CN-121389758-B - Stable non-Gaussian wind pressure simulation method and system based on machine learning reinforced Hermite polynomial model

CN121389758BCN 121389758 BCN121389758 BCN 121389758BCN-121389758-B

Abstract

The invention relates to a steady non-Gaussian wind pressure simulation method and a system based on a machine learning enhanced Hermite polynomial model, which are characterized in that firstly, a multi-input single-output machine learning model is constructed to determine the required statistical moment in a conversion function of the proposed method, then a neural network conversion model from a multi-input multi-output non-Gaussian power spectrum to a potential Gaussian power spectrum is established on the basis, finally, an ML-EHPM model constructed on the basis of a double-machine learning model is respectively predicted to obtain the required statistical moment of the conversion function and the potential Gaussian power spectrum predicted by a Gaussian PSD prediction model, the potential Gaussian power spectrum representation method is adopted to simulate the non-Gaussian wind pressure, and under the conditions of strong non-Gaussian wind pressure, wind pressure exceeding a monotonic area and bimodal wind pressure distribution, the ML-EHPM method can more truly reproduce a target probability density function and a power spectrum density, and the simulation accuracy is remarkably superior to that of the traditional HPM method, and the advantages are particularly remarkable in the strong non-Gaussian samples.

Inventors

JIANG YAN
ZHAO NING
XIE PEILUN
WU FENGBO
XIN JINGZHOU
ZHANG HONG
TANG QIZHI
LIU CHENGLIANG

Assignees

重庆交通大学

Dates

Publication Date: 20260505
Application Date: 20251024

Claims (8)

