Search

CN-122020911-A - Water pump turbine runner approximate model precision improving method based on targeted sample supplementing

CN122020911ACN 122020911 ACN122020911 ACN 122020911ACN-122020911-A

Abstract

The invention discloses a method for improving the precision of a water pump turbine runner approximate model based on a target sample supplementing, which belongs to the technical field of water turbine optimization design and comprises the steps of selecting optimization variables and optimization targets of the water pump turbine runner, and adopting a method based on the target sample supplementing The method comprises the steps of generating an initial sample point set by a criterion improved Latin hypercube sampling method, obtaining CFD calculation data corresponding to the initial sample point set to form an optimized sample set, constructing an initial approximation model by adopting a neural network model, evaluating the precision of the initial approximation model by adopting a comprehensive error parameter, adding sample points in a targeted mode at a position with larger approximation model error by adopting a Markov chain Monte Carlo method until the set approximation model precision requirement is met. The invention avoids the blindness problem of the traditional sample space sample filling, has the characteristics of strong sample utilization sequence, high precision lifting efficiency, controllable calculation cost and the like, and is beneficial to the development of the optimization of the water pump turbine runner design towards modern intellectualization and high efficiency.

Inventors

  • ZHANG WENWU
  • LIANG AO

Assignees

  • 中国农业大学

Dates

Publication Date
20260512
Application Date
20260324

Claims (8)

