite individuals: X new =X elite +σ×N(0,1)×(X best -X elite), and establish an information sharing mechanism within the population, where the best individual transmits 10% of the feature information to the adjacent individuals. N (0, 1) is a standardized random perturbation source, whose core function is to ensure that the update of elite individuals maintains a trend towards the optimal solution while maintaining a certain degree of flexibility through randomness that conforms to the normal distribution. elite is the position of the elite individuals in the current population (usually the individuals in the top k% of fitness ranking). σ is the perturbation intensity coefficient, which is used to control the amplitude of the Gaussian perturbation. Its value range is [0, 1]. A larger σ produces more severe perturbations, while a smaller σ produces more refined local search.

[0082] (6) Adaptive parameter control: dynamically adjust the Lévy index λ = 1.5 + 0.5 × sin (π × p / 2) and introduce an inertia weight ω = 0.9 - 0.5 × p to balance the influence of historical search directions. p is the evolution node indicator.

[0083] In the embodiment of the present invention, individuals are updated through a multi-stage search strategy, adaptive parameter control, and an elite protection mechanism, thereby effectively improving the global search capability and convergence accuracy of the algorithm.

[0084] S4014: If d≤d0, execute S4015; otherwise, execute S4016, where d0 is the threshold for the number of training times.

[0085] In step S4015, the current batch of sample data is fed into each of the m sub-models to obtain m corresponding prediction results. Based on the m prediction results, the loss value for each sub-model is obtained, and the model parameters of each sub-model are updated based on the loss value. Set d = d + 1, and execute step S4014.

[0086] In the embodiment of the present invention, each prediction result includes the arrival delay time prediction value corresponding to each sample data. The loss value of each sub-model can be obtained based on an existing loss function, such as a mean square error loss function.

[0087] S4016, take the hyperparameter combination corresponding to the optimal fitness corresponding to the current m sub-models as the current optimal solution; if the optimal fitness corresponding to the current m sub-models is better than the fitness corresponding to the current global optimal solution, update the current global optimal solution to the current optimal solution, otherwise, do not update the current global optimal solution, set c=c+1, and execute S4012.

[0088] The initial value of the current global optimal solution is the hyperparameter combination corresponding to the optimal fitness of the m initialized sub-models. The optimal fitness corresponding to the current m sub-models is the minimum of the m fitness values.

[0089] S4017: If the current batch is not the last batch, use the sample data of the next batch as the sample data of the current batch and execute S4014; otherwise, that is, the current batch is the last batch, execute S4018.

[0090] S4018: The hyperparameter combination corresponding to the current global optimal solution is used as the target model corresponding to the training data set.

[0091] When flight delay prediction is performed using the target model obtained by the data processing method based on the hybrid bidirectional time-domain convolutional network provided by the embodiment of the present invention, the MAE and RMSE are improved by approximately 15% and 3% respectively compared with the existing prediction scheme.

[0092] The target model obtained in the embodiment of the present invention can be used as a flight delay prediction model in practical applications, specifically a model for predicting flight arrival time delays.

[0093] Furthermore, the method provided in the embodiment of the present invention further includes the following steps:

[0094] S500 , when a data set to be processed is received, a corresponding target model is acquired based on a time range corresponding to the received data set to be processed.

[0095] In this embodiment of the present invention, the dataset to be processed includes basic flight attributes, time-related features, spatial-related features, operational status, and environmental information. It can be scheduled flight data and predicted environmental information for a specified airport within a preset future time period. The length of the preset future time period can be determined based on actual circumstances and can be the current day, the next week, the next month, etc.

[0096] S600: Obtain an input feature vector corresponding to the data set to be processed, input the obtained input feature vector into a corresponding target model, obtain a corresponding prediction result, and output the prediction result.

[0097] Furthermore, the method provided in the embodiment of the present invention further includes the following steps:

[0098] S700, taking the target model with the greatest prediction accuracy among the three target models as the final target model;

[0099] S800, when receiving a data set to be processed, obtaining an input feature vector corresponding to the data set to be processed;

[0100] S900: Input the acquired input feature vector into the final target model to obtain a corresponding prediction result, and output the prediction result.

[0101] Since this embodiment uses the model with the best prediction effect as the target model actually used, it can further improve the prediction accuracy compared with the previous embodiment.

