Cascade performance prediction method and system based on multi-precision sequential deep neural network, storage medium and product

Abstract

The invention belongs to the field of cascade performance prediction, and particularly discloses a cascade performance prediction method and system based on a multi-precision sequential deep neural network, a storage medium and a product, and the method comprises the steps: constructing a low-precision and high-precision data set based on a cascade design variable and a corresponding aerodynamic performance index; pre-training the DNN model by taking the low-precision data set as a training set and the high-precision data set as a verification set; performing fine training on the pre-trained DNN model by taking a part of the high-precision data set as a training set and the other part of the high-precision data set as a verification set to obtain a cascade performance prediction model; in the training process, the verification loss and the loss variable quantity of each iteration are calculated through the verification set; if the loss variation is smaller than a preset value, the early stop index is updated, and otherwise, the early stop index is reset to zero; and when the early stop index reaches a preset value, stopping training and recovering the model to a network state when the verification loss is minimum. According to the invention, the cascade performance prediction precision can be improved by using the advantages of different precision data.

Description

Technical Field

[0001] This invention belongs to the field of leaf cascade performance prediction, and more specifically, relates to a method, system, storage medium and product for predicting leaf cascade performance based on a multi-precision sequential deep neural network. Background Technology

[0002] A compressor is a type of turbine machinery used to pressurize gaseous media. Its aerodynamic design process is typically divided into two stages: preliminary design and detailed design. In the preliminary design stage, accurately predicting blade performance is a crucial step. Therefore, developing accurate and efficient methods for predicting blade performance is of significant engineering value for the aerodynamic design and performance optimization of compressors.

[0003] While traditional high-precision computational fluid dynamics (CFD) simulation methods offer high prediction accuracy, their large computational load and long processing time make them unsuitable for rapid design iterations under various operating conditions and parameters. Therefore, surrogate models have been widely used in cascade performance prediction in recent years. Surrogate models, by learning from limited sample data, can effectively construct alternative models that approximate CFD simulation results. However, existing surrogate models are typically built solely based on high-precision samples, failing to fully utilize low-cost data from other sources, resulting in insufficient information utilization efficiency.

[0004] To reduce reliance on a large number of high-precision samples, researchers have conducted extensive and in-depth studies on multi-precision data fusion methods. The core of these methods is to effectively reduce computational costs by fusing sample data of different precisions. Compared to traditional machine learning methods, deep neural networks (DNNs) possess stronger nonlinear expressive power and adaptive modeling capabilities, gradually becoming a research hotspot in the field of multi-precision data fusion. Among them, multi-precision deep neural networks (MFDNNs), as a representative method, decompose the relationship between high-precision and low-precision data into linear and nonlinear correlations, and introduce a hyperparameter to adjust the contribution weights of the two parts, thereby achieving effective fusion of multi-precision data and reducing computational costs while maintaining prediction accuracy.

[0005] However, the above-mentioned multi-precision data fusion methods still have two limitations in application: First, the model performance is heavily dependent on the selection of hyperparameters, and changes in hyperparameters can easily lead to instability in prediction results, thereby increasing the difficulty of parameter tuning in practical engineering applications; Second, it requires that high-precision inputs must be matched one by one with the corresponding low-precision labels, but in practical engineering applications, low-precision data is usually much more abundant than high-precision data, which greatly limits the information utilization efficiency of low-precision data.

[0006] Therefore, there is an urgent need to develop a simpler and more effective method for multi-precision data fusion. This method can not only make full use of the advantages of data from samples with different precisions, but also avoid the tedious hyperparameter tuning process, thereby improving its applicability and efficiency in engineering practice. Summary of the Invention

[0007] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method, system, storage medium and product for predicting cascade performance based on multi-precision sequential deep neural networks. The purpose is to achieve more convenient and effective multi-precision data fusion and improve the accuracy of cascade performance prediction.

