A d-pod and bi-lstm-based fast prediction method for transient flow field of turbomachinery

Abstract

The application discloses a kind of d-POD and Bi-LSTM-based impeller machine transient flow field fast prediction method. Construct full-cycle flow field snapshot matrix;Through data acquisition code, the physical quantity of flow field area of interest is extracted from simulation data, Hankel augmented matrix is constructed and standardization is completed;The time series augmented original matrix constructed is decomposed by d-POD, and the spatial mode matrix, time coefficient matrix and mode energy matrix are obtained;Bi-LSTM model is trained;The historical time series time coefficient of the predicted working condition is outputed by the model to predict the time coefficient sequence of future time;The physical quantity data reconstructed is matched and bound with the node coordinates and topological structure information of the original flow field grid one by one, to supplement the complete grid information for the restored physical quantity, according to the CFD post-processing standard format.dat File is generated, and the visualization output and subsequent engineering application of predicted flow field are completed.The present application is used to solve the core dilemma that precision and cost are difficult to consider in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of rapid prediction technology of transient flow field in turbomachinery, specifically involving a rapid prediction method of transient flow field in turbomachinery based on d-POD and Bi-LSTM. Background Technology

[0002] Accurate and efficient acquisition of the pressure and temperature fields on the blade surface is a core element in studying the aerodynamic characteristics of turbomachinery and analyzing its complex internal flow mechanisms. It is also a crucial foundation for conducting fluid-structure interaction analysis, fatigue life assessment, and flutter suppression design of blades. As a heat-work conversion component in core equipment for defense and energy sectors such as aero-engines and gas turbines, the internal flow of turbines exhibits complex characteristics of strong nonlinearity and multi-physics coupling. In particular, during the process of variable geometry turbines achieving wide operating condition adaptation through guide vane installation angle adjustment, the spatiotemporal distribution characteristics of blade aerodynamic loads undergo drastic changes, placing extremely high demands on the accuracy of aerodynamic parameter acquisition, response speed, and full-condition adaptability.

[0003] However, current traditional research methods for turbine aerodynamic parameters always face the core dilemma of balancing accuracy and cost: At the experimental measurement level, the mainstream discrete point measurement method is essentially a "sampling snapshot" of continuous dynamic flow. It not only fails to cover key areas with drastic parameter gradient changes, such as the blade leading edge and the suction transition zone, but also cannot fully capture the full-space distribution and transient evolution behavior of aerodynamic parameters under different guide vane angles. Furthermore, the installation of sensors is constrained by the compact turbine structure, and may even damage the integrity of the blade structure and the actual flow state. At the numerical simulation level, although high-fidelity CFD numerical simulation can achieve accurate solutions for the flow field in all dimensions, the mesh refinement and unsteady solution settings required to ensure computational accuracy will result in a huge consumption of computational resources and time costs, making it difficult to support multi-condition aerodynamic optimization design and rapid evaluation of fluid-structure interaction performance.

[0004] The aforementioned technical bottlenecks have become the core obstacles restricting the optimization of variable geometry turbine aerodynamic performance, the precise design of guide vane adjustment strategies, and the improvement of blade structure reliability. There is an urgent need to develop a rapid flow field prediction method that balances accuracy and efficiency. Summary of the Invention

[0005] This invention provides a method for rapid prediction of transient flow fields in turbomachinery based on d-POD and Bi-LSTM, in order to solve the core dilemma of difficulty in balancing accuracy and cost in existing technologies.

[0006] This invention is achieved through the following technical solution: A fast prediction method for transient flow fields in turbomachinery based on d-POD and Bi-LSTM, the prediction method comprising the following steps: Step 1: Adjust the boundary conditions for different operating conditions of the turbomachinery, complete the transient flow field numerical simulation, and construct the full-cycle flow field snapshot matrix; Step 2: Using MATLAB data acquisition code, extract the physical quantities of the region of interest in the flow field from the multiple sets of simulation data generated in Step 1. Based on Takens embedding theorem, perform time delay embedding on the snapshot matrix, construct the Hankel augmented matrix, and complete the normalization process. Step 3: Perform d-POD decomposition on the original temporal augmented matrix constructed in Step 2 based on the SVD singular value decomposition algorithm to obtain the spatial mode matrix, time coefficient matrix and modal energy matrix; Step 4: Train the Bi-LSTM model; Step 5: Use the trained Bi-LSTM model to calculate the historical time series coefficients of the working condition to be predicted, and the model outputs the predicted time series coefficients for future times. Step 6: Match and bind the reconstructed physical quantity data with the node coordinates and topology information of the original flow field mesh one by one, supplement the restored physical quantities with complete mesh information, generate .dat files according to the CFD post-processing standard format, and complete the visualization output of the predicted flow field and subsequent engineering applications.