[0102] In summary, the data processing method based on a hybrid bidirectional time-domain convolutional network provided by the present invention can simultaneously capture the forward and reverse dynamic characteristics and dependencies in time series. It also employs a sliding window mechanism to partition flight delay time series by sliding windows of varying sizes, such as daily, weekly, and monthly. By using overlapping time window data, it can capture short-term fluctuations, cyclical changes, and long-term trends. Secondly, the innovative use of the crested porcupine optimization algorithm to optimize the parameters of the hybrid prediction model can further enhance the predictive performance of the combined model.

[0103] An embodiment of the present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the method described in the embodiment of the present invention.

[0104] An embodiment of the present invention further provides a computer-readable storage medium storing computer-executable instructions, wherein the computer instructions are used to execute the method described in the embodiment of the present invention.

[0105] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present invention can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present invention can be achieved. This is not limited herein.

[0106] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A data processing method based on a hybrid bidirectional time-domain convolutional network, characterized in that: The method comprises the following steps: S100, receiving current model training request information, the model training request information including an original model to be trained and an original data set for training the model; wherein the original model includes a bidirectional time-domain convolutional network module, a bidirectional gated recurrent unit module, a feature weight allocation module, and an output prediction module connected in sequence; the original data set includes multiple pieces of original data, each piece of original data includes input information and output information, the input information including basic attributes of a flight, time-related features, space-related features, operating status, and environmental information; the output information is the actual arrival time of the flight; S200, performing a first preprocessing operation on the input information of the original data to obtain a first preprocessing result as an input feature vector, performing a second preprocessing operation on the output information to obtain a second preprocessing result as a target value; combining the input feature vector and the target value corresponding to each piece of original data into a piece of training data, thereby obtaining a training data set for training the original model; S300, dividing the training data set into a first training data set, a second training data set, and a third training data set based on a first sliding window mechanism, a second sliding window mechanism, and a third sliding window mechanism, respectively; wherein the sliding step length of the first sliding window mechanism is 1 first time dimension, and the window length is L1 first time dimensions, L1 ≥ 1; the sliding step length of the second sliding window mechanism is 1 second time dimension, and the window length is L2 second time dimensions, L2 ≥ 1; the sliding step length of the third sliding window mechanism is 1 third time dimension, and the window length is L3 third time dimensions, L3 ≥ 1, and the duration of the first time dimension is less than the duration of the second time dimension, and the duration of the second time dimension is less than the duration of the third time dimension; S400: Based on each training data set, the original model is trained to obtain a trained original model as a target model corresponding to the training data set.

2. The method according to claim 1, characterized in that The feature weight allocation module is a multi-head attention mechanism module.

3. The method according to claim 2, characterized in that During the training of the original model, the hyperparameters that need to be updated include the learning rate, the number of neurons in the bidirectional gated recurrent unit, the number of attention mechanism keys in the feature weight allocation module, and the L2 regularization coefficient.

4. The method according to claim 3, characterized in that The hyperparameters are updated using the Crested Porcupine optimization algorithm.

5. The method according to claim 1, wherein The first time dimension is day, the second time dimension is week, and the third time dimension is month.

6. The method according to claim 1, wherein The bidirectional time-domain convolutional network module consists of a forward residual module, a data flipping layer, a backward residual module and a feature splicing module; wherein, the forward residual module and the backward residual module have the same structure, and each residual module includes a one-dimensional convolution layer, a first convolution block, a second convolution block and a residual splicing unit; the first convolution block and the second convolution block have the same structure, and each convolution block includes an extended causal convolution layer, a weight normalization layer and a batch normalization layer, an activation function layer and a discard layer connected in sequence, wherein the data flipping layer is connected to the extended convolution layer of the first convolution block of the backward residual module, the discard layer of the first convolution block is connected to the extended causal convolution layer of the second convolution block, the one-dimensional convolution layer of each residual module and the discard layer of the second convolution block are connected to the corresponding residual splicing unit, and the residual splicing unit of each residual module is connected to the feature splicing module.

7. The method according to claim 1, characterized in that The following steps are also included: S500, upon receiving a data set to be processed, obtaining a corresponding target model based on a time range corresponding to the received data set to be processed; the data set to be processed includes basic attributes of the flight, time-related features, space-related features, operating status, and environmental information; S600: Obtain an input feature vector corresponding to the data set to be processed, input the obtained input feature vector into a corresponding target model, obtain a corresponding prediction result, and output the prediction result.

8. An electronic device, characterized in that: including processor and memory; The processor is configured to execute the steps of the method according to any one of claims 1 to 7 by calling the program or instructions stored in the memory.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium is used to store a program or instruction, and the program or instruction enables a computer to execute the steps of the method according to any one of claims 1 to 7.

Images