Flight trajectory prediction method, electronic equipment and storage medium
By introducing multi-layer bit context representation and wavelet transformation technology into the flight trajectory prediction model, the problem of low error accumulation and computational efficiency in multi-time flight trajectory prediction is solved, and more accurate and efficient flight trajectory prediction is achieved.
Patent Information
- Application Number
- CN202411512828.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-28
- Publication Date
- 2025-02-18
AI Technical Summary
Existing flight trajectory prediction methods are prone to error accumulation problems in multi-time prediction tasks, and have low computational efficiency, resulting in real-time application delays.
A method based on a trained flight trajectory prediction model is adopted, which includes a trajectory encoder, a horizontally-aware context generator, a frequency learner based on wavelet transform, a differential prompt decoder and a predictor. This model reduces error accumulation and improves computational efficiency through multi-layer bit context representation and wavelet transformation technology.
Effectively reduce error accumulation, improve the calculation efficiency of multi-level flight trajectory prediction tasks, and reduce the delay in real-time applications.
Smart Images
Figure CN119474836A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of data processing, and in particular to a flight trajectory prediction method, electronic equipment and storage medium. Background Art
[0002] Flight trajectory prediction is an important part of air traffic management, which can help air traffic controllers manage airspace more safely and efficiently, such as traffic flow prediction, conflict detection and arrival time estimation. In this context, flight trajectory prediction technology has received increasing attention from researchers in the past decade and has made significant progress. Generally speaking, flight trajectory prediction in the field of air traffic control aims to predict the flight status in future time steps based on the observed historical flight trajectory. At present, the short flight trajectory prediction task within 1 hour is usually formulated as a time series prediction problem, which predicts the future flight trajectory based on real-time observations. According to the modeling method, flight trajectory prediction methods can be divided into three different categories: physical models, filter-based models and data-driven models. Physical models usually establish a set of mathematical equations based on kinematic and aerodynamic assumptions to estimate the future flight status, while filter-based models iteratively estimate the flight trajectory through a predefined system model by considering real-time measurements. However, these methods are limited by fixed inference rules, are prone to error accumulation problems, and are not suitable for multi-time period flight trajectory prediction tasks. In contrast, data-driven models aim to learn flight transition patterns from historical data, which has become the main method in modern air traffic management applications due to its excellent performance.
[0003] Among data-driven models, FlightBERT is a pioneering work that proposes a binary coded representation to encode the attributes of trajectory points as a set of binary vectors, further formulating the flight trajectory prediction task as multiple binary classification problems. Thanks to the innovative idea of binary coded representation, the FlightBERT framework not only enhances the semantic representation of trajectory points, but also avoids the vulnerabilities brought by the normalization algorithm. However, two major limitations should be further addressed to improve the performance of the flight trajectory prediction task. The limitation of binary coded representation is that high-bit prediction errors will lead to outliers in the prediction. For example, given the binary coded representation "0110 0100" (decimal 100), if the misclassification occurs at the 8th bit ("11100100"), the absolute error is 128 (decimal), while if the prediction error occurs at the 1st bit ("0110 0101"), the absolute error is 1 (decimal). FlightBERT recursively performs multi-temporal prediction, that is, predicts the flight state of the next time step based on the observed values, and iteratively applies the predicted values as pseudo-observations to obtain multi-level predictions. Since the prediction errors in pseudo observations accumulate during the recursive inference process, large cumulative errors are prone to occur. In addition, due to the step-by-step prediction paradigm, the computational efficiency is limited by the prediction horizon. As the prediction horizon increases, the inference speed will experience a severe degradation, resulting in unacceptable delays for real-time applications, such as conflict detection. Summary of the invention
[0004] In view of the above technical problems, the technical solution adopted by the present invention is:
[0005] According to a first aspect of the present invention, a flight trajectory prediction method is provided, the method is implemented based on a trained flight trajectory prediction model, the trained flight trajectory prediction model comprises: a trajectory encoder, a horizontal perception context generator, a frequency learning machine based on wavelet transform, a differential prompt decoder and a predictor; the trajectory encoder comprises a first trajectory point embedding module, a time modeling module and a sequence aggregation module connected in sequence, the horizontal perception context generator comprises a one-hot vector encoding module, a horizontal embedding module, a feature aggregation module and a first feature mapping module, the frequency learning machine based on wavelet transform comprises a domain conversion module, a second feature mapping module and an inverse wavelet transform module, the differential prompt decoder comprises a differential sequence encoding module, a second trajectory point embedding module, a feature fusion module and a third feature mapping module; the input of the trajectory encoder and the differential sequence encoding module is model input data, the output of the sequence aggregation module is used as the input of the one-hot vector encoding module, the output of the inverse wavelet transform module is used as the input of the feature fusion module, and the output of the third feature mapping module is used as the input of the predictor; wherein the method comprises the following steps:
[0006] S100, preprocessing the flight trajectory data to be predicted to obtain a time series data set.