[0007] Furthermore, step 1 specifically involves setting different combinations of boundary conditions for the target turbomachinery, including rotor speed, guide vane installation angle, inlet total temperature, inlet total pressure and outlet back pressure, to complete the full-cycle unsteady CFD numerical simulation. With a fixed time step Δ t Acquire transient flow field data on the surface of the first-stage turbine blades. The number of discrete nodes in the flow field space is: n The total number of consecutive sampling time steps is M It can construct the original transient flow field snapshot matrix. Y :

[0008] In this matrix, each row corresponds to the physical quantity time series data of the same grid node at different time steps, and each column corresponds to the spatial flow field physical quantity distribution of all grid nodes at the same time step.

[0009] Furthermore, step 2 specifically includes: Based on Takens' embedding theorem, a delay number is introduced. d Temporal delay embedding is performed on the original snapshot matrix to construct a Hankel-type delay augmented matrix. H :

[0010] In the formula: d Let n be the number of slips, and let the matrix be n.H Each line is a sequence of original snapshots. Y The delayed copy, of which the first i row correspondence Y Shift to the right i The results after -1 time steps, each column corresponds to a reconstructed phase space state, containing the current time and its preceding time. d A single snapshot of the flow field at consecutive time points directly embeds the temporal evolution information of the flow field into the snapshot matrix; delay times. d The value is determined through preliminary experiments based on the degree of nonlinearity of the flow field, and the range is 3~10.

[0011] Furthermore, the standardization of the Hankle matrix aims to address the numerical instability that may result from the presence of outliers in the data. The first step is to calculate the augmented matrix. H average matrix The formula is as follows:

[0012] in, N For augmented matrix H The total number of columns; Represents an augmented matrix H The j Column vector, To and H A matrix of the same dimension, where each column is the average of all column vectors; The second step is to use an augmented matrix. H Subtract the average matrix from each column The normalized matrix used for d-POD decomposition is obtained. The formula is as follows: .

[0013] Furthermore, step 3 specifically involves: Consistent with the underlying logic of the basic POD method, the essence of the d-POD method is still to find a set of optimal orthogonal bases. This makes the standardized matrix The projection onto it is optimal in the least squares sense, by means of Singular value decomposition is used to efficiently solve the d-POD basis, resulting in three matrices. U , S , V ,Right now: .

[0014] Furthermore, step 4 specifically includes the following steps: Step 4.1: Based on the criterion that the modal energy ratio is >99%, complete the dominant mode truncation and retain the first mode. k The time coefficient sequence corresponding to the dominant mode; Step 4.2: Normalize the truncated time coefficient sequence, construct time-series supervised learning samples, divide the training set and test set, and input Bi... The LSTM bidirectional long short-term memory network was used to complete the model training.

[0015] Furthermore, step 5 specifically involves: employing a rolling prediction method, using only historical data and the model's own previous prediction results as input, without introducing any future real data, and gradually outputting future predictions. S The normalized prediction time coefficients for each time step are obtained; then, through inverse normalization, the prediction time coefficient matrix with true dimensions is restored. .

[0016] Furthermore, step 6 specifically includes the following steps: Step 6.1: Compare the predicted time coefficient series with the retained previous... k The dominant spatial modes are linearly superimposed, and the average flow field matrix is ​​superimposed to reconstruct the complete transient flow field physical quantity distribution at future time. Step 6.2: Match and bind the grid node coordinates and topology information to the obtained physical quantities of the restored flow field, and generate a .dat file that can be directly read by CFD post-processing software.

