A method for optimizing model parameters based on wave energy device model test errors

Abstract

The present invention discloses a method for optimizing model parameters for wave energy device model test errors, specifically relating to the technical field of model parameter optimization. The method comprises step S1, collecting long-term accumulated wave energy device test data and establishing a test database; step S2, orthogonal test design; and step S3, Bayesian parameter optimization: a machine learning method based on a Bayesian Gaussian process is used to train and optimize variable factors, thereby reducing the scale effect of the model test. The present invention optimizes and selects the next model parameters based on existing model test parameters, thereby minimizing the scale effect of the wave energy device model. A small number of selected sample points are used to automatically select iterative sample points through continuous iteration using a machine learning optimization method, thereby reducing the cost of the model test and improving the test efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of model parameter optimization, and in particular to a method for optimizing model parameters based on model test errors of a wave energy device. Background Art

[0002] Wave energy is a key form of ocean energy, possessing abundant reserves and providing the world with a renewable, clean energy source. With the gradual depletion of fossil fuels and the growing prevalence of environmental issues, the development and utilization of ocean energy has become a crucial means of alleviating the global energy crisis. Ocean energy comes in various forms, including tidal energy, current energy, thermal energy, and wave energy. Wave energy, due to its high energy density and relatively stable properties, has become a key focus of research and development. In recent years, research institutions and businesses worldwide have actively engaged in the development and optimization of wave energy conversion devices to improve energy conversion efficiency and device reliability.

[0003] During the development of wave energy devices, model testing is an important means of evaluating their performance. Because prototype construction and testing are expensive and experimental conditions are difficult to fully control, experimental studies are often conducted using scaled models. The key to model testing is ensuring that the scaled model accurately reproduces the behavior of the prototype device under real ocean conditions. Commonly used similarity criteria include the Froude similarity criterion and the Reynolds similarity criterion to ensure mechanical similarity between the scaled model and the prototype with respect to gravity, inertia, viscosity, friction, and other factors. However, these two similarity criteria are often difficult to simultaneously meet under model scaling conditions, especially given the multiple mechanical effects involved in wave energy devices. This leads to scale effects, which adversely affect the accuracy of model test results and, in turn, the performance evaluation of the prototype device. Currently, model tests of wave energy devices primarily utilize the Froude similarity criterion for scaling calculations. Froude similarity is primarily used for gravity-dominated phenomena, such as wave propagation and free surface motion, and is suitable for simulating the primary hydrodynamic behaviors in wave energy devices. However, the effects of viscosity and friction require the Reynolds similarity criterion, and during model scaling, these two similarity criteria cannot be simultaneously met. This results in differences in the performance of the scaled model and the prototype device under the action of viscosity and friction, resulting in scale effects that affect the accurate evaluation of energy conversion efficiency and performance.

[0004] In existing technical solutions, although the use of Froude similarity can better simulate the macroscopic behavior of waves, the Reynolds similarity is difficult to achieve, resulting in errors in the model test results. Especially when the wave energy device involves complex flows, viscous effects and multiple dynamic interactions, this error will significantly affect the reliability of the model test results. The main shortcoming of the existing technical solutions is the lack of an efficient and systematic method to reduce the influence of scale effects. Although numerical simulation and experimental correction can partially solve this problem, due to the complexity of the wave energy device and the limitations of experimental costs, the existing methods still have a lot of room for improvement in efficiency and accuracy. Specifically, the shortcomings of the existing technology include:

[0005] 1. Lack of systematicity: The existing experimental design and analysis methods lack systematic data management, and are unable to comprehensively analyze and summarize the impact of scale effects on model experiments under various variables (such as device structure, PTO system type, model scale ratio, wave conditions, etc.). There is also a lack of means to establish a systematic database to manage experimental data, resulting in poor comprehensiveness of experimental analysis and stability of results.

[0006] 2. Lack of efficient experimental design methods: Existing technologies usually rely on experience and manual adjustments when faced with multiple variables and experimental conditions, making it difficult to design experiments systematically. In order to evaluate the impact of the scale effect on model experiments under different experimental conditions, a large number of repeated experiments are often required, which is time-consuming and costly. The lack of efficient experimental methods such as orthogonal experimental design makes it difficult to achieve rapid and low-cost systematic analysis and quantification of multiple variables.

