Method and system for predicting structural vibration of pumped storage power station based on support vector machine
By optimizing the support vector machine model and performing feature processing, the problems of efficiency, accuracy, and robustness in predicting structural vibration of pumped storage power stations were solved, enabling real-time prediction and uncertainty quantification of structural vibration response, thus meeting engineering safety requirements.
Patent Information
- Application Number
- CN202511702731.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-19
- Publication Date
- 2026-02-27
AI Technical Summary
Existing technologies for predicting structural vibrations in pumped storage power stations suffer from complex modeling, high computational costs, and difficulty in achieving rapid evaluation and design optimization. Furthermore, existing machine learning methods exhibit poor generalization ability and robustness under high-dimensional input, small sample, and strongly nonlinear scenarios, and the prediction results lack uncertainty quantification, making it difficult to meet engineering safety requirements.
Structural vibration prediction is performed using a support vector machine (SVR) model. Training data is generated through parametric simulation, and model parameters are optimized by combining feature selection and dimensionality reduction. Prediction uncertainty quantification is introduced to construct an efficient and accurate prediction model.
It achieves near real-time prediction of structural vibration response, improves the accuracy and robustness of prediction, provides prediction confidence intervals, and meets the needs of engineering safety margin management.
Smart Images
Figure CN121581285A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application generally relates to the technical field of hydraulic structure dynamics and engineering calculation. More specifically, the present application relates to a pumped storage power station structure vibration prediction method and system based on support vector machine. BACKGROUND
[0002] Pumped storage power stations need to be frequently started and stopped, load adjusted and working conditions switched to ensure the stability of the power grid. These transition processes will cause severe hydraulic transients and pressure pulsations in the pressure pipeline, and the hydraulic excitation is the main source of inducing structure vibration of the "pipeline-pow erhouse-mountain" coupling system, which poses a threat to the long-term safety of the project.
[0003] Currently, the evaluation of the structure vibration mainly relies on numerical simulation methods such as transient finite element (FEM) or fluid-structure interaction (FSI). However, such methods have inherent defects of complex modeling and high computational cost, which makes them difficult to perform rapid evaluation and design optimization tasks under multiple working conditions, and are inefficient.
[0004] To replace time-consuming simulation, existing research attempts to introduce machine learning to build a proxy model. However, in the complex scenario of pumped storage power station structure vibration prediction with high-dimensional input (load, material, geometric parameters), small sample (high simulation cost) and strong nonlinearity, existing machine learning methods still face the following serious challenges: first, poor generalization ability and robustness: under limited simulation samples, the model is difficult to learn high-dimensional nonlinear mapping, and the extrapolation prediction is unstable due to insufficient sample coverage. Second, imperfect model construction: there are generally problems such as unsystematic feature engineering and insufficient hyperparameter optimization, which limit the prediction accuracy of the model. Third, lack of uncertainty quantification: the prediction result is usually a deterministic value, which cannot provide a confidence interval, and it is difficult to meet the actual needs of engineering safety margin management.
[0005] Therefore, it is urgent to provide a pumped storage power station structure vibration prediction scheme based on support vector machine to solve the above problems and achieve an effective balance between efficiency, accuracy and robustness. SUMMARY
[0006] To at least solve one or more of the above-mentioned technical problems, the present application proposes a pumped storage power station structure vibration prediction scheme based on support vector machine in multiple aspects.
[0007] In a first aspect, the application provides a support vector machine-based pumped storage power station structure vibration prediction method, comprising: obtaining training samples containing input parameters and structure vibration responses by parameterizing simulation of a pipeline-pow erhouse-mountain coupling model of a pumped storage power station; constructing input feature vectors and structure vibration response target values based on the training samples, and performing standardization processing on the input feature vectors and structure vibration response target values; training an SVR model based on the standardized input feature vectors and structure vibration response target values, and performing optimization processing on the SVR model; receiving input parameters of a working condition to be predicted, performing input feature vector construction and standardization processing thereon, and inputting the same into the optimized SVR model to output a predicted structure vibration response.
[0008] In some embodiments, the input parameters include load characteristics, material characteristics and geometric characteristics; the load characteristics include pressure peak value, main frequency and attenuation coefficient; the material characteristics include elastic modulus, Poisson's ratio and damping ratio; and the geometric characteristics include length of a penstock, branch radius, powerhouse burial depth and cover thickness.