[0017] Furthermore, step 6.1 specifically involves: reconstructing the transient flow field based on the truncated dominant mode and the predicted time coefficient. For t The flow field at any given time is first determined by the predicted time coefficient. With truncated basis modes The reconstructed normalized delay matrix yields the corresponding column vector:

[0018] Then, the physical quantity values ​​of the delay matrix are recovered through inverse centralization, and finally extracted from the reconstructed delay column vector. t Snapshot of the flow field at a given moment This completes the reconstruction of the full flow field distribution at that moment:

[0019] In the formula: This indicates taking the first part of the reconstructed delayed column vector. n Each element corresponds to t Spatial distribution of the flow field at any given time.

[0020] A rapid prediction system for transient flow fields in turbomachinery based on d-POD and Bi-LSTM is disclosed. The prediction system utilizes the aforementioned rapid prediction method for transient flow fields in turbomachinery based on d-POD and Bi-LSTM. The prediction system includes: Matrix construction and standardization module: Adjust the boundary conditions of turbomachinery under different operating conditions, complete the transient flow field numerical simulation, and construct the full-cycle flow field snapshot matrix; Based on MATLAB, the physical quantities of the region of interest in the flow field are extracted from multiple sets of simulation data through data acquisition code. The snapshot matrix is ​​embedded with time delay based on Takens embedding theorem to construct the Hankel augmented matrix and complete the normalization process. d-POD decomposition module: Based on the SVD singular value decomposition algorithm, the constructed time-augmented original matrix is ​​decomposed into d-POD to obtain the spatial mode matrix, time coefficient matrix and modal energy matrix; Training module: Train the Bi-LSTM model; Predicted Time Coefficient Sequence Module: Uses a trained Bi-LSTM model to output the historical time series time coefficients of the working conditions to be predicted, and the Bi-LSTM model outputs the predicted time coefficient sequence for future times. Matching and Binding and Standard Format Generation Module: Matches and binds the reconstructed physical quantity data with the node coordinates and topology information of the original flow field mesh one by one, supplements the restored physical quantities with complete mesh information, and generates .dat files according to the CFD post-processing standard format to complete the visualization output of the predicted flow field and subsequent engineering applications.

[0021] The beneficial effects of this invention are: This invention enables rapid prediction of blade aerodynamic loads under transient flow fields inside turbines. The method improves upon traditional POD (Programmable Flow Analysis) by utilizing Takens' embedding theorem, effectively extracting the spatiotemporal coupling characteristics of the transient flow field. Combined with the temporal prediction advantages of Bi-LSTM neural networks, it significantly enhances the stability and accuracy of long-term flow field prediction, exhibiting superior detail representation capabilities, especially in complex flow scenarios dominated by nonlinearity. This provides a powerful data-driven tool for real-time flow field prediction, flow field control, and optimization design in computational fluid dynamics.

[0022] This invention addresses the fundamental problems of insufficient temporal feature capture and easy accumulation of prediction errors in traditional POD methods by employing the d-POD delayed orthogonal decomposition method, significantly improving the accuracy and stability of long-term prediction of transient flow fields. This invention introduces the Takens embedding theorem to improve the traditional POD method. By constructing a Hankel augmented matrix through time-delay embedding, the historical temporal information of the flow field is integrated into the snapshot matrix. Singular value decomposition is then performed, and the resulting modes are no longer simply spatial features but also contain the spatiotemporal coupling characteristics of the flow field, comprehensively characterizing the temporal dependencies and dynamic properties of the flow field. The prediction model built upon this foundation inherently contains the temporal evolution law of the flow field in its low-dimensional modal coefficients, fundamentally reducing the risk of error accumulation in multi-step predictions.

[0023] This invention employs a Bi-LSTM bidirectional network to construct a time-series prediction model, significantly improving the ability to characterize the details of strongly nonlinear flows, resulting in superior model generalization and robustness. Its bidirectional gating mechanism can simultaneously utilize forward and backward information from the flow field time-series data, fully capturing the periodic characteristics of unsteady turbomachinery flows and the bidirectional coupling correlation between upstream and downstream flow fields. Compared to unidirectional networks, it can more accurately capture the detailed features of strongly nonlinear flows such as flow separation, vortex shedding, and wake interference, and is more adaptable to complex flow scenarios. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the Bi-LSTM neural network structure of the present invention.