[0008] To achieve the above objectives, according to a first aspect of the present invention, a method for predicting leaf cascade performance based on a multi-precision sequential deep neural network is proposed, comprising the following steps: Model building phase: S1. A dataset is constructed based on the cascade design variables and corresponding aerodynamic performance indicators. This dataset is divided into a low-precision dataset and a high-precision dataset according to the data accuracy. S2. Pre-train the DNN model using a low-precision dataset as the training set and a high-precision dataset as the validation set. S3. Use a portion of the high-precision dataset as the training set and another portion of the high-precision dataset as the validation set to fine-train the pre-trained DNN model. The finely trained DNN model is the leaf-passage performance prediction model. During pre-training and fine-training, the validation loss for each iteration and the change in validation loss compared to the previous iteration are calculated using the validation set. If the current change in validation loss is less than a preset value, the early stopping metric is updated. , If the early stopping metric is the same as the one used in the previous iteration, then the early stopping metric is reset to zero. When the early stopping metric reaches the preset value, training is terminated and the model is restored to the network state where the validation loss is minimized. Model application phase: S4. Input the design variables of the blade cascade to be predicted into the blade cascade performance prediction model to obtain the aerodynamic performance index of the blade cascade.

[0009] As a further optimization, during pre-training and fine-training processes: If the current validation loss is greater than the minimum of all previous validation losses, then update the dynamic learning rate adjustment metric. , Adjust the dynamic learning rate metric based on the previous iteration; otherwise, reset the dynamic learning rate metric to zero. Furthermore, if If the preset value is reached, the learning rate will be reduced to 50% of the current learning rate, and the dynamic learning rate adjustment metric will be reset to zero.

[0010] As a further preferred embodiment, the low-precision dataset and the high-precision dataset are constructed in the following ways: Multiple sets of cascade design variables were generated using the Latin hypercube sampling method. Based on the cascade design variables, the corresponding aerodynamic performance indicators are obtained through semi-empirical formulas to construct a low-precision dataset. Based on the cascade design variables, corresponding aerodynamic performance indicators are obtained through numerical simulation, and a high-precision dataset is constructed.

[0011] As a further preferred embodiment, the blade design variables are five-dimensional variables including lift coefficient, maximum relative thickness, incoming Mach number, angle of attack, and consistency.

[0012] As a further preferred embodiment, the aerodynamic performance indicators include the total pressure loss coefficient of the blade cascade and the lag angle.

[0013] As a further preferred embodiment, the verification loss is calculated based on the verification set using the mean absolute error (MAE).

[0014] According to a second aspect of the present invention, a leaf cascade performance prediction system based on a multi-precision sequential deep neural network is provided, comprising a processor for executing the above-described leaf cascade performance prediction method based on a multi-precision sequential deep neural network.

[0015] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described leaf cascade performance prediction method based on a multi-precision sequential deep neural network.

[0016] According to a fourth aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the above-described leaf cascade performance prediction method based on a multi-precision sequential deep neural network.

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages: 1. This invention proposes a simpler and more effective method for multi-precision data fusion, namely, a phased training strategy including a pre-training stage and a fine-training stage: in the pre-training stage, global features are captured by learning from low-precision samples, which significantly improves the information utilization efficiency of low-precision data; in the fine-training stage, the weights and biases of the network are further fine-tuned to fully explore the detailed features in high-precision data, thereby making full use of the data advantages of samples with different precision and effectively improving the prediction accuracy of leaf cascade performance.

[0018] 2. Through an innovative phased training strategy, the pre-training phase can flexibly introduce any number of low-precision samples, thereby completely eliminating the limitation of traditional methods that require high-precision inputs to correspond one by one with low-precision labels, giving full play to the potential value of low-precision data, and greatly improving the utilization efficiency of low-precision data.

[0019] 3. No external hyperparameters need to be introduced during the model construction process, which fundamentally avoids the complex and tedious hyperparameter tuning steps in existing multi-precision data fusion methods, greatly reduces the implementation cost, and significantly enhances the convenience and practicality of this invention in engineering applications. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the leaf cascade performance prediction method based on a multi-precision sequential deep neural network according to an embodiment of the present invention. Figure 2 This is a flowchart of the leaf cascade performance prediction method based on a multi-precision sequential deep neural network according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the distribution of discrete factors under different discrete levels in an embodiment of the present invention; Figure 4 This is a graph showing the response curves of the MAE (Matrices Equation) of the SMFDNN model predicting the target value of the analytical function under different error levels in an embodiment of the present invention, as the number of low-precision samples changes. Figures 5 to 13 This is a graph showing the response curves of the MAE of the MFDNN model predicting the target value of the analytical function as a function of the hyperparameter α under different error levels in this embodiment of the invention. Figures 14 to 22 This is a comparison chart of the prediction accuracy of different models in this invention when predicting the target value of the analytical function under different error levels and different high-precision sample numbers. Figure 23 , Figure 24 When the MFDNN model is used to predict the total pressure loss coefficient and cascade lag angle of the NACA65 series blade cascade in an embodiment of the present invention, the MAE varies with the hyperparameters. Response curve of the change; Figure 25 This is a comparison chart of the prediction accuracy of each model when predicting the total pressure loss coefficient of the NACA65 series blade cascade under different high-precision sample numbers in the embodiments of the present invention. Figure 26 This is a comparison chart of the prediction accuracy of each model when predicting the lag angle of the NACA65 series blades under different high-precision sample numbers in embodiments of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] This invention provides a method for predicting cascade performance based on a multi-precision sequential deep neural network (SMFDNN), such as... Figure 1 and Figure 2 As shown, it includes the following steps: S1: Construct datasets with different levels of precision.

