A fast method for unsteady aerodynamic force calculation based on attention mechanism

Abstract

The application discloses a kind of based on the quick calculation method of unsteady aerodynamic force of attention mechanism, first based on unsteady flow solving method calculation research object in given movement law in unsteady aerodynamic force;Then the movement information of given research object is summarized with the unsteady aerodynamic force obtained by unsteady solver calculation, constructs database;Next corresponding database is further divided into training set, verification set and prediction set, and is loaded to the time series neural network based on attention mechanism;Again, the training set is used to train the parameters of the time series neural network, and the network training condition is tested according to the verification set, until the network accuracy meets the actual engineering application requirement;After that, the trained network is tested using the prediction set, and according to the prediction set test result, the network is selected to be saved or regenerated.The method can provide reference for the unsteady aerodynamic force solving module of aircraft design, flow control and other problems, and shorten the research period of the corresponding problem.

Description

Technical Field

[0001] The present invention belongs to the technical field of aerodynamics, and in particular relates to a method for fast calculation of unsteady aerodynamic forces based on an attention mechanism. Background Art

[0002] In the research of modern fluid and aerospace fields, unsteady aerodynamics is an important component, and the solution of unsteady aerodynamic forces is the main research content of unsteady aerodynamics. On the one hand, the solution of unsteady aerodynamic forces is the core of unsteady flow problems and fluid-structure interaction problems, and is an indispensable link in the corresponding problems; on the other hand, in problems such as flow mechanism analysis, aerodynamic shape optimization, and unsteady flow control, the solution of unsteady aerodynamic forces will directly guide the research of subsequent problems and play a vital guiding role. Therefore, during the research process, the accuracy and efficiency of the solution of unsteady aerodynamic forces have always been one of the key focuses of researchers.

[0003] In current research, the main method for solving unsteady aerodynamic forces for aerodynamic problems and fluid-structure coupling problems is still to use traditional CFD calculations or CFD / CSD coupling calculations. The main advantage of this method being widely used is its strong adaptability to complex unsteady flow problems and the ability to achieve solution accuracy that meets engineering and research requirements by selecting high-order formats. However, with the increasing complexity of engineering and research objects in recent years, the drawbacks of traditional numerical simulation technology have become increasingly apparent: for complex unsteady flow and fluid-structure coupling problems, the calculation cycle of numerical simulation methods is greatly extended compared to simple models due to factors such as the increased difficulty of discretization pre-processing and the extended single-step operation time during iterative calculations. Within a limited time, the demand for computing power is also greatly increased, and the cost of obtaining reliable data is too high, making it difficult to use in engineering and research. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the application provides a non-constant aerodynamic force rapid calculation method based on an attention mechanism, first, a non-constant aerodynamic force of a research object in a given motion law is calculated based on a non-constant flow solving method; then, motion information of the given research object and the non-constant aerodynamic force calculated by the non-constant solver are summarized and arranged to construct a database; next, the corresponding database is further divided into a training set, a verification set and a prediction set, and is loaded into a time series neural network based on an attention mechanism; then, the training set is used for parameter training of the time series neural network, and the network training condition is verified according to the verification set, until the network precision obtained reaches the actual engineering application requirement; after that, the trained network is tested by using the prediction set, and according to the prediction set test result, the network is selected to be saved or regenerated. The method can provide a reference for a non-constant aerodynamic force solving module of an aircraft design, flow control, aerodynamic optimization and the like, and greatly shortens the research period of the corresponding problem.

[0005] The technical scheme adopted by the application to solve the technical problems is as follows:

[0006] Step 1: according to the research object, for a given flow condition, the motion law of the research object is designed and recorded; the non-constant aerodynamic force of the research object in the given motion law is calculated based on a non-constant flow solving method;

[0007] Step 2: the motion information of the given research object and the non-constant aerodynamic force calculated by the non-constant solver are summarized and arranged to construct a database;

[0008] Step 3: the corresponding database is further divided into a training set, a verification set and a prediction set, and is loaded into a time series neural network based on an attention mechanism;

[0009] Step 4: the training set is used for parameter training of the time series neural network, and the network training condition is verified according to the verification set, until the network precision obtained reaches the actual engineering application requirement;

[0010] Step 5: the trained network is tested by using the prediction set, and according to the prediction set test result, the network is selected to be saved or regenerated;

[0011] Step 6: the network that passes the prediction set test is saved, and the network is applied to the calculation of the non-constant aerodynamic force subsequently.