[0025] Figure 2 This is a schematic diagram of the prediction and reconstruction algorithm framework combining d-POD and Bi-LSTM of the present invention. Detailed Implementation

[0026] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0027] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0028] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0029] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0030] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0031] Traditional mode decomposition (POD) models suffer from inherent limitations in capturing temporal features, resulting in poor long-term prediction accuracy and error accumulation in transient flow fields. Traditional POD methods decompose only a snapshot of the spatial flow field at a single moment, focusing primarily on extracting spatial structural features without incorporating temporal evolution information. The resulting spatial modes cannot characterize the temporal dependencies of the flow field, only reconstructing known flow fields and failing to support future flow field predictions. Even when combined with a unidirectional LSTM network to build a prediction model, the lack of spatiotemporal coupling in the underlying POD modes leads to rapid error accumulation over multiple prediction steps, significantly reducing long-term prediction accuracy and making it unsuitable for the full-cycle prediction requirements of unsteady flow fields in turbomachinery. Existing transient flow field order reduction models mostly use one-way LSTM networks for time series prediction. One-way LSTM can only utilize the time series information of historical moments and cannot capture the bidirectional dependence of the flow field time series data. However, the unsteady flow inside turbomachinery has significant periodic characteristics. The evolution of the upstream wake and secondary flow has a strong bidirectional correlation with the downstream flow field state. One-way networks cannot fully capture this periodic coupling feature, resulting in insufficient ability to characterize the details of strongly nonlinear flows such as flow separation and vortex shedding.

[0032] Breaking through the inherent contradiction between accuracy and efficiency in traditional CFD simulation, this method achieves a three-order-of-magnitude improvement in computational speed while maintaining flow field prediction accuracy comparable to high-fidelity CFD simulation, thus meeting the engineering requirements for multi-condition optimization and real-time simulation.

[0033] To address the issue of insufficient temporal feature capture in traditional POD reduced-order models, this paper improves upon traditional POD by using Takens' embedding theorem and constructs a d-POD delayed orthogonal decomposition method. This method integrates the temporal evolution information of the flow field into the modal decomposition process, extracts the spatiotemporal coupling features of the flow field, and suppresses the error accumulation of transient multi-step prediction.

[0034] To enhance the model's ability to characterize details of strongly nonlinear flows and improve its long-term prediction stability, a time-series prediction model is constructed using a Bi-LSTM bidirectional long short-term memory network. This model fully captures the bidirectional dependencies and periodic evolution characteristics of flow field time-series data, further reducing prediction errors and improving the model's generalization ability compared to the traditional unidirectional LSTM model.

[0035] A fast prediction method for transient flow fields in turbomachinery based on d-POD and Bi-LSTM, the prediction method comprising the following steps: Step 1: Adjust the boundary conditions for different operating conditions of the turbomachinery, complete the transient flow field numerical simulation, and construct the full-cycle flow field snapshot matrix; Step 2: Extract the physical quantities of the flow field region of interest from the simulation results in Step 1 using the data acquisition code, perform time delay embedding on the snapshot matrix based on the Takens embedding theorem, construct the Hankel augmented matrix and complete the normalization process; Step 3: Apply the Singular Value Decomposition (SVD) algorithm to the normalized augmented matrix from Step 2 and perform d... POD decomposition yields the spatial mode matrix, time coefficient matrix, and modal energy matrix; Step 4: Based on the criterion that the modal energy ratio is >99%, complete the dominant mode truncation and retain the time coefficient sequence corresponding to the first k dominant modes; The modal energies are accumulated sequentially from highest to lowest percentage. k The energy percentage of each mode is calculated until the total energy percentage exceeds 99%, at which point the final mode cutoff order is determined. k This enables a low-dimensional representation of high-dimensional transient flow fields. The corresponding modal time coefficient matrix... A It can be calculated using the following formula:

[0036] Step 5: Normalize the truncated time coefficient sequence, construct time-series supervised learning samples, divide the training set and test set, and input Bi... The LSTM bidirectional long short-term memory network was used to complete the model training. For the truncated time coefficient matrix A Performing min-max normalization maps all values ​​to the [0,1] interval, eliminating the dimensional differences between coefficients of different orders and improving the stability of neural network training.