  1. 1. The method for improving the accuracy of the water pump turbine runner approximate model based on the target sample supplementing is characterized by comprising the following steps of: s1, selecting optimization variables and optimization targets of a runner of a water pump turbine, and adopting a sample space global uniformity as a constraint condition The method comprises the steps that a Latin hypercube sampling method is improved by criteria, and an initial sample point set is generated in a multidimensional design space; S2, acquiring CFD calculation data corresponding to an initial sample point set, forming an optimized sample set, and constructing an initial approximate model by adopting a neural network model; S3, estimating the precision of an initial approximate model by adopting a comprehensive error parameter, if the comprehensive error is lower than a preset threshold value, executing optimization algorithm optimization based on the approximate model, otherwise, entering S4; S4, constructing a target probability distribution function by using a Markov chain Monte Carlo method according to the error of the approximate model, and generating newly added sample points in a probability guiding mode in a high error area of the approximate model in a targeting manner; and S5, merging the newly added sample points into an optimized sample set, and returning to S2 for iteratively updating the approximate model until the comprehensive error meets the accuracy requirement of the approximate model.
  2. 2. The method for improving the accuracy of the water pump turbine runner approximation model based on the target sample compensation according to claim 1 is characterized in that in S1, variables which are strongly related to runner performance are selected as runner optimization variables, wherein the variables comprise a blade inlet setting angle beta 1 , a blade outlet setting angle beta 2 , a high-pressure side blade inclination angle alpha and a blade wrap angle theta, and pump mode design operating point efficiency eta 1 and turbine mode design operating point efficiency eta 2 of the water pump turbine are selected as optimization targets.
  3. 3. The method for improving the accuracy of the water pump turbine runner approximation model based on the target sample compensation according to claim 2, wherein in S1, the specific logic on which the improved latin hypercube sampling method generates the initial sample point set is based is as follows: determining the size of the initial sample point set based on random Latin hypercube, generating a new matrix by exchanging the order of the two factor levels in the matrix column, and calculating the total spacing of the new points Value, by successive iterations, when Stopping when the value change meets the set condition; ; Where d (X i ,X j ) is the Euclidean distance between sample points X i and X j , p is a positive integer, and N is the initial number of samples.
  4. 4. The method for improving the accuracy of the water pump turbine runner approximation model based on the target sample compensation according to claim 3, wherein in S2, specific logic for constructing an initial approximation model by the neural network model is as follows: s21, carrying out normalization processing on the data of the optimized sample set, dividing the data into a training set and a verification set, wherein a normalization formula is as follows: Wherein, the Respectively representing the normalized blade inlet setting angle, the blade outlet setting angle, the high-pressure side blade inclination angle, the blade wrap angle, the pump mode design operating point efficiency of the water pump turbine and the turbine mode design operating point efficiency; Respectively representing a minimum value of a blade inlet setting angle, a minimum value of a blade outlet setting angle, a minimum value of a blade inclination angle of a high-pressure side, a minimum value of a blade wrap angle, a minimum value of pump mode design working condition point efficiency of a water pump turbine and a minimum value of pump mode design working condition point efficiency of the water turbine in an actual optimized sample set; Respectively represents the maximum value of the blade inlet setting angle, the maximum value of the blade outlet setting angle, the maximum value of the blade inclination angle of the high-pressure side in the actual optimized sample set maximum blade wrap angle, maximum pump mode design operating point efficiency of the water pump turbine and maximum turbine mode design operating point efficiency; S22, dividing the neural network model into an input layer, a hidden layer and an output layer, and obtaining a sample point X i = [ As input to the neural network model input layer, the optimization objective matrix Y i = [ The theoretical output of the neural network model output layer, wherein i is the sample scheme number; Respectively representing blade inlet setting angle, blade outlet setting angle, high-pressure side blade inclination angle, blade wrap angle, pump mode design working condition point efficiency and water turbine mode design working condition point efficiency of ith sample point, selecting clustering algorithm to produce m sample centers as hidden layer function central node, its number range is greater than Lambda is a sample center coefficient, and the range is more than 1 and less than or equal to 2; S23, each hidden layer node h corresponds to a data center X ch , and outputs The distance between the input matrix and the data center is calculated and obtained through Gaussian function transformation: ; wherein X i is the i-th sample point, controlling the width of the basis function Σ h is the average value of the distances between every two data centers; S24, linear mapping is adopted from the hidden layer to the output layer, and the calculation formula is as follows: ; Wherein y ki is the theoretical output of the kth unit of the output layer corresponding to the ith sample scheme, w kh is the connection weight between the kth unit and the h node of the hidden layer, and b k is the bias of the kth unit of the output layer; S25, constructing a hidden layer output matrix phi and a hidden layer-to-output layer parameter matrix W, and defining an error function L which is the square of Euclidean distance between a predicted value matrix phi W and an optimization target matrix Y, wherein the error function L is as follows: ; ; ; In the formula, gamma is a regularization coefficient for balancing fitting precision and model complexity, an optimal value of the regularization coefficient is determined by a cross validation method, and the Euclidean distance adopts an L2 norm A representation; 、 、 W is a parameter matrix from the hidden layer to the output layer, wherein b 1 、b 2 represents the bias of the 1 st unit and the 2 nd unit of the output layer, W 1m represents the weight of the 1 st unit and the m th node of the hidden layer, and W 2m represents the weight of the 2 nd unit and the m th node of the hidden layer; S26, adopting a least square method, and solving a parameter matrix W from a hidden layer to an output layer of the neural network model in a regularization formula: ; ; ; in the formula, I is an identity matrix, and gamma is a regularization coefficient; Is a regularized covariance matrix.
  5. 5. The method for improving accuracy of a water pump turbine runner approximation model based on targeted sample compensation according to claim 4, wherein in S3, the integrated error parameter E consists of a global error E g considering the influence of the number of samples and a local error parameter E l considering the spatial density of samples, and the integrated error parameter is applicable to each optimization objective: ; wherein ζ is global error weight in the integrated error parameters; global error E g , taking into account the influence of the number of samples: ; Wherein E g is a global error parameter, R 2 is a decision coefficient, N T is the number of samples of a test set, f is the dimension of an optimization variable, T i is the true value of the samples of the ith test set, and T i is the response value of the approximation model of the ith sample; Averaging the true values for the test set; The specific calculation logic of the local errors is to arrange the relative errors of all sample points in descending order from large to small, select the first 10% of sample points and calculate by considering the space density of the samples: ; ; Wherein E l is a local error parameter, d i is the distance between the ith sample point and the nearest neighbor of the ith sample point, d max is the maximum spatial distance between the sample points, N TK is 10% of the total number of samples, and w di is the sample spatial density weight.
  6. 6. The method for improving the accuracy of the water pump turbine runner approximation model based on the target sample compensation according to claim 5, wherein in the step S3, the threshold value range of the approximation model error is judged to be 0.08 delta less than or equal to 0.05, and the threshold value of the approximation model error is gradually increased along with the increase of the optimization repetition times of the approximation model.
  7. 7. The method for improving the accuracy of the water pump turbine runner approximation model based on the targeted sample compensation according to claim 6, wherein in S4, specific logic on which the sample points are targeted and added by using a markov chain monte carlo method is as follows: s41, constructing a single sample point local error function The core is that the local density is combined with the average error of adjacent samples: ; In the formula, For new sample points With a new sample point Mean distance representation from s adjacent known samples x i , i.e ; Representing new sample points Average error of s adjacent known sample points, i.e ; S42, constructing a target probability density function p (x): ; Wherein omega is a feasible domain, namely a space formed by optimizing a variable value range, and epsilon is an extremely small positive value; S43, selecting a point with obvious local error from the existing sample points as an initial point x 0 , and avoiding the initial point from deviating from a target area; s44, generating new sample points by adopting Gaussian proposal distribution: ; wherein: representing the generation of new sample points from the current point x 0 Probability of (2); A gaussian distribution with x 0 as a mean value and sigma as a covariance matrix is represented, wherein the element distribution is 1/10 square of each variable value range, and the off-diagonal element is 0; For new sample points Calculating the acceptance probability: ; Randomly extracting a number U from the uniform distribution U (0, 1), and accepting the shift x z+1 = if U is less than or equal to ψ Otherwise, refusing to move x z+1 =x 0 , wherein z is an iteration number; S45, determining the iteration frequency range and the combustion period, after the iteration is finished, removing the combustion period sample from the MCMC chain, selecting a plurality of points with higher error estimation in the residual chain as sample points for final targeting supplement, performing CFD calculation on the new sample points, obtaining a response value, and adding the response value into an optimized sample set for retraining the approximate model.
  8. 8. An executable computer program, wherein the computer program is stored in a computer readable storage medium and when executed by a processor implements the method for improving accuracy of a water pump turbine runner approximation model based on a targeted sample supplement according to any one of claims 1 to 7.