[0012] Preferably, the given flow condition refers to environmental parameters of the corresponding non-constant flow, including a Mach number, a Reynolds number, a temperature, a density and a pressure.

[0013] Preferably, the motion law of the research object refers to displacement, generalized displacement, velocity and generalized velocity of the research object when the research object remains stationary or performs rigid body motion according to a certain motion law.

[0014] Preferably, the unsteady aerodynamic force includes lift, drag, moment, other generalized forces, and their corresponding coefficients.

[0015] Preferably, the specific form of the database is that the displacement of the research object at each time and the aerodynamic force present a one-to-one correspondence relationship, and in the subsequent network training and testing process, the displacement is the input data and the aerodynamic force is the output data.

[0016] Preferably, the division ratio of the training set, the verification set and the test set is 7:2:1 or 8:1:1.

[0017] Preferably, the training set is used to train the time series neural network, and the network training condition is checked according to the verification set. The network internal parameters are updated by continuously introducing the displacement of the training set into the network, and the network training condition is checked in real time by continuously introducing the verification set until the error between the aerodynamic force corresponding to the network output and the actual aerodynamic force is less than a threshold value.

[0018] Preferably, according to the test result of the prediction set, the network is selected to be retained or retrained, specifically: an error threshold value is given in advance according to experience or actual application requirements, then the displacement corresponding to the prediction set is input into the trained network, the aerodynamic force result output by the network is collected, and an error comparison is made with the real aerodynamic force calculated by the solver. If the error is less than the threshold value, the network is saved; otherwise, the network needs to be regenerated.

[0019] Preferably, the time series neural network based on attention mechanism is L-Transformer, which includes:

[0020] The data set loading series module is used to load data into the network in batches, and the network is trained and tested;

[0021] The position coding module PositionalEncoding is used to inject position information into the data loaded into the network;

[0022] The multi-head attention series module MultiHeadAttention with a mask function is used to integrate the internal features of the data after adding position coding and extract the internal relationship features of the data; wherein the mask function is used to eliminate the influence of future data on previous data in the network learning process;

[0023] The residual connection and layer normalization module AddNorm is used to normalize the data processed by the attention mechanism, accelerate the training process, and improve the network performance;

[0024] The position-based feedforward network module PositionWiseFFN is used to capture the long-distance dependence relationship between the data after feature reorganization.

[0025] EncoderBlock, containing a plurality of multi-head attention series modules, residual connection and layer normalization modules, and a position-based feedforward network module, is used for feature induction and learning of input data;

[0026] DncoderBlock, containing a multi-head attention series module with a mask function, a plurality of multi-head attention series modules, residual connection and layer normalization modules, and a position-based feedforward network module, is used to integrate input data information and information processed by the encoder, to realize time series data recognition and prediction;

[0027] TransformerEncoder, containing a linear layer linear, a position encoding module, and a plurality of EncoderBlock, is used to induce feature information of displacement data and guide the decoder to identify aerodynamic force;

[0028] TransformerDecoder, containing a linear layer linear, a position encoding module, a plurality of DecoderBlock, and a hidden state initialization module, is used to obtain aerodynamic force using input displacement information under the guidance of the encoder;

[0029] EncoderDecoder, used to combine the encoder module and the decoder module to build a complete network framework for engineering applications;

[0030] Training series module, used to load training set and validation set data in the database to update network parameters;

[0031] grad_clipping, used to prevent gradient explosion during training and improve network adaptability to data;

[0032] predict_seq2seq, used to load test set to verify whether the network is trained;

[0033] In the data set loading series module, a continuity truncation mechanism is used to process the original data to minimize the loss of input data information and prevent the loss of training efficiency caused by excessive data processing by the network at one time;

[0034] In the data set loading series module, an additional preposition aerodynamic force is added to the aerodynamic force input used in the training phase to force teaching during training and improve training effect;

[0035] In the EncoderBlock, the original mask mechanism in the encoder is removed to ensure complete extraction of input data features;

[0036] In the encoder module and the decoder module, the linear layer linear is used instead of the original embedding layer embedding, which is suitable for dimension conversion of sequence data;

[0037] In the training series module, the exponential mean square error expMSEloss is used instead of the original mean square error MSEloss, so that the training efficiency and network performance are improved;

[0038] In the prediction module, a sliding window mechanism is added, which can adjust the encoder input according to the hysteresis of the input data, and corresponding adjustment parameters are set, so that the network is better applied to the unsteady flow problem containing hysteresis characteristics such as hysteresis loop;

[0039] In the prediction module, a long tail truncation mechanism is introduced, and corresponding adjustment parameters are set, according to the premise that most unsteady aerodynamic forces have limited time delay effect, when updating the decoder hidden state at the current time, the influence of the earlier output data is removed, to prevent the prediction efficiency from decreasing with the sequence length when long sequence prediction.