[0037] Temporal supervised learning samples are constructed based on the normalized time coefficient sequence, and the original... M The size of the snapshot samples was reduced to [number] after data reorganization. M -10-15, the variation in sample size is determined by the following two key aspects: Firstly, in order to construct a phase space trajectory that meets the requirements of delayed embedding theory, a maximum delay number is set. d =10. Each reconstructed state needs to include snapshot data of the current time step and its 10 preceding consecutive time steps to form a complete delay coordinate vector. Therefore, the first 10 time steps in the original sequence are excluded during the reconstruction process because they cannot provide sufficient historical delay information, reducing the initial effective sample size to 10. M -10; On the other hand, the Bi-LSTM network uses 15 consecutive time steps as the input window to predict the state at the next time step. This requires further truncating a continuous subsequence of length 15 from the sample sequence after the above delayed embedding as the model input.

[0038] Following a 7:3 time-series division rule, the samples from the first 70% of the time span are divided into the training set, and the samples from the last 30% of the time span are divided into the test set. The training set is used for updating model weights, and the test set is used to verify the model's generalization ability. This division method strictly follows temporal causality and avoids the temporal travel problem of "training with future data and verifying with past data" caused by random division, ensuring that the model learns the temporal evolution characteristics of the flow field that conform to physical laws.

[0039] A Bi-LSTM bidirectional long short-term memory network model was constructed. The network structure consists of an input layer, two Bi-LSTM hidden layers, and a fully connected output layer connected in series. The Bi-LSTM includes parallel forward LSTM layers and backward LSTM layers, and the processing range of the two layers is strictly limited to the input. L Within the historical time series window, the entire process avoids accessing any real-world data to be predicted for the future, as detailed below: Forward LSTM layer: In the input L Within the history window, sort by time in ascending order (from step 1 to step 2). L (Step 1) Process the input sequence to uncover positive temporal dependencies in historical data; Backward LSTM layer: in the same L Step through the history window, in reverse chronological order (starting from the window's...). L Step 1 in the window: Process the input sequence and mine the reverse correlation features in historical data; The final hidden state at each time step is obtained by concatenating the forward and backward hidden states. Without introducing future real data, it fully captures the bidirectional dependency relationship of time series data within the historical window, adapting to the periodic coupling characteristics of unsteady flow in turbomachinery.

[0040] Step 6: Train the Bi from Step 5 The LSTM model takes historical time series coefficients of the working condition to be predicted as input and outputs a sequence of predicted time coefficients for future times. Step 7: Linearly superimpose the predicted time coefficient sequence with the first k dominant spatial modes retained in Step 4, superimpose the average flow field matrix, and restore the complete transient flow field physical quantity distribution at future time. Step 8: Match and bind the mesh node coordinates and topology information to the physical quantities of the restored flow field obtained in Step 7, and generate a .dat file that can be directly read by CFD post-processing software.

[0041] Extract the historical grid node coordinates and topology information from the original CFD simulation, bind the reconstructed predicted flow field physical quantities to the grid nodes one by one, and generate a directly readable .dat file according to the common format of CFD post-processing software.

[0042] Takens' embedding theorem (also known as the time-delay embedding theorem) was proposed by Dutch mathematician Floris Takens in 1981. This theorem provides a solid mathematical foundation for reconstructing the complete state of an original dynamical system based on a time series of a single observation variable. Its core idea is that by combining the time series with a time-delayed copy of itself, a low-dimensional observation signal can be embedded into a higher-dimensional state space, thereby recovering the system's intrinsic dynamic characteristics.

[0043] Given a dynamical system and an observation function, we can define a delay coordinate mapping that maps the true state of the system at a certain moment to the delay vector that we can observe: Let the actual state of the dynamic system be (Usually unknown), we can only observe a scalar function. Therefore, the time delay embedding is defined as: (1) in: It is the selected time delay, a positive time interval; m It is the embedding dimension, that is, the dimension of the constructed vector space.