[0007] 3. Limited optimization capabilities: The optimization process in existing methods mainly uses empirical formulas or simple exhaustive methods to adjust parameters. There is a lack of systematic means based on mathematical models and optimization algorithms, making it difficult to find the globally optimal parameter settings under complex multivariable conditions. Summary of the Invention

[0008] The purpose of the present invention is to provide a method for optimizing model parameters of wave energy device model test errors, aiming to systematically analyze the main variables affecting the scale effect by combining different means, and reduce the influence of the scale effect on the test results.

[0009] In order to achieve the above-mentioned object, the present invention provides the following technical solution: a method for optimizing model parameters of a wave energy device model test error, comprising the following steps;

[0010] Step S1: collecting long-term accumulated wave energy device test data and establishing a test database;

[0011] Step S2, orthogonal experimental design;

[0012] S2.1, Classification of variable factors;

[0013] S2.2, design orthogonal test table;

[0014] S2.3. Scale Effect Assessment: Compare the mean scale effect differences of the variables to preliminarily determine the importance of the variables in the scale effect. Then, perform variance analysis to determine whether the variables have a significant impact on the scale effect.

[0015] Step S3, Bayesian parameter optimization: Based on the machine learning method of Bayesian Gaussian process, the variable factors are trained and optimized to reduce the scale effect of the model test.

[0016] Preferably, the establishment of the test database in S1 includes:

[0017] S1.1. Database classification:

[0018] S1.1.1. Types of wave energy device structures: including oscillating float type, swinging float type, and horizontal oscillating body type;

[0019] S1.1.2. PTO device types: including hydraulic systems, mechanical systems, and linear generators;

[0020] S1.1.3 Model scaling: The scaling scale should be selected based on the prototype device. Common scales include 1:10, 1:20, or 1:50.

[0021] S1.1.4 Wave conditions: including wave conditions (regular waves, irregular waves) and wave direction (beam waves, oblique waves);

[0022] S1.2 Data extraction:

[0023] S1.2.1. Multi-level screening method based on classification characteristics: Based on the wave energy device structure type, PTO device type, and operating conditions, a multi-level screening method is used to gradually narrow the data range;

[0024] S1.2.2, Keyword matching and filtering: In the database, preliminary filtering of data is performed through keyword matching;

[0025] S1.3 Data Optimization:

[0026] S1.3.1. Statistical method regression analysis: Regression analysis (such as linear regression, polynomial regression) is used to establish the relationship between classification characteristics (such as device structure, PTO device type, model scaling, wave conditions, etc.) and response (scale effect), and weights are assigned to each classification feature. The test data are optimized based on these weights.

[0027] Preferably, the variable factors in S2.1 include the wave energy device structure type, PTO device type, model scaling ratio and wave conditions.

[0028] Preferably, in S2.2, for four factors, the wave energy device structure type and the PTO device type have three variable levels, and the wave working condition and the model scaling ratio each have two variable levels. In order to ensure that the combination of different factors is fully covered, the present invention uses the L18 orthogonal table for experimental design.

[0029] Preferably, the specific steps of S3 are as follows:

[0030] S3.1. Input the initial sample points;

[0031] S3.2, minimize the objective function;

[0032] S3.3. Using Bayesian optimization to establish a surrogate model, which is a Gaussian process, to approximate the objective function;

[0033] S3.4. Set the number of iterations. In each iteration, use the known data to calculate the proxy model through the Gaussian process, and the acquisition function selects the next evaluation point. During the iteration process, the proxy model is continuously updated and the parameters are adjusted according to the posterior distribution until the optimal solution is found, thereby reducing the scale effect of the model test.

[0034] Preferably, the optimization model parameters obtained through machine learning iteration need to be trial-calculated and compared with the true experimental values ​​to verify the accuracy of the Bayesian optimization method calculation.

[0035] Preferably, the model parameters after Bayesian optimization are used to design test conditions based on the elimination method of orthogonal experiment, and the influence of non-similar factors on the model test is eliminated by the elimination method. Thereafter, the number of independent test conditions is increased according to the number of parameters to be measured, and finally the true value response change law that is not interfered by the scale effect is obtained.

[0036] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0037] 1. Improve test efficiency and reduce costs: By establishing a test database, we can collect and organize the long-term accumulated wave energy device test data, classify and filter the test data, and thus reduce the need for repeated tests;

[0038] 2. Systematically analyze the impact of multiple variables: Based on the model test data in the database, an orthogonal experimental design is used to systematically summarize and analyze multiple variables such as different wave energy device structure types, PTO device types, and wave conditions. This allows for a comprehensive understanding of the impact of various variables on the scale effect, thereby optimizing the experimental design and result interpretation.