[0023] For the blade cascade, low-precision and high-precision datasets were constructed using design variables and corresponding aerodynamic performance indicators as samples. The design variables include the lift coefficient. Maximum relative thickness Mach number of incoming flow Angle of attack and consistency The aforementioned variables collectively constitute the five-dimensional input variables of the SMFDNN model, covering the main influencing factors of aerodynamic performance and blade cascade geometry. The predicted aerodynamic performance index is the total pressure loss coefficient of the blade cascade. and backward angle .

[0024] This document describes a specific scheme for generating samples of different precision using the Latin hypercube sampling method. When constructing the low-precision dataset, a semi-empirical formula is used to quickly generate a large number of low-precision samples. When constructing the high-precision dataset, commercial computational fluid dynamics software (such as Fluent) is used for numerical simulation to obtain the corresponding blade cascade performance data, including the total inlet pressure of the blade cascade. Total export pressure Inlet static pressure and airflow exit angle The formula for calculating the total pressure loss coefficient is as follows: The formula for calculating the lag angle is: .

[0025] S2: Pre-train the model and build a network model with the minimum validation loss.

[0026] During the pre-training phase, the deep neural network model uses a low-precision dataset as the training set and a high-precision dataset as the validation set. Mean absolute error (MAE) is used as the validation loss metric, and the initial validation loss value is recorded. During training, dynamic learning rate adjustment and early stopping mechanisms are introduced: the validation loss value is recorded for each iteration during model training. And calculate the change in verification loss. The dynamic learning rate is updated based on the following strategy to determine whether to terminate training: If the current verification loss Greater than the minimum of all previous validation losses, i.e. , n The dynamic learning rate adjustment metric is updated based on the number of iterations. Otherwise, the dynamic learning rate adjustment metric is set to zero. If the current verification loss changes If the value is less than the preset value, the early stop indicator will be updated. Otherwise, the indicator will be set to zero early. If the dynamic learning rate adjustment metric If the preset value is reached, the learning rate will be reduced to 50% of the current value and the dynamic learning rate adjustment metric will be reset. If the early stopping metric reaches the preset value, training is terminated and the model is restored to the network state with the minimum validation loss.

[0027] S3: Refine the model and establish a network model with the minimum validation loss.

[0028] In the fine-training phase, 80% of the high-precision dataset is randomly selected as the training set, and the remaining 20% ​​is used as the validation set. Based on the network model obtained in the pre-training phase, the network weights and biases are further fine-tuned to fully explore the local detailed features in the high-precision data and improve prediction performance. The training strategy in the fine-training phase is consistent with that in the pre-training phase.

[0029] S4: Predict new cascade performance information.

[0030] After the model training is completed, the aerodynamic performance information of the new cascade can be quickly output based on the constructed SMFDNN model by inputting the five-dimensional input variables to be predicted.

[0031] The following are specific examples: To verify the effectiveness and convenience of the method of the present invention, two examples were selected: the analytical function prediction task and the NACA65 series blade aerodynamic performance prediction task. The SMFDNN of the present invention was compared and analyzed with the traditional DNN and MFDNN methods.

[0032] This embodiment generates a sample dataset using the Latin hypercube sampling method within a selected variable range. Specifically, 10,000 test samples are generated based on a 5-dimensional input variable space to objectively evaluate the predictive performance of different models; simultaneously, high-precision samples of sizes ranging from 100 to 1,000 are generated to construct high-precision datasets of various sizes.