[0007] S200, binary encoding is performed on the time series data set to obtain an encoded data set as a prediction data set.
[0008] S300, input the prediction data set into the trained flight trajectory prediction model to obtain the corresponding prediction result as the prediction value of the flight trajectory data to be predicted. According to a second aspect of the present invention, an electronic device is provided, comprising a processor and a memory; the processor is used to execute the steps of the method described in the first aspect of the present invention by calling a program or instruction stored in the memory.
[0009] According to a second 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.
[0010] The present invention has at least the following beneficial effects:
[0011] The flight trajectory prediction method provided by the embodiment of the present invention is implemented based on a flight trajectory prediction model, which includes: a trajectory encoder, a horizontal perception context generator, a frequency learning machine based on wavelet transform, a differential prompt decoder and a predictor connected in sequence; the trajectory encoder includes a first trajectory point embedding module, a time modeling module and a sequence aggregation module connected in sequence, the horizontal perception context generator includes a one-hot vector encoding module, a horizontal embedding module, a feature aggregation module and a first feature mapping module, the frequency learning machine based on wavelet transform includes a domain conversion module, a second feature mapping module and an inverse wavelet transform module, the differential prompt decoder includes a differential sequence encoding module, a feature fusion module and The third feature mapping module; wherein, the trajectory encoder encodes the observation into a probability-level representation, the level-aware context generator can directly generate multi-level predictions (non-autoregressive) instead of performing recursive reasoning, the frequency learning machine based on wavelet transform has a comprehensive understanding of the global flight trends and local motion details in a given trajectory sequence, and achieves the desired performance improvement in predicting future trajectories, the differential prompt decoder is used to perform high-confidence predictions based on the multi-level context representation, and the multi-level context representation uses the differential sequence of observations as prompts. The present invention realizes flight trajectory prediction through this model architecture, which can reduce error accumulation and improve the computational efficiency of the multi-level flight trajectory prediction task.
[0012] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] 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.
[0014] Figure 1 A model framework diagram for implementing a flight trajectory prediction method provided by an embodiment of the present invention;
[0015] Figure 2 A flow chart of a flight trajectory prediction method provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are 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 those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.
[0018] It should be noted that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the steps as sequential processes, many of the steps can be implemented in parallel, concurrently or simultaneously. In addition, the order of the steps can be rearranged. The process can be terminated when its operation is completed, but it can also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.
[0019] The embodiment of the present invention provides a flight trajectory prediction method for predicting the flight trajectory of an aircraft such as a drone. Figure 1 As shown, the method is implemented based on a trained flight trajectory prediction model. Figure 1 As shown, the trained flight trajectory prediction model includes: a trajectory encoder 1, a horizontal perception context generator 2, a frequency learning machine based on wavelet transform 3, a differential prompt decoder 4 and a predictor 5.