[0039] 3. Intelligently optimize key parameters to reduce scale effects: Utilizing a Bayesian optimization algorithm, this approach uses a Gaussian process to proxy model test data, intelligently selecting the optimal parameter points for the next test and gradually optimizing key parameters to minimize the power difference between the scaled model and the prototype. This method can identify globally optimal parameter settings using a small amount of test data, effectively reducing the impact of scale effects on test results and significantly reducing testing costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a flow chart of the method of the present invention;

[0041] Figure 2 is the damping force of the PTO device in the embodiment of the present invention and model output power relationship diagram;

[0042] Figure 3 is the damping force of the PTO device in the embodiment of the present invention Relationship diagram of the objective function power difference E;

[0043] Figure 4 Graph showing the convergence process of the objective function power difference E after 20 iterations in an embodiment of the present invention;

[0044] Figure 5 The third and fourth iterations of the prediction function and the acquisition function distribution diagram in the embodiment of the present invention are as follows: (a) Bayesian optimization Gaussian process prediction diagram for the third iteration; (b) Bayesian optimization acquisition function distribution diagram for the third iteration; (c) Bayesian optimization Gaussian process prediction diagram for the fourth iteration; (d) Bayesian optimization acquisition function distribution diagram for the fourth iteration;

[0045] Figure 6 The following are the prediction functions and acquisition function distribution diagrams for the 6th and 7th iterations in an embodiment of the present invention; (a) Bayesian optimization Gaussian process prediction diagram for the 6th iteration; (b) Bayesian optimization acquisition function distribution diagram for the 6th iteration; (c) Bayesian optimization Gaussian process prediction diagram for the 7th iteration; (d) Bayesian optimization acquisition function distribution diagram for the 7th iteration;

[0046] Figure 7 17 and 18 iterations of the prediction function and the acquisition function distribution graph in the embodiment of the present invention; (a) Bayesian optimization Gaussian process prediction graph for the 17th iteration; (b) Bayesian optimization acquisition function distribution graph for the 17th iteration; (c) Bayesian optimization Gaussian process prediction graph for the 18th iteration; (d) Bayesian optimization acquisition function distribution graph for the 18th iteration;

[0047] Figure 8 This is a distribution diagram of the prediction function after 20 iterations in an embodiment of the present invention; DETAILED DESCRIPTION

[0048] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0049] The present invention provides Figure 1 A method for optimizing model parameters of a wave energy device model test error is shown, comprising the following steps:

[0050] Step S1: Establish a systematic test database to collect and manage various test data of wave energy devices. The test steps are as follows:

[0051] S1.1. Database classification:

[0052] S1.1.1. The main research object of this invention is an oscillating body type wave energy device. The structural types of wave energy devices are divided into the following types according to the degree of freedom of the floating body: including the oscillating float type, which mainly converts energy by up and down swinging, usually with a float (or multiple floats) installed on the sea surface; the swinging float type, which uses the horizontal movement of waves to drive the float to swing back and forth or left and right; the horizontal oscillating body type, which consists of a horizontally placed oscillating body and uses the horizontal movement of the oscillating body to drive a hydraulic or mechanical system to generate electricity;

[0053] S1.1.2. PTO device types: including hydraulic systems, mechanical systems, and linear generators;

[0054] S1.1.3 Model scaling: According to the different scaling scales of the prototype, common ones include 1:10, 1:20 or 1:50 scale ratios;

[0055] S1.1.4. Wave conditions: wave conditions (regular waves, irregular waves) and wave direction (beam waves, oblique waves);

[0056] S1.2 Data extraction:

[0057] S1.2.1. Multi-level screening method based on classification characteristics: Based on the wave energy device structure type, PTO device type, and wave conditions, a multi-level screening method is used to gradually narrow the data range;

[0058] S1.2.2. Keyword matching and filtering: In the database, preliminary filtering of data is performed by keyword matching (e.g., “oscillating float type”, “regular wave transverse wave”, “scaling scale 1 / 10”, etc.);

[0059] S1.3 Data Optimization:

[0060] S1.3.1. Statistical method regression analysis: Regression analysis (such as linear regression and polynomial regression) is used to establish the relationship between classification features (such as wave energy device structure type, PTO device type, model scaling ratio, wave operating conditions, etc.) and response (scale effect), and weights are assigned to each classification feature. According to the weights assigned to each classification feature, the data are optimized based on these weights.