1. The steady non-Gaussian wind pressure simulation method based on the machine learning reinforced Hermite polynomial model is characterized by comprising the following steps of: s1, constructing a BES-XGBoost model with multiple inputs and single outputs, and solving a ten-order origin moment of [% ] ) And the mean value of the transfer function As input feature, the rest conversion function statistical moment [ ] , , , ) Training as output characteristics to obtain a prediction model corresponding to four statistical moments; Assume that Non-Gaussian component process Original PDF of (c), then random procedure The first four statistical moments on the positive side are defined as: (1) (2) (3) (4) Wherein, the Is from the original PDF A median value derived; 、、 And The mean, standard deviation, skewness and kurtosis of the positive side are newly defined, respectively, and similarly, the statistical moment of the negative side is defined as: (5) (6) (7) (8) Wherein the method comprises the steps of , , And The mean, standard deviation, skewness and kurtosis of the negative side are newly defined respectively; The step S1 specifically comprises the following steps: s11, constructing a pairing training data set Estimating a first ten-order origin moment and a target probability density function based on measured data, and obtaining a conversion function mean value (a) through a formula (1) ) As input, then calculate the statistical moment of transfer function through formulas (2), (4), (6), (8) respectively , , , ) As output data, constructing a training data set, and providing basic data for model training; s12, training BES-XGBoost model Four statistical moments for transfer function , , , ) In the training process, optimizing key super parameters of XGBoost, namely iteration times, tree depth and learning rate by using a BES algorithm to improve prediction accuracy and model generalization capability; s13, establishing a nested input structure To further improve the stability and accuracy of the prediction, a nested input structure is introduced, namely, the BES-XGBoost model is utilized to predict the standard deviation of the positive and negative sides , ) The standard deviation obtained by prediction is used as a new input characteristic, and the same BES-XGBoost model is further used for predicting kurtosis , ); S14, predicting the statistical moment of the transfer function After training, the original point moment of the first ten steps of the target is calculated ) And the corresponding conversion function mean value ) Inputting into trained BES-XGBoost model, directly outputting statistical moment of transfer function , , , ); S2, constructing a multi-input multi-output BPNN model, and averaging a non-Gaussian PSD and a conversion function Sum transfer function statistical moment [ ] , , , ) As input characteristics, potential Gaussian PSD is used as output characteristics, and a prediction model of the potential Gaussian PSD is obtained after training; S3, based on the ML-EHPM model constructed in the step S1 and the step S2, respectively predicting the statistical moment required by the conversion function , , , ) And potential Gaussian power spectrum predicted by Gaussian PSD prediction model Using the potential Gaussian power spectrum The representation simulates non-gaussian wind pressure.
2. The method for simulating smooth non-Gaussian wind pressure according to claim 1, wherein the step S1 is specifically to simulate the target first ten-order origin moment of the point [ ] ) And the mean value of the transfer function Inputting the model into the nested BES-XGBoost model trained in the step S1, and predicting the statistical moment required by the conversion function , , , )。
3. The method for simulating smooth non-Gaussian wind pressure according to claim 2, wherein step S2 is specifically performed by normalizing the target non-Gaussian PSD of the simulated point using equation (16) , (16) Normalized non-Gaussian PSD power spectrum The conversion function statistical moment of the simulation point predicted by the BES-XGBoost model in the step S1 , , , ) And the mean value of the transfer function Inputting into a trained BPNN model to obtain a required potential Gaussian power spectrum 。
4. The method for simulating smooth non-Gaussian wind pressure according to claim 2, wherein the origin moment of step S1 is [ ] ) Mean value of conversion function To the statistical moment of the conversion function , , , ) The construction principle of the prediction model is that the original probability density function is divided into a positive side and a negative side by taking the median of the original probability density function as a limit, the original probability density function is called PDF for short, the positive side is larger than the median, the negative side is smaller than the median, two new symmetrical PDFs are respectively constructed so as to keep the same with the conversion function of the original PDFs on the corresponding sides, and the calculated statistical moment is not restricted by monotonicity because the deviation of the two symmetrical PDFs is always zero.
5. The method for simulating smooth non-Gaussian wind pressure according to claim 2, wherein the transformation function statistical moment is converted in the step S1 , , , ) According to different combinations of positive and negative side kurtosis, conversion function models of four different conditions are constructed, namely two softening processes, representing condition 1, half-softening half-hardening process, representing condition 2, half-hardening half-softening process, representing condition 3 and two hardening processes, representing condition 4, which are given by the following formulas: in case 1 of the case where the number of the cells is not equal, >3, >3: (9) In case 2 of the case where the number of the cells is not equal, >3, <3: (10) In case 3 of the case where the number of the cells is not equal, <3, >3: (11) In case 4 of the case where the number of the cells is not equal, <3, <3: (12) Wherein, the 、、 And Is determined using the newly defined statistical moment and using the following equation: (13) While 、、 And Is determined using the following equation: (14) (15) from the above equation, it can be seen that in each case the piecewise transfer function is in The points are all continuous, and therefore, Is a memory-free and monotonically increasing transfer function.
6. The method for simulating smooth non-Gaussian wind pressure of claim 5, wherein step S3 is performed by performing step S2 on the potential Gaussian power spectrum The matrix is subjected to Cholesky decomposition, and Gaussian time Cheng Yangben is generated by a spectrum representation method of a multivariate smooth random process of the formula (17) , , (17) Conversion function statistical moment of simulation point based on prediction in step S1 , , , ) Gaussian samples of the transformation function model under different conditions are obtained using equations (9) - (12) Conversion to non-Gaussian samples Will not be a Gaussian sample Inverse normalization to obtain final non-Gaussian sample 。
7. A steady non-gaussian wind pressure simulation system based on a machine learning enhanced Hermite polynomial model, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steady non-gaussian wind pressure simulation method according to any of the preceding claims 1 to 6 when executing the program.
8. A storage medium having stored thereon a computer program, which when executed by a processor, implements the steps of the stationary non-gaussian wind pressure simulation method according to any of claims 1 to 6.