[0020] Among them, the trajectory encoder is used to encode the observation into a probability level representation, the horizontal perception context generator is designed to generate a multi-layer context representation of the prediction level, the frequency domain learning machine based on wavelet transform has a comprehensive understanding of the global flight trend and local motion details in a given trajectory sequence, and achieves the desired performance improvement in predicting future trajectories, and the differential prompt decoder is used to perform high-confidence prediction based on a multi-level context representation, which uses the differential sequence of observations as a prompt. Specifically, the trajectory encoder 1 includes a first trajectory point embedding module 101, a time modeling module 102 and a sequence aggregation module 103 connected in sequence, the horizontal perception context generator 2 includes a unique hot vector encoding module 201, a horizontal embedding module 202, a feature aggregation module 203 and a first feature mapping module 204, the frequency learning machine based on wavelet transform 3 includes a domain conversion module 301, a second feature mapping module 302 and an inverse wavelet transform module 303, and the differential prompt decoder 4 includes a differential sequence encoding module 401, a second trajectory point embedding module 402, a feature fusion module 403 and a third feature mapping module 404. The input of the trajectory encoder and the differential sequence encoding module is the model input data, i.e., the data input into the model; the output of the sequence aggregation module is used as the input of the one-hot vector encoding module; the output of the inverse wavelet transform module is used as the input of the feature fusion module; and the output of the third feature mapping module is used as the input of the predictor.
[0021] Furthermore, the flight trajectory prediction method provided by the embodiment of the present invention may include: Figure 2 Steps shown:
[0022] S100, preprocessing the flight trajectory data to be predicted to obtain a time series data set.
[0023] In an embodiment of the present invention, the flight trajectory data can be provided by a third-party provider such as China's actual air traffic control system, and the flight trajectory prediction combines the data of multi-source secondary surveillance radar (SSR) and automatic dependent surveillance broadcast (ADS-B). The proposed method is verified by fusing multi-source flight trajectories, and the data is preprocessed to improve the quality.
[0024] In an embodiment of the present invention, the flight trajectory data to be predicted may be historical data of a set time period, such as data within 30 minutes. Preprocessing may include data inspection, multi-source data analysis, track point and flight information registration, multi-source data alignment, data filtering and fusion, etc. The update interval of the trajectory data is 20 seconds, and the trajectory attributes are standardized by the maximum-minimum normalization method to ensure the consistency of the data scale and range (longitude and latitude are in degrees, altitude is in 10 meters, and speed is in kilometers per hour).
[0025] In the embodiment of the present invention, the flight trajectory data includes longitude, latitude, altitude and speed. The obtained time series data set X can be defined as: X = [x t ] t＝1……m , x t =(Lon t , Lat t , Alt t , Vx t , Vy t , Vz t ) T , x t is the t-th trajectory point in X, t ranges from 1 to m, and m is the number of trajectory points; Lon t , Lat t , Alt t are the longitude, latitude and altitude of the t-th trajectory point, Vx t , Vy t , Vz t The velocity components of the velocity of the t-th trajectory point on the x-axis, y-axis, and z-axis respectively.
[0026] S200, binary encoding is performed on the time series data set to obtain an encoded data set as a prediction data set.
[0027] Those skilled in the art know that any method of binary encoding each data in a time series data set falls within the protection scope of the present invention.
[0028] S300, inputting the prediction data set into a trained flight trajectory prediction model to obtain a corresponding prediction result as a prediction value of the flight trajectory data to be predicted.
[0029] Furthermore, S300 may specifically include:
[0030] S301: Embed the prediction data set into a first high-dimensional embedding space using the first trajectory point embedding module to extract distinguishing spatial features in the prediction data set to obtain a first feature vector.
[0031] In the embodiment of the present invention, the dimension of the first high-dimensional embedding space is greater than the dimension of the prediction data set, and can be set based on actual needs, for example, can be set to 64 dimensions.
[0032] In the embodiment of the present invention, the first trajectory point embedding module 101 may be composed of a one-dimensional convolution layer. The structure of the one-dimensional convolution layer may be an existing structure, and the convolution kernel size of the convolution layer may be 3. Those skilled in the art know that the first feature vector is a feature obtained after performing a one-dimensional convolution on the prediction data set, and the method of using a one-dimensional convolution layer to perform a convolution operation on the input feature vector to obtain the first feature vector may belong to the prior art.
[0033] S302, using the time modeling module to perform time extraction on the time feature in the first feature vector to obtain a second feature vector, and using the sequence aggregation module to perform a weighted sum operation on the second feature vector in the time dimension to obtain a third feature vector.