[0061] Step S2: Orthogonal experimental design. Orthogonal experiment is a systematic experimental design method that can be used to analyze the impact of multiple factors on a certain result and obtain the combined effect of multiple factors and multiple levels with a smaller number of experiments. The specific steps are as follows:

[0062] S2.1. Classify the variable factors, including the wave energy device structure type, PTO device type, wave conditions, and model scaling;

[0063] S2.2. Design an orthogonal test table. For each of the four factors, wave energy device structure type and PTO device type each have three variable levels, and wave conditions and model scaling each have two variable levels. To ensure sufficient coverage of different factor combinations, an L18 (3^2 × 2^2) orthogonal table was used for the experimental design. This design effectively reduces the number of experiments and ensures sufficient coverage of each factor combination. The orthogonal experimental design is shown in the following table:

[0064] Test number PTO device type Wave energy device structure type Wave conditions Model scale 1 hydraulic system Oscillating float type rule 1 / 10 2 hydraulic system Swing float type irregular 1 / 20 3 hydraulic system Horizontal Oscillation Pose rule 1 / 20 4 Mechanical system Oscillating float type irregular 1 / 10 5 Mechanical system Swing float type rule 1 / 10 6 Mechanical system Horizontal Oscillation Pose irregular 1 / 20 7 Linear generator system Oscillating float type rule 1 / 20 8 Linear generator system Swing float type irregular 1 / 10 9 Linear generator system Horizontal Oscillation Pose rule 1 / 10 10 hydraulic system Oscillating float type irregular 1 / 20 11 hydraulic system Swing float type rule 1 / 20 12 hydraulic system Horizontal Oscillation Pose irregular 1 / 10 13 Mechanical system Oscillating float type rule 1 / 10 14 Mechanical system Swing float type irregular 1 / 20 15 Mechanical system Horizontal Oscillation Pose rule 1 / 20 16 Linear generator system Oscillating float type irregular 1 / 10 17 Linear generator system Swing float type rule 1 / 10 18 Linear generator system Horizontal Oscillation Pose irregular 1 / 20

[0065] S2.3. Assessment of scale effect:

[0066] Orthogonal test tables can be used to evaluate the impact of each variable factor on the final scale effect. For example, to evaluate the impact of PTO device type on the scale effect, the 18 test data mentioned above are classified and grouped according to PTO device type: test numbers for hydraulic systems: 1, 2, 3, 10, 11, 12; test numbers for mechanical systems: 4, 5, 6, 13, 14, 15;

[0067] Test numbers for PTO device type linear generator system: 7, 8, 9, 16, 17, 18;

[0068] For all tests of each PTO device type, the average response value of the scale error due to the scale effect is calculated as follows:

[0069] Average scale error of PTO device in hydraulic system ;

[0070] Average scale error of mechanical system PTO device ;

[0071] Average scale error of PTO device of linear generator system ;

[0072] in, : The number is the scale error in the i-th trial.

[0073] Comparing the average scale errors corresponding to different PTO device types, if the difference between these average scale error response values ​​is large, it means that the scale effect is greatly affected by the PTO device type. Therefore, by comparing the differences in average scale errors, we can preliminarily determine whether the PTO device type has a significant impact on the scale effect;

[0074] S2.4. To further quantify the effect of PTO device type on the scale effect, an analysis of variance (ANOVA) was performed. The main steps are as follows:

[0075] First, the variance between PTO device types (between-group variance) was calculated to quantify the extent to which PTO device type affects the size effect. , : the mean scale error of the i-th group of experiments, : Mean square error of all test samples; mean square between groups , the degree of freedom between groups in the designed experiment is 2;

[0076] Then, the within-group variance is calculated to represent the scale effect variation caused by other factors (wave energy device structure type, wave conditions, model scaling). , : The scale error value of the jth trial in the i-th group; Mean Square Within groups , the degree of freedom within the group in the designed experiment is 15;

[0077] Subsequently, the F value was calculated and the F-ratio was used to determine whether the influence of the PTO device type on the scale effect was significant. , the calculated F value and critical value Comparison, assuming a significance level , indicating that there is a 5% probability of making mistakes, calculated by statistical software It is approximately equal to 3.68. If the calculated F value is greater than 3.68, it means that the difference in scale effect between different PTO device types is significant at the 95% confidence level.

