Method, system and related device for improving and optimizing blade angle distribution of supercritical carbon dioxide compressor

Abstract

The invention provides a supercritical carbon dioxide compressor blade angle distribution improvement and optimization method and system and a related device. The method comprises the following steps: step 1, constructing a sample space of a specified dimension; step 2, selecting uniformly distributed Latin hypercube sampling results from the sample space by adopting a translation propagation algorithm; step 3, training the constructed feedforward back propagation neural network model according to the selected Latin hypercube sampling result to obtain a trained feedforward back propagation neural network model; 4, solving the trained feedforward back propagation neural network model by adopting a fast non-dominated sorting genetic algorithm to obtain optimal blade angle distribution; according to the method, the optimal blade angle distribution design in a given pre-optimization parameter space can be obtained only by depending on a small amount of prior under the condition of not depending on a large number of compressor modeling design sample experience libraries.

Description

Technical Field

[0001] The present invention belongs to the field of supercritical carbon dioxide compressors, and in particular relates to a method, system and related devices for improving and optimizing the blade angle distribution of a supercritical carbon dioxide compressor. Background Art

[0002] The design of supercritical carbon dioxide compressors is an important issue in the current research on supercritical carbon dioxide cycle components. The blade angle distribution of the compressor blades has a significant impact on the isentropic efficiency and total pressure ratio of the compressor. Therefore, accurately determining the optimal blade angle distribution is of great significance to the design of supercritical carbon dioxide compressors.

[0003] Based on existing research on compressor performance parameter optimization, the weight coefficients used in current multi-objective compressor design optimization methods still rely on the designer's experience. For complex single-stage SCO2 compressors with splitter blades and a vaneless-vaned-vaneless diffuser composite, traditional methods can achieve a certain degree of compressor efficiency. However, these methods involve many artificially selected key hyperparameters, including weightings between objective parameters. This results in over-reliance on the designer's prior knowledge, potentially leading to omissions in the search target space.

[0004] Specifically, as the only components that perform work on the fluid, the three-dimensional geometry of centrifugal compressor impeller blades has a significant impact on the aerodynamic performance of the centrifugal compressor. The impeller blade angle distribution plays a key role in the three-dimensional design of centrifugal compressor blades, as it is related to other modeling parameters such as the blade wrap angle distribution and blade pitch angle distribution. Therefore, the blade angle distribution must be carefully selected to ensure the three-dimensional geometry of the centrifugal compressor blades is excellent and achieves efficient compression.

[0005] When the hub blade angle distribution is determined, the rim blade angle distribution has a significant impact on the centrifugal compressor tip clearance leakage flow. A Bezier curve determined by six equidistant control points is used to give the compressor rim side blade angle distribution curve. Among them, the starting control point T1 and the final control point T6 are both determined by the results of the low-dimensional aerodynamic and thermodynamic design of the centrifugal compressor. The vertical coordinates of the remaining four control points form four optimization variables, forming a four-factor multi-objective optimization problem:

[0006] According to the experience of fluid machinery optimization design, the multi-objective optimization problem pointed out by the above formula has complex characteristics such as nonlinearity, multi-peaks, and huge computational cost. Considering the economic cost and time cost of numerical calculation, in order to most accurately solve the Pareto frontier of the above multi-objective optimization problem and give the blade angle distribution in high-dimensional parameter space ( b - MThe optimal solution Yopt B of the curve control point coordinate group will be obtained by using a global stochastic optimization algorithm based on the meta-model. The main process of this method is as follows Figures 1 to 3 shown.

[0007] Although this method can comprehensively cover the geometric parameters of the compressor, it lacks guidance on the distribution and control of specific parameters, which leads to poor portability of its optimization results and, in turn, difficulties in large-scale design and application in a wider range of scenarios. Summary of the Invention

[0008] The purpose of the present invention is to provide a method, system and related devices for improving and optimizing the blade angle distribution of a supercritical carbon dioxide compressor, so as to solve the defect that the existing method for optimizing the blade angle distribution of supercritical carbon dioxide compressor blades has inaccurate optimization results.

[0009] In order to achieve the above object, the technical solution adopted in the present invention is: In a first aspect, the present invention provides a method for improving and optimizing the blade angular distribution of a supercritical carbon dioxide compressor, comprising the following steps: Step 1: construct a sample space of specified dimension; Step 2: Use the translation propagation algorithm to select uniformly distributed Latin hypercube sampling results from the sample space; Step 3: training the constructed feedforward back propagation neural network model according to the selected Latin hypercube sampling result to obtain a trained feedforward back propagation neural network model; Step 4: Use the fast non-dominated sorting genetic algorithm to solve the trained feedforward back propagation neural network model to obtain the optimal blade angle distribution.