[0034] In an embodiment of the present invention, the time modeling module may include multiple time modeling units, the input end of each time modeling unit is connected to the first trajectory point embedding module, the output end of each time modeling unit is connected to the sequence aggregation module, and each time modeling unit includes a multi-head attention layer, a normalization layer and a feedforward neural network layer connected in sequence, and the multi-head attention layer, the normalization layer and the feedforward neural network layer are all existing structures.
[0035] In an embodiment of the present invention, the temporal modeling module may include 16 parallel temporal modeling units. The number of heads in the multi-head attention layer may be 4.
[0036] The step of extracting the time feature in the first feature vector by using the time modeling module to obtain the second feature vector may specifically include the following steps:
[0037] S3021, using the multi-head attention layer of each temporal modeling unit to extract the attention coefficient of the first eigenvector.
[0038] Those skilled in the art know that the method of extracting the attention coefficient of the feature vector using a multi-head attention layer may be prior art.
[0039] S3022, using the normalization layer of each temporal modeling unit to normalize the received attention coefficient to obtain a normalized coefficient. The normalized coefficient is between 0 and 1.
[0040] S3023, using the feedforward neural network layer of each time modeling unit to perform feature extraction on the received normalized coefficients to obtain a corresponding second feature vector.
[0041] Those skilled in the art know that the working principle of the feedforward neural network layer can be an existing method.
[0042] In an embodiment of the present invention, the sequence aggregation module includes a linear layer and a Gaussian error linear unit connected in sequence. The linear layer is used to map the input linear transformation to a new space, changing the dimension of the data to a higher dimension, such as 128 dimensions. In an embodiment of the present invention, the output results of multiple time modeling units are spliced together as the input of the sequence aggregation module, and the input is processed in sequence by a linear layer and a Gaussian error linear unit to obtain a third eigenvector.
[0043] S303: Encode at least one category label corresponding to the third feature vector by using the one-hot vector encoding module to obtain at least one corresponding one-hot vector.
[0044] In an embodiment of the present invention, the category label can be determined based on the category of the flight status, and the category of the flight status may include ascending, circling, cruising and descending, etc. Each category label is represented by a corresponding identification value, for example, using 0-3, etc.
[0045] S304, using the horizontal embedding module to embed the at least one one-hot vector into a second high-dimensional embedding space to obtain at least one corresponding horizontal embedding representation, and using the feature aggregation module to splice the third feature vector and each horizontal embedding representation in the at least one horizontal embedding representation to obtain at least one spliced vector.
[0046] In an embodiment of the present invention, the horizontal embedding module may be a one-dimensional convolution layer with a convolution kernel size of 3. The dimension of the second high-dimensional embedding space is greater than the dimension of the first high-dimensional embedding space, for example, it may be 128. It is known to those skilled in the art that the horizontal embedding is represented as a feature after convolution processing by a one-dimensional convolution layer.
[0047] In this embodiment of the present invention, each concatenation vector is a feature obtained by concatenating the third feature vector before the corresponding horizontal embedding representation, which can be expressed as hc=Contat(F3, he), where hc represents the concatenation vector, F3 is the third feature vector, he represents the horizontal embedding representation, and Contat() represents concatenation.
[0048] S305: Use the first feature mapping module to perform a mapping operation on the at least one concatenated vector to obtain a corresponding mapping result as a fourth feature vector.
[0049] In an embodiment of the present invention, the first feature mapping module is a multilayer perceptron, which may be composed of a first fully connected layer, a Gaussian error linear unit, a second fully connected layer, and a random dropout layer connected in sequence. The multilayer perceptron is used to perform high-dimensional embedding on the received concatenated vector and generate a final multi-level context representation. It is known to those skilled in the art that the working principle of the multilayer perceptron may be prior art.
[0050] In the embodiment of the present invention, a cascade operation of the third feature vector plus the horizontal embedding representation and a multi-layer perceptron are used to fuse these vectors and generate a multi-level context representation. The third feature vector can be regarded as a high-dimensional representation, which implies the global characteristics of the observation, while the horizontal embedding representation provides different semantic representations for each prediction horizon.