[0078] Step S3: Bayesian parameter optimization: After this, assuming that the orthogonal test method verifies that the PTO device type has a significant impact on the scale effect in the model test, the Bayesian optimization algorithm is then used to optimize the key parameters of the PTO device to reduce the impact of the scale effect and improve the accuracy of the model test. The specific steps of the test are as follows:

[0079] (1) Establishment of objective function: In the study of model scaling, the PTO system is usually scaled according to Froude similarity. Under the Froude similarity condition, assuming that λ is the geometric scale scaling factor, the force scaling ratio is λ 3 , the speed scaling factor is At this time, the energy conversion efficiency of the scaled model can be ensured to be similar to that of the prototype, that is, the similarity of power can be ensured, and the energy absorption characteristics of the floating body in the model are consistent with those of the prototype; the power generation power of the PTO device ,in, The damping force applied by the PTO device, is the velocity of the floating body in the model test; the power scaling ratio is ; Due to the velocity of the floating body in the model test It is usually difficult to control, and the energy loss (such as friction loss, leakage of hydraulic system, etc.) may also be different from the prototype. Therefore, by controlling the damping force of the PTO device The energy loss in the scaled model can be matched with the prototype to ensure the consistency of energy conversion efficiency; based on the Bayesian optimization algorithm, the damping force of the PTO device is dynamically adjusted , in order to minimize the defined objective function E, its expression form is as follows:

[0080]

[0081] in and The power outputs of the PTO devices of the model and prototype are respectively. By minimizing the objective function E, the consistency of the power outputs of the model and prototype is achieved. To simulate the actual situation, the following assumptions and steps are made:

[0082] 1. Assume that the PTO device damping force and model output power There is a nonlinear relationship between Figure 2 As shown, its true expression is a complex black box function; PTO damping force The value range of is between 0 and 630 N / (m / s); the transparent shaded area in the figure is the 95% confidence interval, which simulates the uncertainty in the actual model test by introducing a noise that conforms to the normal distribution;

[0083] 2. Assuming the power output of the prototype PTO device , geometric scaling factor . Figure 3 Demonstrates the PTO device damping force The relationship between the power difference E of the objective function and the transparent shaded area is also the 95% confidence interval;

[0084] (2) Iterative optimization: Call the gp_minimize function in the scikit-optimize library (other optimization libraries can be used instead, such as Optuna or Hyperopt. For example, Optuna can perform adaptive sampling and has efficient parallel optimization features, which is suitable for large-scale optimization tasks; while Hyperopt provides a search method based on distribution estimation, which is suitable for dealing with optimization problems with complex search spaces) and minimize the objective function E based on a surrogate model called Gaussian Process (GP) (other alternative methods such as random forest, support vector regression (SVR) or neural network can be used. These methods have different advantages when processing different types of data. For example, random forest performs well in processing high-dimensional and nonlinear data, while neural networks are very powerful in modeling complex relationships). The main steps are as follows:

[0085] 1. Randomly select 3 known sample points within the PTO damping force parameter range and set , as the input of the initial data, construct a proxy model of the Gaussian process; Gaussian process is a probability model used to model uncertain functions, which regards the value of the function as a random variable and assumes that the joint distribution of these random variables in the domain of definition is a multivariate normal distribution; Gaussian process defines a mean function and a covariance function To describe the mean and uncertainty of the objective function; when known sample data is input, the Gaussian process uses Bayes' theorem to update the prior distribution to the posterior distribution, and obtains the posterior prediction of the objective function, thereby quantifying the uncertainty of the unobserved area;

[0086] 2. After the Gaussian process predicts the objective function based on the existing data, Bayesian optimization can intelligently select the next input PTO damping force parameter to optimize the objective function. This is achieved through an acquisition function called α. In the algorithm, the present invention selects a method called Expected Improvement (EI) as the acquisition function, namely ; The acquisition function is used to select the next evaluation point to maximize the expected improvement of the objective function, ensuring a balance between global search and local development in each iteration;