[0040] Preferably, the error of the corresponding network output aerodynamic force and the actual aerodynamic force is less than the error in the threshold, the evaluation error includes mean square error MSEloss, mean absolute error MAEloss, relative error REloss and exponential mean square error expMSEloss.

[0041] The beneficial effects of the present application are as follows:

[0042] The method of the present application can be directly used for rapid solution of unsteady aerodynamic force of related problems such as aerospace, civil engineering and wind engineering, and can also provide reference for unsteady aerodynamic force solution module of problems such as aircraft design, flow control and aerodynamic optimization, greatly shortening the research period of corresponding problems. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The flow chart of the rapid calculation method proposed in the present application is shown in the figure;

[0044] Figure 2 The flow field grid diagram of the embodiment of the present application is shown in the figure;

[0045] Figure 3 The architecture diagram of the time sequence neural network L-Transformer based on attention mechanism of the present application is shown in the figure. DETAILED DESCRIPTION

[0046] The present application will be further described below in combination with the drawings and embodiments.

[0047] The purpose of the present application is to solve the problems in the prior art and provide a fast calculation method of unsteady aerodynamic force based on attention mechanism. The method obtains training data through fluid dynamics simulation, trains a neural network based on attention mechanism as a proxy model, and realizes its engineering application. For unsteady flow problems, the model can quickly solve the unsteady aerodynamic force in the problem of the research object being static or rigid body motion by only giving the known motion information of the research object.

[0048] Embodiments:

[0049] The present application selects a single degree of freedom oscillation working condition of NACA0012 airfoil at a certain transonic speed condition, and the corresponding flow field grid node distribution is as shown in Figure 2 The specific details of the working condition are as follows:

[0050] Airfoil geometric parameters: airfoil chord length 1, airfoil initial angle of attack 0°, training set and verification set oscillation amplitude [0°, 6°], training set and verification set oscillation reduction frequency [0.06, 0.12], predicted working condition oscillation amplitude 1°, predicted working condition oscillation reduction frequency 0.09;

[0051] Uniform incoming flow Mach number 0.755, temperature 300K, pressure 101325Pa.

[0052] Referring to Figure 1 The specific flow details of a fast calculation method of unsteady aerodynamic force based on attention mechanism are as follows:

[0053] First, the angle of attack of the airfoil at each time under single degree of freedom oscillation is selected as the generalized displacement, and random signals with certain amplitude and reduction frequency and sinusoidal signals under the predicted working condition are designed as training signals and prediction signals, respectively. The unsteady aerodynamic force under the corresponding signals is solved by using a pressure-based unsteady solver, and the lift coefficient and moment coefficient at each time are extracted as the generalized aerodynamic force to be modeled.

[0054] Then, each generalized displacement and the corresponding generalized aerodynamic force of the current time displacement are one-to-one corresponding, and the database required by the network is constructed.

[0055] Wherein, the training signal and its corresponding generalized aerodynamic force are divided by 8:2 as the data source of the training set and the validation set, and further, the data in the training set and the validation set are truncated continuously according to the adjacent time for the corresponding generalized displacement-generalized aerodynamic force pair, and the length of each segment is a hyperparameter. At the same time, in order to further improve the training effect of the network, for the generalized aerodynamic force part, another pre-item is constructed to store the generalized aerodynamic force at the previous time of each breakpoint, and it is paired with the original generalized displacement-generalized aerodynamic force data pair one by one to form a generalized displacement-generalized aerodynamic force-pre-item aerodynamic force data pair for subsequent teaching training. For the prediction signal and its corresponding generalized aerodynamic force as the data source of the prediction set. Then, the divided training set is loaded into the L-Transformer time series neural network.

[0056] The specific internal architecture of the L-transformer neural network is shown in Figure 3 , and the flow after the training set is loaded into the network is as follows:

[0057] Firstly, the training set is unpacked into three parts of generalized displacement, generalized aerodynamic force and pre-item aerodynamic force before being input into the network in batches.