[0051] S306, using the domain conversion module to perform wavelet transform processing on the fourth eigenvector to obtain low-frequency components and high-frequency components corresponding to the fourth eigenvector, and using the second feature mapping module to perform mapping operations on the low-frequency components and high-frequency components corresponding to the fourth eigenvector to obtain corresponding mapping results as the fifth eigenvector.
[0052] In the embodiment of the present invention, a discrete wavelet transform is used to perform time-frequency analysis to extract high-level abstract trajectory embeddings with rich time features, thereby providing a robust representation of the input trajectory sequence. In the embodiment of the present invention, the context embedding can be converted from the time-frequency domain to the wavelet domain by performing a wavelet transform on the fourth eigenvector, and then down-sampling is performed to obtain the approximate coefficients (low-frequency components) and detail coefficients (high-frequency components) of the signal. The high-frequency components are regarded as the local motion details of the track, and the low-frequency classification is regarded as the overall trend of the track operation.
[0053] Those skilled in the art know that the method for obtaining the low-frequency component and the high-frequency component of the fourth eigenvector may be an existing technology.
[0054] In the embodiment of the present invention, the low-frequency component A can be expressed as: A=Re(A)+i×Im(A), Re(A) is the real part of A, Im(A) is the imaginary part of A, and i is the imaginary unit. The high-frequency component B can be expressed as: B=Re(B)+i×Im(B), Re(B) is the real part of B, and Im(B) is the imaginary part of B.
[0055] In an embodiment of the present invention, the second feature mapping module includes n feature mapping layers for capturing potential patterns of time series with global view and energy compression.
[0056] Among them, the output y of the jth feature mapping layer i The following conditions must be met:
[0057] y j =σ(Re(y j-1 )×W aj )-Im(y j-1 )×W bj +B aj )+σ(Re(y L j-1 )×W aj )-Im(y j-1 )×Wbj +B bj );
[0058] Among them, Re(y j-1 ) is the real part of the output obtained by the j-1th feature mapping layer, Im(y j-1 ) The imaginary part of the output obtained by the j-1th feature mapping layer, W aj is the real weight matrix corresponding to the jth feature mapping layer, W bj is the imaginary weight matrix corresponding to the jth feature mapping layer, B aj is the real bias value corresponding to the jth feature mapping layer, B bj is the imaginary bias value corresponding to the jth feature mapping layer, σ() represents the activation function, j ranges from 1 to n, the input of the first feature mapping layer is the low-frequency component and high-frequency component of the fourth eigenvector, y j for y L j or H j ,y L j is the output based on the low-frequency component of the fourth eigenvector, y H j is the output obtained based on the high frequency component of the fourth eigenvector, that is, each feature mapping layer will obtain two outputs, one is the output corresponding to the high frequency component, and the other is the output corresponding to the low frequency component. In the embodiment of the present invention, σ() is a Gaussian error linear unit.
[0059] S307, using the inverse wavelet transform module to perform an inverse wavelet transform operation on the sixth eigenvector to obtain a corresponding operation result as the seventh eigenvector.
[0060] In the embodiment of the present invention, an inverse wavelet transform module is used to perform an inverse wavelet transform operation on the sixth eigenvector to obtain a wavelet multi-time domain context embedding as the seventh eigenvector. Those skilled in the art will appreciate that the inverse wavelet transform operation may be implemented in the prior art.
[0061] S308, using the differential sequence encoding module to obtain the differential sequence of the prediction data set and binary encode the obtained differential sequence to obtain a differential sequence encoding feature vector, using the second trajectory point embedding module to embed the differential sequence encoding feature vector into the first high-dimensional embedding space to obtain a seventh feature vector, and using the feature fusion module to splice the seventh feature vector and the sixth feature vector to obtain a corresponding splicing result as an eighth feature vector.
[0062] In the embodiment of the present invention, the difference sequence is obtained by subtracting the previous trajectory point data from the latter trajectory point data of two adjacent trajectory point data in the prediction data set. The second trajectory point embedding module is a one-dimensional convolution layer.
[0063] S309: Use the third feature mapping module to perform a mapping operation on the time in the eighth feature vector to obtain a corresponding mapping result as a ninth feature vector.