[0087] 3. Then, set the number of iterations , represents the maximum number of iterations of Bayesian optimization; Figure 4 The convergence of the objective function after 20 iterations is shown. During each iteration, Bayesian optimization intelligently selects the next sample point within the search domain by setting a search metric for the acquisition function α. ​​The acquisition function α performs a fine search around the optimal value of the current objective function while simultaneously finding new sample points in unexplored areas, thus maintaining a balance between exploration and exploitation. Figure 5 The third and fourth iterations of Bayesian optimization are shown from top to bottom. The red dots in the left figure are known sample points or new sample points updated based on the acquisition function α. ​​The green dotted line represents the prediction of the true objective function by the Gaussian process proxy model based on the existing sample points. The transparent shaded area is the 95% confidence interval. The right figure is the acquisition function distribution diagram. The point corresponding to its maximum value is the sample point that may have the greatest improvement on the objective function. Similarly, Figure 6 and Figure 7 The process of the 6th, 7th, 17th and 18th iterations are shown respectively. In each iteration, the acquisition function α continuously and intelligently selects sample points to update the objective function value, thereby improving the proxy model and adjusting the parameters based on the posterior distribution. After 20 iterations, the optimal objective function value obtained by Bayesian optimization is as follows Figure 8 As shown; when the damping force of the PTO device , objective function power difference Although the same prototype power There is still a small error compared to the original, but after a limited number of iterations, this error is small enough, indicating that the model has approached the optimal solution; this method can obtain the optimal parameters close to the true value through a small amount of test data, which greatly reduces the cost of model testing and improves test efficiency.

[0088] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

Claims

1. A method for optimizing model parameters based on model test errors of a wave energy device, characterized in that: The method includes the following steps: Step S1: collecting long-term accumulated wave energy device test data and establishing a test database; S1.

1. Database classification: S1.1.

1. Types of wave energy device structures: including oscillating float type, swinging float type, and horizontal oscillating body type; S1.1.

2. PTO device types: including hydraulic system, mechanical system and linear generator type; S1.1.3 Model scaling: Select the scaling scale based on the prototype device. S1.1.4 Wave conditions: including wave conditions and wave direction; S1.2 Data extraction: S1.2.

1. Multi-level screening method based on classification features: Based on the wave energy device structure type, PTO device type, model scaling ratio, and wave conditions, a multi-level screening method is used to gradually narrow the data range; S1.2.2, Keyword matching and filtering: In the database, preliminary filtering of data is performed through keyword matching; S1.3 Data Optimization: S1.3.

1. Statistical Methods Regression Analysis: Regression analysis is used to establish the relationship between classification features and responses, and weights are assigned to each classification feature. The test data are then optimized based on these weights. Step S2, orthogonal experimental design; S2.

1. Classify the variable factors, including the wave energy device structure type, PTO device type, model scaling ratio, and wave conditions; S2.

2. Design an orthogonal test table for four factors. The wave energy device structure type and PTO device type have three variable levels, and the wave conditions and model scaling each have two variable levels. To ensure that different combinations of factors are fully covered, an L18 orthogonal table is used for the experimental design. S2.

3. Scale Effect Assessment: Compare the mean scale effect differences of the variables to preliminarily determine the importance of the variables in the scale effect. Then, perform variance analysis to determine whether the variables have a significant impact on the scale effect. Step S3, Bayesian parameter optimization: Based on the machine learning method of Bayesian Gaussian process, the variable factors are trained and optimized to reduce the scale effect of the model test; S3.

1. Input the initial sample points; S3.2, minimize the objective function; S3.

3. Using Bayesian optimization to establish a surrogate model, which is a Gaussian process, to approximate the objective function; S3.

4. Set the number of iterations. In each iteration, use the known data to calculate the proxy model through the Gaussian process, and the acquisition function selects the next evaluation point. During the iteration process, the proxy model is continuously updated and the parameters are adjusted according to the posterior distribution until the optimal solution is found, thereby reducing the scale effect of the model test.

2. A method for optimizing model parameters based on model test errors of a wave energy device according to claim 1, characterized in that: The optimization model parameters obtained through machine learning iteration need to be trial-calculated and compared with the true experimental values ​​to verify the accuracy of the Bayesian optimization method calculation.

3. The method for optimizing model parameters based on model test errors of a wave energy device according to claim 1, characterized in that: After Bayesian optimization, the model parameters are used to design test conditions based on the elimination method of orthogonal experiments. The influence of non-similar factors on the model test is eliminated by the elimination method. Then, according to the number of parameters to be measured, the number of independent test conditions is increased, and finally the true value response change law that is not interfered by the scale effect is obtained.

Images