[0033] We select a five-dimensional analytic equation that simultaneously contains nonlinear terms, cross terms, and periodic functions as the object of study, and its expression is: Each input variable The range of values ​​is During the effectiveness verification process, the dependent variable of this analytical equation was used. f Identify high-precision labels.

[0034] To simulate the characteristics of low-precision data, the error term of the low-precision data is defined to follow a normal distribution, which is used to generate corresponding low-precision sample labels. This embodiment sets three discrete levels and three bias levels to comprehensively study the impact of error levels on the performance of different models. The discrete levels correspond to the standard deviations. The deviation levels correspond to the system deviations respectively. First, based on the aforementioned standard deviation parameter... Generate the corresponding random discrete factor Combined with system bias With high-precision labels The corresponding low-precision label can then be generated, and the formula is as follows: The distribution of discrete factors corresponding to different discrete levels is as follows: Figure 3 As shown.

[0035] Figure 4 This paper demonstrates the impact of the number of low-precision samples on the prediction performance of the proposed SMFDNN model. The number of high-precision samples is fixed at 100, while low-precision datasets of different sizes are constructed, and the model's prediction accuracy is evaluated on the same test set. As can be observed from the figure, the overall prediction error of the model decreases with the increase of the number of low-precision samples, indicating that increasing the number of low-precision samples helps improve the model's prediction performance. Furthermore, different error levels show significant differences in their impact on the model's prediction accuracy: with the same number of low-precision samples, lower error levels (…) result in higher prediction accuracy. The prediction accuracy at the time of ) is significantly better than that at a higher error level ( This illustrates that the lower the error level of low-precision data, the more significant its contribution to improving model performance. It's worth noting that once the number of low-precision samples exceeds a certain threshold (e.g., 30,000), further increasing the size of low-precision samples no longer significantly improves model performance, exhibiting a diminishing marginal return. This phenomenon indicates that although the model's prediction error continuously decreases with increasing low-precision data size, excessively increasing the number of low-precision samples is not economical and may even lead to a waste of computational resources. Therefore, considering both the model performance improvement and computational resource consumption, this embodiment selects 100,000 samples as the low-precision sample size for the SMFDNN model pre-training stage in all examples.

[0036] Figures 5 to 13 Different discrete levels were displayed respectively. ), different deviation levels ( ) and different numbers of high-precision samples ( Under these conditions, the MAE of the MFDNN model on the test set varies with hyperparameters. The trend of change. As can be observed from the figure, the MFDNN model's hyperparameters... It exhibits strong sensitivity. The curves in each group generally show a typical "U"-shaped or U-shaped distribution, indicating that only when the hyperparameter... The model achieves better predictive performance only when the value of falls within a specific interval. Furthermore, the optimal hyperparameters corresponding to the minimum MAE obtained by the MFDNN model on the test set are... The summary is as follows: when hour, The corresponding optimal hyperparameters are 0.391, 0.409, and 0.566, respectively.

[0037] when hour, The corresponding optimal hyperparameters are 0.169, 0.714, and 0.039, respectively.

[0038] when hour, The corresponding optimal hyperparameters are 0.001, 0.454, and 0.582, respectively.

[0039] when hour, The corresponding optimal hyperparameters are 0.526, 0.529, and 0.357, respectively.

[0040] when hour, The corresponding optimal hyperparameters are 0.178, 0.436, and 0.397, respectively.

[0041] when hour, The corresponding optimal hyperparameters are 0.033, 0.483, and 0.541, respectively.

[0042] when hour, The corresponding optimal hyperparameters are 0.462, 0.526, and 0.17, respectively.

[0043] when hour, The corresponding optimal hyperparameters are 0.194, 0.39, and 0.41, respectively.

[0044] when hour, The corresponding optimal hyperparameters are 0.291, 0.198, and 0.475, respectively.

[0045] Figures 14 to 22 Further demonstration of different discrete levels ( ) and different deviation levels ( Under different high-precision sample numbers, DNN, MFDNN, SMFDNN, and pre-trained models... Comparison of prediction accuracy under different conditions. Among them, the hyperparameters used in the MFDNN model... Corresponding to the above text regarding The determined optimal hyperparameters.