[0064] In an embodiment of the present invention, the third feature mapping module includes a causal self-attention mechanism layer, a multi-layer perceptron and a second normalization layer connected in sequence.
[0065] Further, S309 may specifically include:
[0066] S3091, use the causal self-attention mechanism layer to obtain the attention coefficient of the eighth eigenvector and obtain the corresponding attention result.
[0067] It is known to those skilled in the art that the working principle of the causal self-attention mechanism layer may be prior art.
[0068] S3092, use the first normalization layer to normalize the attention result obtained by the first causal self-attention mechanism layer to obtain the first normalized feature vector.
[0069] In this embodiment of the present invention, the first normalization layer is used to normalize the results of the causal self-attention mechanism layer to capture the dependencies in the sequence.
[0070] S3093, using a multilayer perceptron to perform feature extraction on the feature vector after the first normalization process to obtain a corresponding extraction result.
[0071] S3094, using a second normalization layer to normalize the extraction result of the multilayer perceptron to obtain a feature vector after the second normalization.
[0072] In the embodiment of the present invention, the extraction results of the multilayer perceptron are normalized, which can stabilize and accelerate the model training process.
[0073] In an embodiment of the present invention, the embedded representation of the differential sequence in the prediction data set and the previously obtained multi-time domain context representation are connected together as prompt information for learning the differential sequence transfer mode, which can reduce the high-order prediction error when using binary coding representation.
[0074] S310, using the predictor to predict the eighth eigenvector, and obtaining a corresponding prediction result as a flight trajectory data prediction value corresponding to the prediction data set.
[0075] In an embodiment of the present invention, the predictor may be composed of a fully connected layer and a Sigmoid activation function.
[0076] In an embodiment of the present invention, the flight trajectory data prediction value may be the flight trajectory data within a set future time period. The duration of the set future time period may be set based on actual needs, and in an exemplary embodiment, may be 30 minutes.
[0077] Furthermore, in an embodiment of the present invention, the trained flight trajectory prediction model is obtained through the following steps:
[0078] S10, obtaining historical flight trajectory data of a set time period and performing preprocessing to obtain a historical time series data set.
[0079] In the embodiment of the present invention, the set time period can be set based on actual needs, such as the flight trajectory data of a certain flight within 45 days.
[0080] S20, binary encoding the historical time series data set, using the encoded data set as a sample data set, and dividing the sample data set into a training set, a validation set, and a test set according to a preset ratio.
[0081] In an embodiment of the present invention, the division ratio of the training set, the validation set and the test set can be determined based on actual needs. In the case where the data set is 45 days of data, the trajectory data of the first 40 days can be used as the training set, the trajectory data of the 41st day can be used as the validation set, and the trajectory data of the last four days can be used as the test set.
[0082] S30, using the training set to train multiple initial flight trajectory prediction models with different hyperparameters, to obtain multiple trained initial flight trajectory prediction models.
[0083] In an embodiment of the present invention, the daily flight trajectory data in the training set can be divided into a plurality of training sample data according to the first set time window, each training sample data includes an input feature variable and a target variable, the input feature variable is the flight trajectory data within the first set time window, and the target variable is the flight trajectory data within the second set time window after the first set time window. The duration of the second set time window can be less than or equal to the duration of the first set time window.
[0084] In the training process of each initial flight trajectory prediction model, sample training data can be input in batches for training, and the specific training process can be the existing technology. The model processing process of the training sample data in the training process can operate the steps defined in S301 to S310.
[0085] S50, using the verification set to verify the prediction effects of multiple trained initial flight trajectory prediction models, obtain corresponding verification results, and select the trained initial flight trajectory prediction model with the best prediction effect as the trained flight trajectory prediction model based on the verification results.
[0086] The validation process of the validation set for the model can be the same as the training process. The present invention uses root mean square error (MSE), mean absolute error (MAE) and mean deviation error (MDE) to evaluate the prediction effect of the prediction model.
[0087] S60, inputting the test set into the trained flight trajectory prediction model to obtain a flight trajectory data prediction value, and obtaining the prediction accuracy of the trained flight trajectory prediction model based on the obtained flight trajectory data prediction value and the actual flight trajectory data value, so as to evaluate the prediction performance of the trained flight trajectory prediction model.