[0044] Takens' embedding theorem is rigorously proven that, under mild assumptions (the system is smooth, the observation function is smooth and general), as long as... ( d (This is the dimension of the primordial attractor), this mapping This allows the attractor differential homeomorphism of the system to be embedded into In space, the "differential homeomorphism" is the mathematical core of the theorem. It means: ① Topological equivalence—the trajectory in the reconstructed space has the same topological properties as the attractor of the original system, such as connectivity and number of holes; ② Dynamical equivalence: the evolution of the trajectory is preserved. The invariants of the system remain unchanged in the reconstructed space.

[0045] For the flow field The original snapshot matrix Y, i.e., equation (2): (2) The Hankle matrix H obtained after the slip is as follows (3): (3) In the formula: For vector signals, m The number of vectors, d The number of slides.

[0046] For a given one-dimensional time series Y, if the dimension is 1× m Its trajectory matrix H can be constructed through time-delay embedding. Each row of H is a delayed copy of Y, where the th row is the tagged matrix of Y. i The line is a rightward shift of Y. The sequence after n time steps. Ultimately, H is a sequence with dimension n. The matrix.

[0047] The standardization of the Hankle matrix is ​​shown in Equation (4), which aims to solve the problem of numerical instability that may be caused by the presence of singular values ​​in the data.

[0048] (4) In the formula: It is an average matrix. r The number of samples.

[0049] Consistent with the underlying logic of the basic POD method, the delayed proper orthogonal decomposition (d-POD) method is essentially still about finding a set of optimal orthogonal bases. , making The projection onto it is optimal in the least squares sense, which can be achieved by adjusting the matrix. Singular value decomposition is used to achieve efficient solution of the POD basis, i.e. (5) In the formula: U is the left singular vector matrix, S is the singular value matrix, and ΦU is the fundamental mode. Then, the time coefficient... A n It can be obtained through the following equation (6): (6) in: a It is a column vector. N The number of vectors. Based on the previous... k The reconstruction method of the dominant mode can capture the key dynamics of the flow field with high accuracy and improve computational efficiency by significantly reducing the data dimensionality. However, it should be noted that the above equations can only reconstruct the known original flow and cannot predict the future flow conditions. We need a Bi-LSTM network as a framework to complete the prediction process.

[0050] The core idea of ​​Bi-LSTM is to construct two independent LSTM layers in opposite directions in parallel: one is a forward LSTM layer that processes the sequence in forward time (t=1 to t=T) to capture the influence of historical moments on the current moment; the other is a backward LSTM layer that processes the sequence in reverse time (t=T to t=1) to capture the influence of future moments on the current moment.

[0051] The hidden states of the two LSTM layers are independent of each other. At each time step, the forward and backward hidden state are concatenated (or weighted summed) to obtain the final output of the Bi-LSTM at that time, achieving complete extraction of bidirectional temporal features of the sequence. The overall network structure of Bi-LSTM is as follows: Figure 1 As shown, for a time series of length T, the output of each time step contains both forward and backward time series information.

[0052] The modal coefficient matrix obtained from d-POD decomposition is used as the supervised learning target to construct a coupled prediction framework of "d-POD feature extraction + Bi-LSTM time series prediction". Specifically, the modal coefficients of the historical time step are used as input and the modal coefficients of the future time step are used as output. The Adam gradient descent algorithm is used to train the Bi-LSTM bidirectional long short-term memory network to establish a nonlinear mapping relationship from the historical flow field state to the modal coefficients of the future time step. The bidirectional gating mechanism is used to fully capture the bidirectional dependence and periodic evolution characteristics of the flow field time series data.