[0046] As can be observed from the figure, the pre-trained SMFDNN model exhibits low sensitivity to discrete errors: when the discrete level increases from 0.35 to 0.75, the prediction accuracy of the pre-trained model hardly changes significantly. However, the increase in systematic bias significantly reduces the prediction performance of the pre-trained model. For example, when the systematic bias reaches 0.4 and 0.8, respectively, the mean absolute error of the pre-trained model on the test set rises sharply to 0.3 and 0.6, respectively. Based on this model, the network weights and biases are further optimized in the fine training stage, making the prediction performance of the SMFDNN model superior to the DNN and MFDNN models under all error levels and high-precision data scale conditions. It is worth noting that even when the pre-trained model produces large prediction errors, the SMFDNN model can still achieve satisfactory prediction accuracy, indicating that the fine training stage can effectively correct systematic biases in low-precision data. This further highlights the stability and robustness of the proposed method.

[0047] When the number of high-precision samples is small, the MFDNN model outperforms the DNN model in prediction. However, as the number of high-precision samples increases, its prediction accuracy weakens compared to the DNN model. This indicates that the MFDNN model is affected by hyperparameters. It exhibits high sensitivity, particularly susceptible to changes in data distribution and initial conditions. Under different sample sizes, the hyperparameters need to be re-optimized. .

[0048] The predictive performance of DNN models is highly dependent on the number of high-precision samples, exhibiting stability and good prediction accuracy under large sample conditions. However, results across different error levels show that the prediction error of DNN models is significantly larger when the number of samples is small. Although the prediction accuracy gradually improves with increasing sample size, the overall performance remains limited by the scale of high-precision samples.

[0049] In this embodiment, when constructing high-precision datasets for both the loss and lag angle models, the Latin hypercube sampling method is first used to generate 100 to 1000 high-precision samples of different sizes to construct high-precision datasets of various sizes. An additional 2000 samples are generated as a test set. After removing samples where numerical calculations did not converge, the actual number of usable high-precision samples are 92, 190, 285, 385, 477, 578, 662, 765, 875, 968, and 1909, respectively.

[0050] Figure 23 and Figure 24 This demonstrates the MAE of the MFDNN model on the test set as a function of hyperparameters under different high-precision sample sizes. The response curve shows the changes. As can be seen from the figure, regardless of whether the predicted object is the total pressure loss coefficient or the lag angle, the MFDNN model responds well to these changes. It exhibits significant sensitivity. The optimal hyperparameters corresponding to the minimum MAE achieved by MFDNN on the test set are summarized below: For loss models, The corresponding optimal hyperparameters are 0.179, 0.001, and 0.001.

[0051] For the lag angle model The corresponding optimal hyperparameters are 0.121, 0.062, and 0.016.

[0052] Figure 25 and Figure 26 The prediction error curves of various prediction models (including semi-empirical formulas, DNN models, MFDNN models with different optimal hyperparameters, pre-trained models, and SMFDNN models) on the test set are shown under different high-precision sample sizes.

[0053] First, it can be observed that, compared with the traditional semi-empirical formula, the pre-trained model achieves the initial fusion of multi-precision data information by using high-precision samples as the validation set and restoring to the minimum validation loss training strategy, which effectively improves the accuracy of the total pressure loss coefficient and the prediction of the lag angle.

[0054] Furthermore, the MAE curves of DNN, SMFDNN, and MFDNN models all show a decreasing trend with the increase of the number of high-precision samples, indicating that high-precision samples play a decisive role in the model's prediction accuracy. Among them, the SMFDNN model outperforms the DNN and MFDNN models in prediction accuracy at all high-precision sample sizes. When the number of high-precision samples is less than 200, the prediction error of the SMFDNN model is reduced by more than 40% compared to the DNN model, fully demonstrating its advantage under the condition of scarce high-precision samples.

[0055] MFDNN model with different hyperparameters The MAE curves under different conditions show significant differences, which further verifies that the MFDNN model is effective for hyperparameters. The strong dependence on high-precision samples indicates that the MFDNN model requires repeated hyperparameter tuning for different scales of high-precision samples, while the SMFDNN model does not require the introduction of additional external hyperparameters. Its parameter-free characteristic significantly reduces implementation costs in engineering applications. In summary, the SMFDNN model can fully utilize low-precision information and significantly reduce errors when high-precision samples are scarce, while slightly outperforming the DNN model when samples are abundant. Furthermore, it avoids the strong dependence of the MFDNN model on hyperparameters, exhibiting higher overall predictive performance and stability.