[0010] Preferably, a sample space of a specified dimension is constructed by: Set the desired number of sample points in the sample space of a specified dimension, where each sample point includes four dimensions; Set the value range for each dimension.

[0011] Preferably, a translation propagation algorithm is used to select uniformly distributed Latin hypercube sampling results from the sample space, and the specific method is: Set the initial seed module in the sample space; Determine the number of partitions of the sample space based on the set initial seed module; Reshape the initial seed module using the determined number of partitions to obtain a reshaped seed module; The reshaped seed module is translated and updated according to the number of dimensions and the filling vector until all partitions in the sample space are filled to obtain the result matrix; Perform Latin hypercube sampling on the result matrix to obtain a Latin hypercube sampling result.

[0012] Preferably, the number of partitions of the sample space is determined according to the set initial seed module, specifically by: The number of sample points in the set sample space is divided by the seed value of the seed module to obtain the sample point ratio; Take the square root of the obtained sample point ratio and get the area value of the seed module size of the corresponding dimension; Round up the obtained area value to get the rounded partition number; The number of rounded partitions is compared with the regional value of the seed module size in each dimension. If the number of rounded partitions is greater than the regional value of the seed module size in each dimension, the number of partitions is the power of the number of sample dimensions. Otherwise, the number of partitions is the ratio of sample points, and the sample space size is the number of partitions multiplied by the number of sample points in the seed module.

[0013] Preferably, in step 3, the expression of the constructed feedforward back propagation neural network model is:

[0014] Where y represents the neuron input / output, f represents the activation function, w represents the weight, b represents the bias, k represents the current hidden layer of the neural network, i represents the i-th hidden layer neuron, and j represents the j-th neuron in the previous layer; The loss function is:

[0015] in, The last output of the hidden layer neuron Predicted attribute values, Indicates the first attribute values; represents the loss function; Q represents the size of the training set data.

[0016] Preferably, a fast non-dominated sorting genetic algorithm is used to solve the trained feedforward back propagation neural network model to obtain the optimal blade angle distribution. The specific method is: S1, initialize the genetic algorithm to obtain the initial random population; S2, using the trained neural network to predict the attribute value of each individual in the initial random population to obtain the attribute value of the initial population; S3, merging the initial random population with the initial population attribute value to obtain an updated random population; S4, using the population non-dominated sorting algorithm to calculate the dominance level and crowding distance of each individual in the random population, output the population arranged according to the dominance level according to the number of individuals in the population, and obtain the new parent population; S5, using the selection algorithm to select individuals that meet the set requirements from the new parent population to obtain the parent population; S6, performing crossover and mutation processing on the individuals in the selected parent population to obtain the offspring population; S7, using the obtained offspring population as the input of the trained feedforward back-propagation neural network model to predict the attribute value of each individual in the offspring population and obtain a new offspring population; S8, merge the new offspring population and the parent population in S5 into a fusion population; S9, using the population non-dominated sorting algorithm to calculate the dominance level and crowding distance of each individual in the fused population, output the populations sorted according to the dominance level according to the number of individuals in the population, and obtain the latest parent population; S10, repeat S2 to S9 with the latest parent population until the number of iterations is reached; S11, inverting the attribute value of the final population obtained to obtain the optimal value of the compressor blade angle distribution.

[0017] In a second aspect, the present invention provides a supercritical carbon dioxide compressor blade angle distribution improvement and optimization system, comprising: Sample space construction unit, used to construct a sample space of specified dimensions; A sampling unit, used to select uniformly distributed Latin hypercube sampling results from the sample space using a translation propagation algorithm; A model training unit, configured to train a constructed feedforward back propagation neural network model according to the selected Latin hypercube sampling result, and obtain a trained feedforward back propagation neural network model; The model solving unit is used to solve the trained feedforward back propagation neural network model using a fast non-dominated sorting genetic algorithm to obtain the optimal blade angle distribution.

[0018] In a third aspect, the present invention provides an electronic device comprising a processor and a memory, wherein the memory stores computer instructions, and when the computer instructions are executed by the processor, the electronic device executes the described method.

[0019] In a fourth aspect, the present invention provides a computer program product, which includes computer-executable instructions, and the computer-executable instructions implement the method when executed.

[0020] In a fifth aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions implement the described method when executed by a processor.