Description

Stable non-Gaussian wind pressure simulation method and system based on machine learning reinforced Hermite polynomial model Technical Field The invention belongs to the technical field of non-Gaussian wind pressure simulation, and relates to a steady non-Gaussian wind pressure simulation method and a steady non-Gaussian wind pressure simulation system based on a machine learning reinforced Hermite polynomial model. Background The wind pressure of the low building roof is mainly guided by the ridge separation shear layer and the vortex in the roof corner region, the instantaneous negative pressure peak value is short in duration and high in amplitude, and the on-site actual measurement and the wind tunnel tests show that the probability distribution of the wind pressure coefficient of the roof is obviously negative biased, thick tail and even double peak, and the probability distribution exceeds the assumption of the traditional Gaussian random process. The existing non-Gaussian wind pressure simulation mostly adopts a conversion process method, namely a potential Gaussian process is firstly assumed, and then a Gaussian signal is mapped to target non-Gaussian edge distribution through monotone Hermite polynomial transformation (HPM). However, when the measured wind pressure kurtosis is greater than 3 or a double peak appears, the monotone transformation area is limited, so that the conversion function needs to be introduced with high-order moment estimation, the error is rapidly amplified, and meanwhile, the potential Gaussian power spectrum (UGPSD) cannot be obtained in an analytic way, and only can be repeatedly and iteratively corrected, so that the calculation is time-consuming and easy to distort. Therefore, a new method is needed that is independent of the monotonic assumption, can automatically match any kurtosis/skewness combination, and can directly and rapidly invert the gaussian power spectrum (UGPSD) from the non-gaussian power spectrum (TNPSD) so as to improve the simulation precision and efficiency of the strong non-gaussian roof wind pressure field. Disclosure of Invention In view of the above, the invention provides a steady non-Gaussian wind pressure simulation method and a system based on a machine learning enhanced Hermite polynomial model, which are used for fast and accurate simulation of steady non-Gaussian wind pressure, aiming at solving the problems that the prior Hermite polynomial transformation has large high-order moment error, slow inversion of potential Gaussian power spectrum and time-consuming and easy distortion caused by monotonic constraint under strong non-Gaussian and double-peak wind pressure. Based on the traditional HPM simulation method based on moment, two machine learning models are established to improve the simulation precision and efficiency. In order to achieve the above purpose, the present invention provides the following technical solutions: A steady non-Gaussian wind pressure simulation method based on a machine learning reinforced Hermite polynomial model comprises the following steps: s1, constructing a BES-XGBoost model with multiple inputs and single outputs, and solving a ten-order origin moment of [% ] ) And the mean value of the transfer functionAs input feature, the rest conversion function statistical moment [ ],,,) Training as output characteristics to obtain a prediction model corresponding to four statistical moments; s2, constructing a multi-input multi-output BPNN model, and averaging a non-Gaussian PSD and a conversion function Sum transfer function statistical moment [ ],,,) As input characteristics, potential Gaussian PSD is used as output characteristics, and a prediction model of the potential Gaussian PSD is obtained after training; S3, based on the ML-EHPM model constructed in the step S1 and the step S2, respectively predicting the statistical moment required by the conversion function ,,,) And potential Gaussian power spectrum predicted by Gaussian PSD prediction modelUsing the potential Gaussian power spectrumThe representation simulates non-gaussian wind pressure. Further, the step S1 is specifically that the original point moment of the target front ten steps of the simulation point is calculated) And the mean value of the transfer functionInputting the model into the nested BES-XGBoost model trained in the step S1, and predicting the statistical moment required by the conversion function,,,)。 Further, step S2 is specifically performed by normalizing the target non-Gaussian PSD of the simulation point using equation (16), (16) Normalized non-Gaussian PSD power spectrumThe conversion function statistical moment of the simulation point predicted by the BES-XGBoost model in the step S1,,,) And the mean value of the transfer functionInputting into a trained BPNN model to obtain a required potential Gaussian power spectrum。 Further, the step S3 is specifically to the step S2 potential Gaussian power spectrumThe matrix is subjected to C