[0053] A rapid prediction system for transient flow fields in turbomachinery based on d-POD and Bi-LSTM is disclosed. The prediction system utilizes the aforementioned rapid prediction method for transient flow fields in turbomachinery based on d-POD and Bi-LSTM. The prediction system includes: Matrix construction and standardization module: Adjust the boundary conditions of turbomachinery under different operating conditions, complete the transient flow field numerical simulation, and construct the full-cycle flow field snapshot matrix; Based on MATLAB, the physical quantities of the region of interest in the flow field are extracted from multiple sets of simulation data through data acquisition code. The snapshot matrix is ​​embedded with time delay based on Takens embedding theorem to construct the Hankel augmented matrix and complete the normalization process. d-POD decomposition module: Based on the SVD singular value decomposition algorithm, the constructed time-augmented original matrix is ​​decomposed into d-POD to obtain the spatial mode matrix, time coefficient matrix and modal energy matrix; Training module: Train the Bi-LSTM model; Predicted Time Coefficient Sequence Module: Uses a trained Bi-LSTM model to output the historical time series time coefficients of the working conditions to be predicted, and the Bi-LSTM model outputs the predicted time coefficient sequence for future times. Matching and Binding and Standard Format Generation Module: Matches and binds the reconstructed physical quantity data with the node coordinates and topology information of the original flow field mesh one by one, supplements the restored physical quantities with complete mesh information, and generates .dat files according to the CFD post-processing standard format to complete the visualization output of the predicted flow field and subsequent engineering applications.

[0054] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for rapid prediction of transient flow fields in turbomachinery based on d-POD and Bi-LSTM, characterized in that, The prediction method includes the following steps: Step 1: Adjust the boundary conditions for different operating conditions of the turbomachinery, complete the transient flow field numerical simulation, and construct the full-cycle flow field snapshot matrix; Step 2: Using MATLAB data acquisition code, extract the physical quantities of the region of interest in the flow field from the multiple sets of simulation data generated in Step 1. Based on Takens embedding theorem, perform time delay embedding on the snapshot matrix, construct the Hankel augmented matrix, and complete the normalization process. Step 3: Perform d-POD decomposition on the original temporal augmented matrix constructed in Step 2 based on the SVD singular value decomposition algorithm to obtain the spatial mode matrix, time coefficient matrix and modal energy matrix; Step 4: Train the Bi-LSTM model; Step 5: Use the trained Bi-LSTM model to calculate the historical time series coefficients of the working condition to be predicted, and the model outputs the predicted time series coefficients for future times. Step 6: Match and bind the reconstructed physical quantity data with the node coordinates and topology information of the original flow field mesh one by one, supplement the restored physical quantities with complete mesh information, generate .dat files according to the CFD post-processing standard format, and complete the visualization output of the predicted flow field and subsequent engineering applications.

2. The method for rapid prediction of transient flow field in turbomachinery according to claim 1, characterized in that, Step 1 specifically involves setting different combinations of boundary conditions for the target turbomachinery, including rotor speed, guide vane installation angle, inlet total temperature, inlet total pressure and outlet back pressure, to complete the full-cycle unsteady CFD numerical simulation. With a fixed time step Δ t Acquire transient flow field data on the surface of the first-stage turbine blades. The number of discrete nodes in the flow field space is: n The total number of consecutive sampling time steps is M It can construct the original transient flow field snapshot matrix. Y : In this matrix, each row corresponds to the physical quantity time series data of the same grid node at different time steps, and each column corresponds to the spatial flow field physical quantity distribution of all grid nodes at the same time step.

3. The method for rapid prediction of transient flow field in turbomachinery according to claim 1, characterized in that, Step 2 specifically involves: Based on Takens' embedding theorem, a delay number is introduced. d Temporal delay embedding is performed on the original snapshot matrix to construct a Hankel-type delay augmented matrix. H : In the formula: d Let n be the number of slips, and let the matrix be n. H Each line is a sequence of original snapshots. Y The delayed copy, of which the first i row correspondence Y Shift to the right i The results after -1 time steps, each column corresponds to a reconstructed phase space state, containing the current time and its preceding time. d A single snapshot of the flow field at consecutive time points directly embeds the temporal evolution information of the flow field into the snapshot matrix; delay times. d The degree of nonlinearity of the flow field is determined through preliminary experiments.