[0021] Compared with the prior art, the present invention has the following beneficial effects: To address the technical problem that existing supercritical CO2 compressor blade design methods rely on designer experience to select blade angles, resulting in blade angle distributions that fail to achieve optimal impeller pressure ratio and efficiency, hindering widespread application in various design scenarios, the present invention provides an improved optimization method for supercritical CO2 compressor blade angle distribution. This method uses a translational propagation algorithm to rapidly select design sample points to achieve uniform sample selection. A feedforward back-propagation neural network is rapidly trained using known sample points to achieve full coverage of the sample space. A fast non-dominated sorting genetic algorithm is then used to search the target space for non-dominated solutions to determine the optimal blade angle distribution. This method, with only prior values of pre-optimized parameters and upper and lower limits of the parameter variation range based on the prior values, can determine the optimal pre-optimized parameter-optimization target result. This method can determine the optimal blade angle distribution design within a given pre-optimized parameter space, relying on only a small amount of prior information, without relying on a large library of compressor model design samples. Consequently, this method is applicable to the rapid determination of impeller blade angle distributions in the design of supercritical CO2 compressors with varying external characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a schematic diagram of the wheel rim blade angle curve; Figure 2 It is the variable range setting diagram of the blade angle curve control point; Figure 3 is a schematic diagram of the process of Example 1 of the present invention; Figure 4 is a schematic diagram of a flow chart of Example 2 of the present invention; Figure 5 This is a comparison chart of the test set efficiency prediction results; Figure 6 This is a comparison chart of the test set pressure ratio prediction results; Figure 7 This is a scatter plot of 300 individuals whose dominance level is all 1; Figure 8 Schematic diagram of determining the Pareto front using the non-dominated sorting algorithm. DETAILED DESCRIPTION

[0023] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0024] Example 1 This embodiment provides a method for improving and optimizing the blade angle distribution of a supercritical carbon dioxide compressor, characterized by comprising the following steps: Step 1: construct a sample space of specified dimension; Step 2: Use the translation propagation algorithm to select uniformly distributed Latin hypercube sampling results from the sample space; Step 3: training the constructed feedforward back propagation neural network model according to the selected Latin hypercube sampling result to obtain a trained feedforward back propagation neural network model; Step 4: Use the fast non-dominated sorting genetic algorithm to solve the trained feedforward back propagation neural network model to obtain the optimal blade angle distribution.

[0025] To address the technical problem that existing supercritical CO2 compressor blade design methods rely on designer experience to select blade angles, resulting in blade angle distributions that fail to achieve optimal impeller pressure ratio and efficiency, hindering widespread application in design scenarios, this application provides an improved optimization method for supercritical CO2 compressor blade angle distribution. This method uses a translational propagation algorithm to rapidly select design sample points to achieve uniform sample selection. It then uses known sample points to rapidly train a feedforward back-propagation neural network to achieve full coverage of the sample space. A fast non-dominated sorting genetic algorithm then searches the target space for non-dominated solutions to determine the optimal blade angle distribution. This method, with only prior values of pre-optimized parameters and upper and lower limits on the parameter variation range based on the prior values, can determine the optimal pre-optimized parameter-optimization target result. This method can determine the optimal blade angle distribution within a given pre-optimized parameter space, relying on only a small number of priors, without relying on a large library of compressor model design samples. Consequently, this method is applicable to the rapid determination of impeller blade angle distributions in the design of supercritical CO2 compressors with varying external characteristics.

[0026] Example 2 This embodiment provides a method for improving and optimizing the blade angle distribution of supercritical carbon dioxide compressor blades, comprising the following steps: Step 1: construct a sample space of specified dimension; like Figure 1As shown, the expected number of sample points in the sample space of the specified dimension is set (np = 100), each sample point includes four dimensions (nv = 4), and the value range of each dimension in each sample point is:

[0027] Step 2: Use the translation propagation algorithm to select uniformly distributed Latin hypercube sample points from the specified dimensional sample space. Specifically: S201, set the seed module seed in the sample space to:

[0028] S202, determining the number of partitions of the sample space according to the set initial seed module, the specific method is: The area value of the seed module size of the first dimension corresponding to the sample space is calculated according to the set number of seeds of the seed module (ns = 3). The area value of the seed module size of the first dimension is calculated according to the following formula: nd=(np / ns) (1 / nv) =(100 / 3) (1 / 4) =2.4028; Round up the calculated region value of the seed module size of the first dimension to get the rounded partition number (ndstar = 3), and determine that ndstar>nd. Compare the rounded partition number (ndstar) with the seed module size area value (nd) of each dimension. If ndstar>nd, it is considered necessary to create a larger number of partitions (nb). The number of partitions (nb) is the power of the number of sample dimensions (nv) of the rounded partition number (ndstar), that is, nb = ndstar nv = 3 4 = 81; otherwise, ndstar = nd, eliminating the need to create a larger number of partitions. For a given seed module (seed), the number of sample points (ns) within it is known; in this example, ns = 3. Therefore, shifting the seed module (seed) across nb partitions yields a sampling result. The final number of sample points (npstar) is: npstar = ns·nb = 3·81 = 243.