[0058] Among them, the generalized displacement as the input item in training is loaded into the encoder module TransformerEncoder of the network in batches, which is used to extract internal features to guide the prediction of the generalized aerodynamic force in the decoder. After the generalized displacement item is input into the encoder module, it is first passed through the modified linear layer linear to deconstruct the underlying features of the time series data; then, the deconstructed data is passed through the position encoding module PositionalEncoding to add position encoding to each feature dimension, which gives the parallel input data position information for subsequent time series feature learning, wherein the position encoding can adopt a trigonometric function type:

[0059]

[0060] In the formula, i is the time order of the data at the current position in the current batch, 2j and 2j+1 are the feature order of the data at the current position in the current batch for even and odd items respectively, d is the total number of features of the current batch data, and p is the position encoding value of the data at the current position.

[0061] Then, the data after adding position encoding is sequentially passed through a plurality of encoder basic unit blocks EncoderBlock to summarize and extract underlying features, wherein the number of encoder basic unit blocks is a hyperparameter. In theory, the more the number of blocks, the stronger the ability of the encoder to extract underlying features of the data, so in use it needs to be reasonably set according to the data size.

[0062] The data entering the encoder basic unit block is sequentially processed through four steps. First, the data passes through the MultiHeadAttention module with a mask-free mechanism. This module uses a multi-head self-attention mechanism to self-reorganize the time dimension of the input data. In a single batch example, the expression can be represented as:

[0063] y i =f(x i ,(x1,x1),...(x n ,x n )) (2)

[0064] In the formula, x1.....x n represent an input sequence with d feature dimensions, n represents the sequence length, and f represents an attention aggregation function, the general expression of which can be represented as:

[0065]

[0066] In the formula, q, k, and v represent query, key, and value, respectively, which are vectors of q, k, and v dimensions. The attention mechanism is the core part, and its working mechanism is to obtain the weight coefficient by solving the similarity of the query and the key, so as to determine the weighted sum of the value. Further, a can be represented as a function of the attention score function a:

[0067]

[0068] In the commonly used attention mechanism-based Transformer series network, the scaled dot-product attention is usually selected as the scoring function. Its essence is to directly perform inner product operation on the query and the key, and to obtain the corresponding weight by using feature dimension standardization. The expression is:

[0069]

[0070] In the L-Transformer network encoder, the multi-head self-attention module is used to extract the internal relationship features of the data from the input generalized displacement itself.

[0071] Then, the data integrated through the multi-head self-attention passes through the AddNorm module, which can further normalize the data processed by the functional layer, and is used to assist in building an effective deep architecture to improve network performance.

[0072] The normalized data then passes through the PositionWiseFFN module based on the position. This module can be regarded as a two-layer feedforward neural network layer, which is used to capture long-distance dependency relationships in the data. This module determines that the architecture can be naturally extended to long sequence recognition or prediction problems.

[0073] Subsequently, the data processed by the position feedforward processing is further normalized by the residual connection and layer normalization module AddNorm. At this time, the processing process of the single encoder basic unit module is completed, and the output result of the module can be further used to input the next basic unit block or directly introduced into the multi-head attention layer of the decoder module TransformerDecoder to guide the decoder to identify or predict the generalized aerodynamic force.

[0074] Unlike the parallel features of the encoder, the decoder of the L-Transformer is a time mechanism constructed by attention, which has the typical sequence characteristics of the time network. In use, the generalized aerodynamic force and the pre-item aerodynamic force are taken as the input items of the decoder module TransformerDecoder. In order to further improve the training efficiency and accuracy of the network, the teacher forcing method is used to train the decoder of the L-Transformer network. Among them, for a single batch of data, the last item of the input generalized aerodynamic force is truncated in the time dimension, and the pre-item aerodynamic force is added at the beginning of the dimension, and the generalized aerodynamic force is taken as the output item. The purpose of this operation is to enable the decoder to learn the ability to infer the next item according to the information of the previous several items.

[0075] Similar to the encoder part, the spliced decoder input item is sequentially processed by the linear layer linear and the position encoding module PositionalEncoding, respectively used to deconstruct the underlying features and add position information.

[0076] Then, for the data after adding position encoding, it is sequentially processed by several decoder basic unit blocks DecoderBlock. Under the guidance of the generalized displacement feature information transmitted by the encoder, this module realizes the prediction of the generalized aerodynamic force. Compared with the encoder basic unit block, there are mainly two different units in this part, the multi-head attention series module MultiHeadAttention with a mask mechanism and the multi-head attention series module MultiHeadAttention with external input. To avoid repetition, only the two modules are introduced here, and the other parts have the same functions as the encoder, so they are not described again.