Description

Water pump turbine runner approximate model precision improving method based on targeted sample supplementing Technical Field The invention relates to the technical field of water turbine optimization design, in particular to a method for improving the accuracy of a water pump water turbine runner approximate model based on a target sample supplementing. Background The water pump turbine is used as core power equipment of the pumped storage power station, and plays key roles of peak regulation and valley filling, frequency modulation and phase modulation and emergency standby of a power grid, and the operation efficiency and stability of the water pump turbine directly determine the energy conversion efficiency of the power station. The runner is used as a core component of the water pump turbine, and the geometric structure parameter of the runner is closely related to the flow state of water flow in the runner. Therefore, the optimal design of the rotating wheel is a key link for improving the overall performance of the water pump turbine. However, existing methods of constructing approximation models have significant drawbacks. Firstly, the traditional Latin hypercube sampling can ensure layering uniformity of single variable dimension, but sample uneven distribution is easy to occur in a variable space of multidimensional multi-parameter optimization, especially sample blank or dense sample can exist in a key variable region sensitive to the rotating wheel efficiency, so that fitting errors of an initial approximate model in the regions are obviously increased, secondly, when the model precision is insufficient, a commonly adopted uniform encryption or random sample supplementing strategy lacks pertinence, precious computing resources are wasted in a region where the model is fitted well, and the correction is insufficient in a key region with larger error, so that the model precision is slowly improved, and the optimization efficiency is low. Disclosure of Invention The invention aims to provide a method for improving the accuracy of a water pump turbine runner approximate model based on target sample supplementing, which adopts a method based on the following stepsThe improved Latin hypercube sampling method of the criterion generates an initial sample set, ensures the even distribution of samples in a multidimensional design space from the source, lays a foundation for constructing a high-precision initial approximate model, constructs comprehensive error parameters to evaluate the model precision by a system, creatively introduces a Markov chain Monte Carlo method, intelligently identifies a region with larger error of the approximate model, encrypts the target samples, and avoids the blindness problem of sample supplementing in the traditional sample space. In order to achieve the purpose, the invention provides a method for improving the accuracy of a water pump turbine runner approximation model based on a target sample supplementing, which comprises the following steps: s1, selecting optimization variables and optimization targets of a runner of a water pump turbine, and adopting a sample space global uniformity as a constraint condition The method comprises the steps that a Latin hypercube sampling method is improved by criteria, and an initial sample point set is generated in a multidimensional design space; S2, acquiring CFD calculation data corresponding to an initial sample point set, forming an optimized sample set, and constructing an initial approximate model by adopting a neural network model; S3, estimating the precision of an initial approximate model by adopting a comprehensive error parameter, if the comprehensive error is lower than a preset threshold value, executing optimization algorithm optimization based on the approximate model, otherwise, entering S4; S4, constructing a target probability distribution function by using a Markov chain Monte Carlo method according to the error of the approximate model, and generating newly added sample points in a probability guiding mode in a high error area of the approximate model in a targeting manner; and S5, merging the newly added sample points into an optimized sample set, and returning to S2 for iteratively updating the approximate model until the comprehensive error meets the accuracy requirement of the approximate model. Preferably, in S1, variables with strong correlation with the performance of the runner are selected as runner optimization variables, wherein the variables comprise a blade inlet setting angle beta 1, a blade outlet setting angle beta 2, a high-pressure side blade inclination angle alpha and a blade wrap angle theta, and pump mode design operating point efficiency eta 1 and turbine mode design operating point efficiency eta 2 of the water pump turbine are selected as optimization targets. Preferably, in S1, the specific logic on which the improved latin hypercube sampling method generates the initial sample point s