[0029] S203: Reshape the initial seed module using the determined number of partitions to obtain a reshaped seed module. The seed module is a matrix consisting of ns nv-dimensional row vectors, with the values of the elements in each row vector required to be in the interval [0, 1]. The upper and lower limits of the seed module are calculated based on the number of partitions (nb) and the maximum number of sample points (npstar). The upper and lower limits of the seed module are then looped through the dimensions (nv). The seed module is updated by multiplying the upper limit by the row vector of each seed module plus the lower limit to obtain the reshaped seed module. Specifically: The upper and lower limits of the value range of each dimension in each sample point are calculated according to the following formula:

[0030]

[0031] Among them, upper represents the upper limit, lower represents the lower limit, np represents the sample space size, nd represents the number of rounded partitions, nv represents the number of dimensions, and ns represents the number of sample points in the seed module.

[0032] In this embodiment, the calculated upper limit value is [36, 36, 36, 36], and the lower limit value is [-35, -35, -35, -35].

[0033] Multiply each seed value by the upper limit and add the lower limit, then round it off to get the reshaped seed module as follows:

[0034] S204 , performing translation update on the reshaped seed module according to the number of dimensions (nv) and the filling vector until all partitions in the sample space are filled, thereby obtaining a result matrix.

[0035] When performing translation updates according to the number of dimensions and the fill vector, loop through the dimensions (nv) one by one, changing one dimension at a time. Specifically: Initialize the fill vector d = ones(1,nv); (ones represents a row vector whose elements are all 1, and the vector length is the number of dimensions nv) c1 = 1; (update starts from the first dimension, c1 takes positive integer values from 1 to nv, and the sampling result is obtained when c1 = nv iteration ends) update seed = seed0; (update seed module, get update seed) d(1:(c1-1)) = ndstar (c1-2) ; (Assign the first (c1-1) elements in the d vector to ndstar (c1-2)In this example, c1 = 1 means assigning the first 0 elements, and there is no change before and after the assignment) d(c1) = npstar / ndstar; (Assign the c1th element in the d vector to npstar / ndstar. For this example, when c1 = 1, d(1) = 243 / 3 = 81); d((c1+1):end) = ndstar (c1-1) ; (Assign all elements from c1+1th to the last element of the d vector to ndstar (c1-1) , for this example, d(2:4) = 3 0 = 1); Therefore, for the first dimension, we get the filling vector d = [81,1,1,1]; Add the padding vector d to the update seed in the first dimension to get seed1:

[0036] Add the filling vector d to seed1 in sequence to get seed2:

[0037] Merge the two new seed modules with the reshaped seed module, and the result matrix A after the first cycle is as follows:

[0038] Update the resulting matrix A as a new seed module, i.e. update seed = A (a matrix of size 9·4); at this time c1 = 2; Update the filling vector d according to the above method, and the result is d = [1,81,3,3]; Add the filling vector d to the update seed in the second dimension to get seed1:

[0039] Add the filling vector d to seed1 in sequence to get seed2:

[0040] Merge the two new seed modules with the reshaped seed module to obtain the result matrix A after the second cycle , the size of the matrix is (27·4) Repeat the above steps until c1 = nv, and the iteration ends.

[0041] Generally, according to the above method, the size of the result matrix A after the c1th iteration is (ns·ndstar c1 ) · nv, in this example, when c1 = nv = 4, the resulting matrix A is of size (3·3 4 ) ·4 means a matrix of size 243·4.

[0042] Repeat the above cycle until c1 = nv S205: Since npstar > np, sample points need to be removed to ensure the target number of sample points is met. First, the center of the sample space is calculated as [121.5, 121.5, 121.5, 121.5]. For each sample in the result matrix A, the Euclidean distance from the center of the space is calculated to obtain a distance vector d (243x1). Each distance value is sorted and assigned an index value, where the smallest distance value is assigned the smallest index value and the largest distance value is assigned the largest index value. The row vectors corresponding to the first np index values in the result matrix A are extracted to form a new matrix X.