[0056] Based on the above analysis, the SMFDNN method proposed in this invention has the following significant advantages: (1) SMFDNN, through its innovative phased training strategy, can flexibly introduce any number of low-precision samples during the pre-training phase. This successfully avoids the strict limitation in traditional multi-precision data fusion methods that high-precision inputs need to correspond one-to-one with low-precision labels, fully leverages the global information value of low-precision data, and significantly improves the utilization efficiency of low-precision data.

[0057] (2) No external hyperparameters need to be introduced during the construction of the SMFDNN model, which fundamentally avoids the tedious and complex hyperparameter tuning process in the traditional multi-precision data fusion process, greatly reduces the implementation cost, and significantly improves the applicability and stability of the model in actual engineering. (3) The SMFDNN model achieves efficient multi-precision data information fusion. It demonstrates excellent prediction performance and robustness in all examples of this patent embodiment. It outperforms traditional DNN and MFDNN models under different low-precision error levels and high-precision sample sizes, highlighting its significant technical advantages in practical engineering applications.

[0058] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for cascade performance prediction of a turbine cascade based on a multi-precision sequential deep neural network, characterized in that, The application relates to a cascade performance prediction method based on a multi-precision successive deep neural network. The method comprises the following steps: S1. Constructing a data set based on cascade design variables and corresponding aerodynamic performance indexes, wherein the data set is divided into a low-precision data set and a high-precision data set according to data precision; S2. Pre-training a DNN model by taking the low-precision data set as a training set and the high-precision data set as a verification set; S3. Fine training the pre-trained DNN model by taking part of the high-precision data set as a training set and another part of the high-precision data set as a verification set, wherein the DNN model after fine training is a cascade performance prediction model; In the pre-training and fine-training processes, the validation loss of each iteration is calculated through a validation set, and the change amount of the validation loss compared with the last iteration; if the current change amount of the validation loss is less than a preset value, the early stopping index is updated , is the early stopping index of the last iteration, otherwise the early stopping index is reset to zero; when the early stopping index reaches a preset value, the training is terminated and the model is restored to the network state at the time when the validation loss is the smallest. S4. Inputting cascade design variables to be predicted into the cascade performance prediction model to obtain aerodynamic performance indexes of the cascade. During the pre-training and fine training processes:

2. The multi-precision sequential deep neural network based cascade performance prediction method of claim 1, wherein, The construction method of the low-precision data set and the high-precision data set comprises the following steps: If the current validation loss is greater than the minimum of all previous validation losses, then update the dynamic learning rate adjustment indicator , is the dynamic learning rate adjustment indicator for the previous iteration, otherwise reset the dynamic learning rate adjustment indicator to zero; Further, if If the preset value is reached, the learning rate is reduced to 50% of the current learning rate, and the dynamic learning rate adjustment indicator is reset to zero.

3. The multi-precision sequential deep neural network based cascade performance prediction method of claim 1, wherein, A Latin hypercube sampling method is adopted to generate multiple groups of cascade design variables; Corresponding aerodynamic performance indexes are obtained based on the cascade design variables through semi-empirical formulas to construct the low-precision data set; Corresponding aerodynamic performance indexes are obtained based on the cascade design variables through numerical simulation to construct the high-precision data set. The cascade design variables are five-dimensional variables including a lift coefficient, a maximum relative thickness, a Mach number of incoming flow, an attack angle and a solidity.

4. The multi-precision sequential deep neural network based cascade performance prediction method of claim 1, wherein, The aerodynamic performance indexes include a total pressure loss coefficient and a lag angle of the cascade.

5. The multi-precision sequential deep neural network based cascade performance prediction method of claim 1, wherein, The verification loss is calculated based on the verification set by adopting a mean absolute error (MAE).

6. The multi-precision sequential deep neural network based cascade performance prediction method of any one of claims 1-5, wherein, The application further discloses a processor for executing the cascade performance prediction method based on the multi-precision successive deep neural network.

7. A cascade performance prediction system based on multiple precision successive deep neural networks, characterized in that, The application further discloses a computer program for implementing the cascade performance prediction method based on the multi-precision successive deep neural network when executed by a processor.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The application further discloses a computer program product for implementing the cascade performance prediction method based on the multi-precision successive deep neural network when executed by a processor.

9. A computer program product, characterised in that, ​

Images