[0088] In the embodiments of the present invention, all experiments are built and executed based on the PyTorch1.4 framework. In the implementation, the wavelet transform based on PyTorch is used to support the operations of discrete wavelet transform and inverse discrete wavelet transform. The optimizer selects Adam to update the trainable parameters of the model. The initial learning rate is set to 0.001, and the learning rate is decayed at a rate of 0.5 every 10 epochs. The model configuration is trained and optimized on a smaller data set.
[0089] The present invention uses the same data set to perform prediction using the flight trajectory prediction module, WTFTP, TimeVQVAE and FreTST provided by the present invention, and the prediction results are shown in Table 1 below:
[0090] Table 1
[0091]
[0092] Among them, WTFTP is the abbreviation of Wavelet Transform-based Flight Trajectory Prediction, and its Chinese name is flight trajectory prediction based on wavelet transform. TimeVQVAE is the alias of Vector Quantized Time Series Generation with a Bidirectional Prior Model, and its Chinese name is vector quantized time series generation model based on bidirectional prior model. FreTST is the abbreviation of Frequency-domain MLPs of time series technology, and its Chinese name is frequency domain MLP technology of time series.
[0093] As can be seen from Table 1, the flight trajectory prediction model provided by the present invention surpasses other benchmark models in all indicators of longitude, latitude and altitude, and shows the best overall performance, except that the root mean square error in altitude is slightly inferior. This shows the performance advantage of the model of the present invention in the task of flight trajectory prediction, and proves the effectiveness of time-frequency analysis, that is, the time-frequency representation of wavelet components enables it to fully examine the slow and fast dynamic characteristics of the flight trajectory. With the ability of time-frequency analysis and deep modeling of flight trends and motion details, the model provided by the present invention reduces the average absolute error and average relative error in longitude and latitude by at least 30% and 35%, respectively, compared with the benchmark model. In addition, the model provided by the present invention significantly improves the practicality, reduces the average distance error by about 16%, and makes the prediction error less than 6.71 kilometers. This MDE result makes this method have high potential in practical applications because it has a smaller deviation when predicting trajectory points.
[0094] 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.
[0095] The 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.
[0096] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.
[0097] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A flight trajectory prediction method, characterized in that: The method is implemented based on a trained flight trajectory prediction model, which includes: a trajectory encoder, a horizontal perception context generator, a frequency learning machine based on wavelet transform, a differential prompt decoder and a predictor; the trajectory encoder includes a first trajectory point embedding module, a time modeling module and a sequence aggregation module connected in sequence, the horizontal perception context generator includes a one-hot vector encoding module, a horizontal embedding module, a feature aggregation module and a first feature mapping module, the frequency learning machine based on wavelet transform includes a domain conversion module, a second feature mapping module and an inverse wavelet transform module, and the differential prompt decoder includes a differential sequence encoding module, a second trajectory point embedding module, a feature fusion module and a third feature mapping module; the inputs of the trajectory encoder and the differential sequence encoding module are model input data, the output of the sequence aggregation module is used as the input of the one-hot vector encoding module, the output of the inverse wavelet transform module is used as the input of the feature fusion module, and the output of the third feature mapping module is used as the input of the predictor; Wherein, the method comprises the following steps: S100, preprocessing the flight trajectory data to be predicted to obtain a time series data set; S200, binary encoding the time series data set to obtain an encoded data set as a prediction data set; S300, inputting the prediction data set into a trained flight trajectory prediction model to obtain a corresponding prediction result as a prediction value of the flight trajectory data to be predicted.