[0043] S205. For each dimension (column vector), the minimum value is calculated. First, the row vectors in X are sorted from smallest to largest along the first dimension (first column). For each column, the minimum value of that column is subtracted and 1 is added to ensure that all values in X are non-negative. The first column is compared with the column vectors from 1 to np. If an element is equal, it is assigned 0; if it is not equal, it is assigned 1. This forms a logical vector, and a determination is made whether the logical vector is non-zero. If it is non-zero, the loop is exited and the operation for the next dimension is performed. Otherwise, the logical vector is subtracted from the first column until the set number of iterations is reached. Then, the row vectors in X are sorted from smallest to largest along the second dimension (second column), and the above operation is repeated for all four dimensions.

[0044] S206 , the Latin hypercube sampling result obtained at this time is a matrix B with a size of np·nv and all elements having a value range of [1 / np,1].

[0045]

[0046] Each row of the matrix B corresponds to a sample point, and each column corresponds to a variable.

[0047] For different application scenarios, the value ranges of the variables in different dimensions for each sample point vary. Matrix B needs to be projected onto the value ranges of the variables in different dimensions for the sample points in the actual problem. The range of variation of the variables in different dimensions for the sample points in the actual problem is denoted as range. The first column of the range matrix serves as the lower limit of the variation of the sample points in different dimensions and is denoted as range(:,1). The second column of the range matrix serves as the upper limit of the variation of the sample points in different dimensions and is denoted as range(:,2). For the compressor blade angle distribution problem, the optimal parameter range prior is given based on a prior. Prior represents the "baseline value" for the possible optimal value of the compressor blade angle distribution, and range fluctuates around prior. Therefore, for the compressor blade angle optimization problem, the final Latin hypercube sampling result C = (range(:,2)-range(:,1)) · B + range(:,1) + prior.

[0048]

[0049] Step 3: Use the obtained Latin hypercube sampling points to train the constructed feedforward back propagation neural network to obtain the trained feedforward back propagation neural network. Specifically: S301 , using a computational fluid dynamics method to obtain two attribute values of np sampling points: isentropic efficiency and total pressure ratio.

[0050] By matching the sample point set with the attribute value set, we get the initial data set O1 (100x6), where the first 4 columns represent the sample point set and the last 2 columns represent the attribute value set.

[0051] S302, randomly shuffle the initial data set to set the training set and the test set, where the percentage of the training set to the initial data set is w=0.8, indicating that the first 80% of the data is used to train the neural network, and the remaining 20% of the data is used to test the neural network.

[0052] S303: Normalize the training and test data sets and input the training data into the newff function. Set the following training parameters: 2 hidden layers, 7 neurons, 4 neurons; logsig activation function for the first hidden layer, tansig activation function for the second hidden layer; trainlm training method; 2500 iterations, 1e-9 target error, 0.05 learning rate, and 1e-9 minimum performance gradient.

[0053] The construction of the feedforward back propagation neural network model is:

[0054] Among them, y represents the neuron input / output, f represents the activation function, w represents the weight, b represents the bias, k represents the current hidden layer of the neural network, i represents the i-th hidden layer neuron, and j represents the j-th neuron in the previous layer.

[0055] The Trainlm method uses the Levenberg-Marquardt algorithm to solve the optimal values of w and b in the above-mentioned neural network. The optimized loss function is:

[0056] in, The last output of the hidden layer neuron Predicted attribute values, Indicates the first attribute values. Q represents the size of the training set data (w=0.8, np=100, then Q=w*np=80); represents the loss function.

[0057] S304, using the test set to test the trained neural network, and calculating the coefficient of determination using the following formula:

[0058] in represents the attribute value in the sample point, represents the average value of the attribute value of the sample point, is the attribute value of the sample point predicted by the neural network, R 2 is the coefficient of determination.

[0059] S305, such as Figure 2 As shown in the figure, after training and testing, the determination coefficients obtained were 0.98114 and 0.97928 respectively, indicating that the trained neural network can highly fit the initial data set O1 (100x6).

[0060] Step 4: Use the fast non-dominated sorting genetic algorithm to solve the trained feedforward back propagation neural network and search for the Pareto front in the target sample space, that is, the supercritical carbon dioxide compressor blade angle distribution. Specifically: S401, initialize the genetic algorithm, set the number of optimization targets M = 2, the sample dimension V = 4, the number of individuals in the population pop = 300, and set the upper and lower limits to range + prior:

[0061] Set the maximum number of iterations gen=1000, the crossover distribution index h=11, the mutation distribution index r=15, and the mutation probability pm=0.04. Generate a random number matrix of size popxV and multiply the corresponding variable in the matrix by the value of the corresponding variable in the width of range+prior:

[0062] And add the first column of the range+prior matrix (the lower limit of all variables) range to get the initial random population.