4. The method for rapid prediction of transient flow field in turbomachinery according to claim 3, characterized in that, Standardization of the Hankle matrix: The first step is to calculate the augmented matrix. H average matrix The formula is as follows: in, N For augmented matrix H The total number of columns; Represents an augmented matrix H The j Column vector, To and H A matrix of the same dimension, where each column is the average of all column vectors; The second step is to use an augmented matrix. H Subtract the average matrix from each column The normalized matrix used for d-POD decomposition is obtained. The formula is as follows: 。 5. The method for rapid prediction of transient flow field in turbomachinery according to claim 1, characterized in that, Step 3 specifically involves: Consistent with the underlying logic of the basic POD method, the essence of the d-POD method is still to find a set of optimal orthogonal bases. This makes the standardized matrix The projection onto it is optimal in the least squares sense, by means of Singular value decomposition is used to efficiently solve the d-POD basis, resulting in three matrices. U , S , V ,Right now: 。 6. The method for rapid prediction of transient flow field in turbomachinery according to claim 1, characterized in that, Step 4 specifically includes the following steps: Step 4.1: Based on the criterion that the modal energy ratio is >99%, complete the dominant mode truncation and retain the first mode. k The time coefficient sequence corresponding to the dominant mode; Step 4.2: Normalize the truncated time coefficient sequence, construct time-series supervised learning samples, divide the training set and test set, and input Bi... The LSTM bidirectional long short-term memory network was used to complete the model training.

7. The method for rapid prediction of transient flow field in turbomachinery according to claim 1, characterized in that, Step 5 specifically involves: employing a rolling prediction method, using only historical data and the model's own previous prediction results as input, without introducing any future real data, and gradually outputting future predictions. S Normalized prediction time coefficients for each time step; Then, through inverse normalization, the prediction time coefficient matrix with true dimensions is obtained. .

8. The method for rapid prediction of transient flow field in turbomachinery according to claim 6, characterized in that, Step 6 specifically includes the following steps: Step 6.1: Compare the predicted time coefficient series with the retained previous... k The dominant spatial modes are linearly superimposed, and the average flow field matrix is ​​superimposed to reconstruct the complete transient flow field physical quantity distribution at future time. Step 6.2: Match and bind the grid node coordinates and topology information to the obtained physical quantities of the restored flow field, and generate a .dat file that can be directly read by CFD post-processing software.

9. The method for rapid prediction of transient flow field in turbomachinery according to claim 8, characterized in that, Step 6.1 specifically involves: reconstructing the transient flow field based on the truncated dominant mode and the predicted time coefficient. For t The flow field at any given time is first determined by the predicted time coefficient. With truncated basis modes The reconstructed normalized delay matrix yields the corresponding column vector: Then, the physical quantity values ​​of the delay matrix are recovered through inverse centralization, and finally extracted from the reconstructed delay column vector. t Snapshot of the flow field at a given moment This completes the reconstruction of the full flow field distribution at that moment: In the formula: This indicates taking the first part of the reconstructed delayed column vector. n Each element corresponds to t Spatial distribution of the flow field at any given time.

10. A rapid prediction system for transient flow fields in turbomachinery based on d-POD and Bi-LSTM, characterized in that, The prediction system uses the rapid prediction method for transient flow fields in turbomachinery based on d-POD and Bi-LSTM as described in any one of claims 1-9. The prediction system includes: Matrix construction and standardization module: Adjust the boundary conditions of turbomachinery under different operating conditions, complete the transient flow field numerical simulation, and construct the full-cycle flow field snapshot matrix; Based on MATLAB, the physical quantities of the region of interest in the flow field are extracted from multiple sets of simulation data through data acquisition code. The snapshot matrix is ​​embedded with time delay based on Takens embedding theorem to construct the Hankel augmented matrix and complete the normalization process. d-POD decomposition module: Based on the SVD singular value decomposition algorithm, the constructed time-augmented original matrix is ​​decomposed into d-POD to obtain the spatial mode matrix, time coefficient matrix and modal energy matrix; Training module: Train the Bi-LSTM model; Predicted Time Coefficient Sequence Module: Uses a trained Bi-LSTM model to output the historical time series time coefficients of the working conditions to be predicted, and the Bi-LSTM model outputs the predicted time coefficient sequence for future times. Matching and Binding and Standard Format Generation Module: Matches and binds the reconstructed physical quantity data with the node coordinates and topology information of the original flow field mesh one by one, supplements the restored physical quantities with complete mesh information, and generates .dat files according to the CFD post-processing standard format to complete the visualization output of the predicted flow field and subsequent engineering applications.

Images