[0077] First is the multi-head attention series module MultiHeadAttention with a mask mechanism. Compared with the non-mask module, the module adds a mask mechanism, mainly in the training link, so that the multi-head attention module calculates the weight only to the previous time when the data at a certain time is calculated. When calculating the attention of the future time, the weight is zeroed. This mechanism can ensure that the decoder meets the actual physical mechanism when predicting, only considering the data feature information obtained.

[0078] In addition, the multi-head attention series module MultiHeadAttention contains external input. Unlike other self-attention mechanism modules based on current input sequence information, this module receives generalized displacement feature information from the encoder as the key-value pair, and uses the feature of the input data of the previous layer of the encoder as the query, which can realize the introduction of encoder information, and thus achieve the effect of experience guidance and information fusion.

[0079] The input data of the decoder is processed by several decoder basic unit blocks DecoderBlock, and then output through the linear layer linear. At this time, the output dimension is set to 1, and the network predicted generalized aerodynamic force can be obtained. The error is calculated with the real generalized aerodynamic force of the current batch, and the relationship with the threshold value is used to determine whether the next batch of data needs to be input for network training. At the same time, according to the error calculation result, the deviation is updated through the backpropagation through time (BPTT) mechanism to update the gradient of each parameter in the network. Here, the learning rate, batch size, training iteration step, and other hyperparameters can be adjusted to adjust the training speed to obtain the final required network model.

[0080] After a period of training, the validation set can be loaded into the network, and its data transmission process in the network is exactly the same as that of the training set. Its role is to determine whether the network training is complete. If the error calculated by the validation set is less than the threshold value, the update of the network parameters can be stopped immediately, and the obtained network is the required proxy model.

[0081] Then, for the trained network, the prediction set is used to test the network performance. The difference between the prediction set and the training set and the validation set is that the prediction set only has the initial time generalized aerodynamic force information and the generalized displacement information at each time when passing through the network layer, and does not contain additional real generalized aerodynamic force information. Therefore, the data passing through the network layer is different from the training set and the validation set as follows:

[0082] On the one hand, in the prediction stage, the L-Transformer network introduces a sliding window mechanism, which can select the generalized displacement at different times through the sliding window to predict the corresponding generalized aerodynamic force. For the hysteresis loop problem selected in this example, the time of the selected generalized displacement can be appropriately moved backward, thereby better capturing the hysteresis effect.

[0083] On the other hand, in the prediction stage, to update the hidden state of the decoder, the L-Transformer network uses a long tail truncation mechanism. This mechanism believes that the influence time of unsteady effects is limited and only affects the unsteady aerodynamic force output within a limited time. Therefore, the generalized aerodynamic force data of the early output can be truncated. This strategy can effectively prevent the need for all previous generalized displacement inputs to predict the generalized aerodynamic force at a new time during the prediction process, greatly improving the prediction efficiency of long sequences.

[0084] For networks verified by the test set, they can be saved and used for subsequent practical engineering applications. If the network itself does not pass the test, it needs to return to the previous several times, check various problems, and retrain. Possible problems include the accuracy of the data set, the setting of the network hyperparameters, the rationality of the network structure, and whether the network training times are sufficient.