[0063] S402, using the trained neural network to predict the attribute value of each individual in the initial random population, merging the initial random population with the initial population attribute value to obtain an updated random population; S403: Use the population non-dominated sorting algorithm to output the dominance level and crowding distance of each individual in the updated random population. For each attribute value, the crowding distance can be calculated according to the following formula:

[0064] Among them, f represents the attribute value, x represents the sample point, and cd represents the crowding distance.

[0065] S404: Add the crowding distances of all attribute values to obtain the crowding distance of the individual. Specifically, assign infinite crowding distances to the two endpoints of each level. Output the populations arranged by dominance hierarchy to obtain the new parent population. The following shows the first 10 rows of the arrangement results (decimals have been retained for ease of viewing):

[0066] Among them, the first four columns represent individuals (sample points), columns 5-6 are attribute values, column 7 is the dominance level, and column 8 is the crowding distance.

[0067] The dominance level indicates that in the multi-objective optimization process, if the sample point A has at least one attribute value f If A is better than sample point C, and all attribute values of sample point A are not worse than those of sample point C, then sample point A is said to dominate sample point C. In the entire sample set, if no other sample point dominates sample point A, then the dominance level of sample point A is 1 (the highest level), and all such sample points A and B constitute the Pareto frontier.

[0068] S405: Input the new parent population into the selection algorithm, compare the dominance level and crowding distance of each individual pairwise, and prioritize individuals with low dominance level and large crowding distance to obtain the parent population.

[0069] S406, performing crossover and mutation on the individuals selected in S405, performing crossover and mutation operations according to the following method: Compare the parent individuals pairwise. If the two individuals are completely equal or the generated random number is greater than 0.5, the offspring individual is directly set as the parent individual; otherwise, the following calculation is performed:

[0070] Among them, p2 represents the element of the rth dimension of individual 2, p1 represents the element of the rth dimension of individual 1, and x l Represents the minimum value of the rth dimension, x u represents the maximum value of the rth dimension, and h represents the cross-distribution index.

[0071] Before calculating this formula, you need to compare the sizes of p2 and p1 and set the larger element to p2. Calculation:

[0072] Among them, rand represents a random number, and the inequality result outputs a logical vector. Based on this result, the new offspring individuals can be calculated as follows:

[0073] Among them, c1 represents offspring individual 1, and c2 represents offspring individual 2.

[0074] Perform a crossover operation on all parent populations to update the population.

[0075] Perform the following mutation operations on the updated population:

[0076] Among them, r represents the mutation distribution index, pm represents the mutation probability, x represents the individual, and c represents the offspring population obtained after the parent population is mutated.

[0077] S407, record the population after the above selection, crossover, and mutation operations as the offspring population, use the trained feedforward back propagation neural network model to predict the attribute value of each individual in the offspring population, and obtain a new offspring population; Merge the new offspring population and the parent population into a fused population.

[0078] S408: In the fused population, a population non-dominated sorting algorithm is used to output the dominance level and crowding distance of each individual. The population arranged according to the dominance level is output as the latest parent population according to the number of individuals in the population, and the remaining individuals are eliminated.

[0079] The following shows the arrangement results of the first 10 rows of the new parent population after iterating the above genetic process 6 times (the decimal part is retained as necessary to make the results easier to view):

[0080] It can be seen that after only 6 iterations, the dominant individuals (high dominance level) have begun to dominate the population. After 1000 iterations, the results of columns 5-6 are output in reverse order, and we get Figure 7 Results shown. Figure 7 The red line in the middle is a scatter plot of 300 individuals whose dominance levels are all 1. This plot can be output as a Pareto solution set, and any individual in the solution set can be used as the optimal value of the compressor blade angle distribution.

[0081] By adopting the blade angle distribution improvement and optimization method of the present application, when optimizing the compressor blade angle, there is no restriction on the number of optimization curve points. In this embodiment, a 6-degree-of-freedom Bezier curve is used as the blade angle distribution curve. The first and last two points are determined by the aerodynamic constraints of the compressor and therefore do not participate in the optimization. The vertical coordinate positions of the four middle points constitute the optimization problem of this example. This method can also be applied to Bezier curve optimization problems with more degrees of freedom. Furthermore, it can also meet the optimization of other design spline curve control points of the compressor. There is no restriction on the type of compressor, and it can realize the rapid optimization of the compressor design line control points. It is easy to implement and has low optimization cost. It is a key technology for realizing compressor line optimization in large-scale design applications and has broad application prospects.