2. The method according to claim 1, characterized in that in, S300 specifically includes: S301, embedding the prediction data set into a first high-dimensional embedding space using the first trajectory point embedding module to extract distinguishing spatial features in the prediction data set to obtain a first feature vector; S302, using the time modeling module to perform time extraction on the time feature in the first feature vector to obtain a second feature vector, and using the sequence aggregation module to perform a weighted sum operation on the second feature vector in the time dimension to obtain a third feature vector; S303, using the one-hot vector encoding module to encode at least one category label corresponding to the third feature vector to obtain at least one corresponding one-hot vector; S304, using the horizontal embedding module to embed the at least one one-hot vector into a second high-dimensional embedding space to obtain at least one corresponding horizontal embedding representation, and using the feature aggregation module to concatenate the third feature vector and each horizontal embedding representation in the at least one horizontal embedding representation to obtain at least one concatenated vector; S305, using the first feature mapping module to perform a mapping operation on the at least one concatenated vector to obtain a corresponding mapping result as a fourth feature vector; S306, using the domain conversion module to perform wavelet transform processing on the fourth eigenvector to obtain a low-frequency component and a high-frequency component corresponding to the fourth eigenvector, and using the second eigenmapping module to perform a mapping operation on the low-frequency component and the high-frequency component corresponding to the fourth eigenvector to obtain a corresponding mapping result as a fifth eigenvector; S307, using the inverse wavelet transform module to perform an inverse wavelet transform operation on the sixth eigenvector to obtain a corresponding operation result as the seventh eigenvector; S308, using the differential sequence encoding module to obtain the differential sequence of the prediction data set and binary encoding the obtained differential sequence to obtain a differential sequence encoding feature vector, using the second trajectory point embedding module to embed the differential sequence encoding feature vector into the first high-dimensional embedding space to obtain a seventh feature vector, and using the feature fusion module to splice the seventh feature vector and the sixth feature vector to obtain a corresponding splicing result as an eighth feature vector; S309, using the third feature mapping module to perform a mapping operation on the time in the eighth feature vector to obtain a corresponding mapping result as a ninth feature vector; S310, using the predictor to predict the ninth eigenvector, and obtaining a corresponding prediction result as a flight trajectory data prediction value corresponding to the prediction data set.
3. The method according to claim 2, characterized in that The first trajectory point embedding module, the second trajectory point embedding module and the horizontal embedding module are all composed of a one-dimensional convolutional layer.
4. The method according to claim 2, characterized in that: The temporal modeling module includes multiple temporal modeling units, the input end of each temporal modeling unit is connected to the first trajectory point embedding module, the output end of each temporal modeling unit is connected to the sequence aggregation module, and each temporal modeling unit includes a multi-head attention layer, a normalization layer and a feedforward neural network layer connected in sequence.
5. The method according to claim 2, characterized in that: The sequence aggregation module includes linear layers and Gaussian error linear units connected in sequence.
6. The method according to claim 2, characterized in that The first feature mapping module is a multi-layer perceptron, and the third feature mapping module includes a causal self-attention mechanism layer, a first normalization layer, a multi-layer perceptron, and a second normalization layer connected in sequence.
7. The method according to claim 2, characterized in that The second feature mapping module includes n feature mapping layers, wherein the output y of the jth feature mapping layer i The following conditions must be met: y j σ(Re(y j-1 )×W aj )-and(y j-1 )×W bj +B aj )+σ(Re(y L j-1 )×W aj )-and(y j-1 )×W bj +B bj ): Among them, Re(y j-1 ) is the real part of the output obtained by the j-1th feature mapping layer, Im(y j-1 ) is the imaginary part of the output obtained by the j-1th feature mapping layer, W aj is the real weight matrix corresponding to the jth feature mapping layer, W bj is the imaginary weight matrix corresponding to the jth feature mapping layer, B aj is the real bias value corresponding to the jth feature mapping layer, B bj is the imaginary bias value corresponding to the jth feature mapping layer, σ() represents the activation function, j ranges from 1 to n, the input of the first feature mapping layer is the low-frequency component and high-frequency component of the fourth eigenvector, y j for y L j or H j ,y L j is the output based on the low-frequency component of the fourth eigenvector, y H j is the output obtained based on the high frequency component of the fourth eigenvector.
8. The method according to claim 1, characterized in that: The flight trajectory data includes longitude, latitude, altitude and speed.
9. An electronic device, characterized in that: including a processor and a memory; The processor is used to execute the steps of the method according to any one of claims 1 to 8 by calling the program or instruction stored in the memory.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store a program or an instruction, wherein the program or the instruction enables a computer to execute the steps of the method according to any one of claims 1 to 8.