Claims

1. A method for fast computation of unsteady aerodynamic forces based on attention mechanism, characterized in that, Comprise the following steps: Step 1: according to the research object, for a given flow condition, design and record the motion law of the research object; based on the unsteady flow solving method, the unsteady aerodynamic force of the research object in the given motion law is calculated; Step 2: the motion information of the given research object and the unsteady aerodynamic force calculated by the unsteady solver are summarized and arranged to build a database; The specific form of the database is that the displacement of the research object at each time and the aerodynamic force present one-to-one correspondence, and in the subsequent network training and testing process, the displacement is the input data, and the aerodynamic force is the output data; Step 3: the corresponding database is further divided into training set, validation set and prediction set, and loaded into the time series neural network based on attention mechanism; The time series neural network based on attention mechanism is L-Transformer, comprising: Data set loading series module, used for loading data into network by batch, and network training and testing; Positional encoding module PositionalEncoding, used for injecting position information into the data loaded into the network; Multi-head attention series module MultiHeadAttention with mask function, used for internal feature integration and extraction of internal relationship features of the data after adding position coding; wherein the mask function is used to eliminate the influence of future data on previous data in the network learning process; Residual connection and layer normalization module AddNorm, used for normalizing the data processed by the attention mechanism, accelerating the training process and improving the network performance; Position-based feedforward network module PositionWiseFFN, used for capturing long-distance dependence relationship between data after feature reorganization; Encoder basic unit block EncoderBlock, containing a plurality of multi-head attention series modules, residual connection and layer normalization modules and position-based feedforward network modules, used for feature induction and learning of input data; Decoder basic unit block DncoderBlock, containing multi-head attention series module with mask function, a plurality of multi-head attention series modules, residual connection and layer normalization modules and position-based feedforward network modules, used for integrating input data information and information processed by the encoder to realize time series data identification and prediction; Encoder module TransformerEncoder, containing linear layer linear, position coding module and a plurality of encoder basic unit blocks, used for inducing feature information of displacement data and guiding the decoder to identify aerodynamic force; Decoder module TransformerDecoder, containing linear layer linear, position coding module, a plurality of decoder basic unit blocks and hidden state initialization module, used for obtaining aerodynamic force by input displacement information under the guidance of the encoder; Encoder-decoder combination module EncoderDecoder, used for combining the encoder module and the decoder module to build a complete network framework for engineering application; Training series module, used for loading training set and validation set data in the database to update network parameters; The gradient clipping module grad_clipping is used to prevent gradient explosion during training and improve the network's adaptability to data. The prediction module predict_seq2seq is used to load the test set to verify whether the network is trained. In the data set loading series module, a continuity truncation mechanism is used to process the original data to minimize the loss of input data information and prevent the network from processing excessive data at once, which may cause a loss of training efficiency. In the data set loading series module, a preposition aerodynamic force is added to the aerodynamic force input for the training phase to improve the training effect. In the encoder basic unit block, the original mask mechanism in the encoder is removed to ensure complete extraction of input data features. In the encoder module and the decoder module, the linear layer linear is used instead of the original embedding layer embedding to adapt to the dimension conversion of sequence data. In the training series module, the exponential mean square error expMSEloss is used instead of the original mean square error MSEloss to improve the training efficiency and network performance. In the prediction module, a sliding window mechanism is added to adjust the encoder input according to the input data's hysteresis, and the corresponding adjustment parameters are set to make the network better apply to non-steady flow problems with hysteresis characteristics such as hysteresis loops. In the prediction module, a long tail truncation mechanism is introduced, and the corresponding adjustment parameters are set. Based on the premise that most non-steady aerodynamic forces have limited time delay effects, the influence of early output data is removed when updating the decoder hidden state at the current time, preventing the prediction efficiency from decreasing with the sequence length during long sequence prediction. Step 4: Use the training set to train the time series neural network, and use the validation set to verify the network training until the network accuracy meets the actual engineering application requirements. Step 5: Test the trained network using the prediction set, and based on the prediction set test results, choose to save or regenerate the network. Step 6: Save the network that passes the prediction set test, which will be applied to non-steady aerodynamic force calculation in the future.

2. The method of claim 1, wherein, The given flow condition refers to the environmental parameters of the corresponding non-steady flow, including Mach number, Reynolds number, temperature, density, and pressure.

3. The method for fast calculation of unsteady aerodynamic forces based on attention mechanism according to claim 1, characterized in that: The motion law of the research object refers to the displacement, generalized displacement, velocity, and generalized velocity of the research object when it remains stationary or moves according to a certain motion law.

4. The method of claim 1, wherein, The non-steady aerodynamic force includes lift, drag, moment, other generalized forces, and their corresponding coefficients.

5. The method of claim 1, wherein, The training set, validation set, and test set division ratio is 7:2:1 or 8:1:

1.

6. The method of claim 1, wherein, The use of the training set to train the time series neural network and the use of the validation set to verify the network training refer to continuously introducing training set displacement into the network to update the network's internal parameters, while continuously introducing the validation set to check the network training in real time until the error between the aerodynamic force output by the current network and the actual aerodynamic force is less than the threshold.

7. The method of claim 1, wherein, The method comprises the following steps: according to the prediction set test result, selecting to reserve or retrain the network, specifically: an error threshold is given in advance according to experience or actual application requirement, then the corresponding displacement input of the prediction set is input into the trained network, the aerodynamic force result output by the network is collected, and the error is compared with the real aerodynamic force calculated by the solver, if the error is less than the threshold, the network is saved, otherwise the network needs to be regenerated.

8. The method of claim 6, wherein the method is characterized by, The error in the error less than the threshold between the corresponding network output aerodynamic force and the actual aerodynamic force adopts an evaluation error, including mean square error MSEloss, mean absolute error MAEloss, relative error REloss and exponential mean square error expMSEloss.

Images