[0082] The sampling-model-optimization framework formed by this embodiment establishes a proxy model between the pre-optimization parameters and the optimization target, given a defined range of pre-optimization parameter values. This framework then globally optimizes the proxy model to determine the optimal value of the optimization target within the pre-optimization parameter space. This framework allows for optimal pre-optimization parameter-optimization target results based on the given prior (priro) values of the pre-optimization parameters and the upper and lower limits of the parameter variation range. For determining the blade angle distribution of a compressor, this method relies on a small amount of prior information, rather than a large empirical library of compressor model design samples, to determine the optimal blade angle distribution design within a given pre-optimization parameter space.

[0083] Example 3 This embodiment provides a supercritical carbon dioxide compressor blade angle distribution improvement and optimization system, comprising: Sample space construction unit, used to construct a sample space of specified dimensions; A sampling unit, used to select uniformly distributed Latin hypercube sampling results from the sample space using a translation propagation algorithm; A model training unit, configured to train a constructed feedforward back propagation neural network model according to the selected Latin hypercube sampling result, and obtain a trained feedforward back propagation neural network model; The model solving unit is used to solve the trained feedforward back propagation neural network model using a fast non-dominated sorting genetic algorithm to obtain the optimal blade angle distribution.

[0084] Example 4 This embodiment also provides a computing device. The computing device includes a bus, a processor, a memory, and a communication interface. The processor, the memory, and the communication interface communicate with each other via the bus. The computing device can be a server or a terminal device. It should be understood that this application does not limit the number of processors and memories in the computing device.

[0085] A bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. Buses can be categorized as address buses, data buses, control buses, and so on. For ease of presentation, a bus can include the pathways that transmit information between various components of a computing device (e.g., memory, processor, and communication interfaces).

[0086] The processor may include any one or more of a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a microprocessor (MP), or a digital signal processor (DSP).

[0087] The memory may include volatile memory, such as random access memory (RAM). The processor may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0088] The memory stores executable program code, and the processor executes the executable program code to implement the functions of each of the aforementioned units, thereby implementing, for example, the method described in Example 2. That is, the memory may store instructions for the methods and functions of the computing device described in any of the above embodiments.

[0089] The communication interface uses a transceiver module such as, but not limited to, a network interface card or a transceiver to implement communication between the computing device and other devices or a communication network.

[0090] Example 5 This embodiment further provides a computer-readable storage medium having computer instructions stored thereon. When a processor executes the instructions, the processor executes the methods and functions related to the computing device in any of the above embodiments.

[0091] In general, various embodiments of the present disclosure may be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects may be implemented in hardware, while other aspects may be implemented in firmware or software, which may be executed by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of the present disclosure are shown and described as block diagrams, flow charts, or using some other pictorial representation, it should be understood that the blocks, devices, systems, techniques, or methods described herein may be implemented as, by way of non-limiting example, hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or a controller or other computing device, or some combination thereof.

[0092] Example 6 The present embodiment provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which are executed in a device on a real or virtual processor of a target to perform the process / method described above with reference to the accompanying drawings. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided between program modules as needed. The machine-executable instructions for the program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0093] The computer program code for implementing the disclosed method can be written in one or more programming languages. These computer program codes can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device so that the program code, when executed by the computer or other programmable data processing device, causes the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer or entirely on a remote computer or server.

[0094] In the context of the present disclosure, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, acoustic, or other forms of propagated signals, such as carrier waves, infrared signals, and the like.

[0095] A computer-readable medium may be any tangible medium containing or storing a program for or relating to an instruction execution system, apparatus, or device, or a data storage device such as a data center containing one or more available media. A computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. A computer-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination thereof. More detailed examples of computer-readable storage media include an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0096] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A method for improving and optimizing the blade angle distribution of a supercritical carbon dioxide compressor, characterized in that: The following steps are involved: Step 1: construct a sample space of specified dimension; Step 2: Use the translation propagation algorithm to select uniformly distributed Latin hypercube sampling results from the sample space; Step 3: training the constructed feedforward back propagation neural network model according to the selected Latin hypercube sampling result to obtain a trained feedforward back propagation neural network model; Step 4: Use the fast non-dominated sorting genetic algorithm to solve the trained feedforward back propagation neural network model to obtain the optimal blade angle distribution.

2. The method for improving and optimizing the blade angle distribution of a supercritical carbon dioxide compressor according to claim 1, characterized in that: Construct a sample space of specified dimension. The specific method is: Set the desired number of sample points in the sample space of a specified dimension, where each sample point includes four dimensions; Set the value range for each dimension.

3. The method for improving and optimizing blade angle distribution of a supercritical carbon dioxide compressor according to claim 1, characterized in that: The translation propagation algorithm is used to select uniformly distributed Latin hypercube sampling results from the sample space. The specific method is: Set the initial seed module in the sample space; Determine the number of partitions of the sample space based on the set initial seed module; Reshape the initial seed module using the determined number of partitions to obtain a reshaped seed module; The reshaped seed module is translated and updated according to the number of dimensions and the filling vector until all partitions in the sample space are filled to obtain the result matrix; Perform Latin hypercube sampling on the result matrix to obtain a Latin hypercube sampling result.

4. The method for improving and optimizing the blade angle distribution of a supercritical carbon dioxide compressor according to claim 3, characterized in that: The number of partitions of the sample space is determined based on the set initial seed module. The specific method is: The number of sample points in the set sample space is divided by the seed value of the seed module to obtain the sample point ratio; Take the square root of the obtained sample point ratio and get the area value of the seed module size of the corresponding dimension; Round up the obtained area value to get the rounded partition number; The number of rounded partitions is compared with the regional value of the seed module size in each dimension. If the number of rounded partitions is greater than the regional value of the seed module size in each dimension, the number of partitions is the power of the number of sample dimensions. Otherwise, the number of partitions is the ratio of sample points, and the sample space size is the number of partitions multiplied by the number of sample points in the seed module.

5. The method for improving and optimizing blade angle distribution of a supercritical carbon dioxide compressor according to claim 1, characterized in that: In step 3, the expression of the constructed feedforward back propagation neural network model is: Where y represents the neuron input / output, f represents the activation function, w represents the weight, b represents the bias, k represents the current hidden layer of the neural network, i represents the i-th hidden layer neuron, and j represents the j-th neuron in the previous layer; The loss function is: in, The last output of the hidden layer neuron Predicted attribute values, Indicates the first attribute values; represents the loss function; Q represents the size of the training set data.

6. The method for improving and optimizing blade angle distribution of a supercritical carbon dioxide compressor according to claim 1, characterized in that: The fast non-dominated sorting genetic algorithm is used to solve the trained feedforward back-propagation neural network model to obtain the optimal blade angle distribution. The specific method is: S1, initialize the genetic algorithm to obtain the initial random population; S2, using the trained neural network to predict the attribute value of each individual in the initial random population to obtain the attribute value of the initial population; S3, merging the initial random population with the initial population attribute value to obtain an updated random population; S4, using the population non-dominated sorting algorithm to calculate the dominance level and crowding distance of each individual in the random population, output the population arranged according to the dominance level according to the number of individuals in the population, and obtain the new parent population; S5, using the selection algorithm to select individuals that meet the set requirements from the new parent population to obtain the parent population; S6, performing crossover and mutation processing on the individuals in the selected parent population to obtain the offspring population; S7, using the obtained offspring population as the input of the trained feedforward back-propagation neural network model to predict the attribute value of each individual in the offspring population and obtain a new offspring population; S8, merge the new offspring population and the parent population in S5 into a fusion population; S9, using the population non-dominated sorting algorithm to calculate the dominance level and crowding distance of each individual in the fused population, output the populations sorted according to the dominance level according to the number of individuals in the population, and obtain the latest parent population; S10, repeat S2 to S9 with the latest parent population until the number of iterations is reached; S11, inverting the attribute value of the final population obtained to obtain the optimal value of the compressor blade angle distribution.

7. A supercritical carbon dioxide compressor blade angle distribution improvement and optimization system, characterized in that: include: Sample space construction unit, used to construct a sample space of specified dimensions; A sampling unit, used to select uniformly distributed Latin hypercube sampling results from the sample space using a translation propagation algorithm; A model training unit, configured to train a constructed feedforward back propagation neural network model according to the selected Latin hypercube sampling result, and obtain a trained feedforward back propagation neural network model; The model solving unit is used to solve the trained feedforward back propagation neural network model using a fast non-dominated sorting genetic algorithm to obtain the optimal blade angle distribution.

8. An electronic device, characterized in that: The electronic device comprises a processor and a memory, wherein computer instructions are stored in the memory. When the computer instructions are executed by the processor, the electronic device executes the method according to any one of claims 1 to 6.

9. A computer program product, characterized in that The computer program product contains computer-executable instructions, which implement the method according to any one of claims 1 to 6 when executed.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the method according to any one of claims 1 to 6.

Images