[0009] In some embodiments, the structure vibration response target values are statistical quantities extracted from at least one of a tailrace pipe outlet, a powerhouse floor and a mountain amplification area, wherein the statistical quantities include one or more of maximum acceleration, mean value, standard deviation and a preset quantile.
[0010] In some embodiments, during the standardization processing of the input feature vectors, feature selection or feature dimension reduction is performed. In the feature selection process, the first M input feature vectors with the highest contribution to the structure vibration response target values are screened by mutual information, or input feature vectors with a contribution less than a contribution threshold to the structure vibration response target values are iteratively eliminated by recursive feature elimination; in the feature dimension reduction process, the input feature vectors are reduced in dimension by principal component analysis, and the cumulative variance contained in the reduced input feature vectors accounts for at least a preset percentage of the total variance contained in the original input feature vectors.
[0011] In some embodiments, the SVR model adopts a radial basis kernel function; and in the optimization processing of the SVR model, the penalty factor, kernel parameter and insensitive coefficient of the SVR model are optimized to optimize the SVR model.
[0012] In some embodiments, in the process of optimizing the penalty factor, the kernel parameter and the insensitive coefficient of the SVR model to optimize the SVR model, the following steps are performed: a wide range of grid search is performed on the parameter combination of the penalty factor, the kernel parameter and the insensitive coefficient to obtain an initial evaluation sample and locate an effective optimization interval; the next parameter combination to be evaluated is selected in the effective optimization interval based on a Gaussian process regression proxy model; the selected parameter combination is evaluated by stratified k-fold cross-validation to obtain a performance indicator of the selected parameter combination; the current optimal parameter combination and its confidence interval are updated based on the performance indicator of the selected parameter combination; it is determined whether the optimization stopping condition is reached; in response to reaching the optimization stopping condition, the current optimal parameter combination is taken as the optimal parameter combination of the SVR model; in response to not reaching the optimization stopping condition, the step of selecting the next parameter combination to be evaluated in the effective optimization interval based on the Gaussian process regression proxy model is returned to.
[0013] In some embodiments, in the process of training the SVR model, the prediction uncertainty of the SVR model is evaluated.
[0014] In some embodiments, in the process of evaluating the prediction uncertainty of the SVR model, the corresponding prediction confidence interval and median prediction value are output by training the corresponding SVR model at the preset quantile point respectively.
[0015] In some embodiments, in the process of evaluating the prediction uncertainty of the SVR model, the following steps are performed: k SVR sub-models are trained by k-fold cross-validation; k prediction results are obtained by inputting the corresponding standardized input feature vectors into the k SVR sub-models respectively; the variance of the k prediction results is calculated, and the prediction confidence interval is obtained based on the variance of the k prediction results.
[0016] In a second aspect, the present application provides a pumped storage power station structure vibration prediction system based on a support vector machine, which performs pumped storage power station structure vibration prediction based on a support vector machine by using the pumped storage power station structure vibration prediction method based on a support vector machine of any one of the embodiments of the first aspect. The system comprises: a data generation module for obtaining training samples containing input parameters and structure vibration responses by parameterizing simulation of a pipeline-powerhouse-mountain coupling model of a pumped storage power station; a feature processing module for constructing input feature vectors and structure vibration response target values based on the training samples, and performing standardization processing on the input feature vectors and structure vibration response target values; a model training module for training an SVR model based on the standardized input feature vectors and structure vibration response target values, and performing optimization processing on the SVR model; and a prediction interface module for receiving input parameters of a working condition to be predicted, performing input feature vector construction and standardization processing thereon, and inputting the same into the optimized SVR model to output a predicted structure vibration response.
[0017] Through the pumped storage power station structure vibration prediction scheme based on the support vector machine provided above, the embodiment of the application uses the SVR model to replace the traditional time-consuming pumped storage power station coupling system simulation calculation, shortens the originally time-consuming complex simulation process to seconds, and realizes the near real-time prediction of the structure vibration response. The training data is generated through systematic parameterized simulation, effectively overcoming the problem of sparse field measurement data and incomplete working condition coverage. By combining the powerful nonlinear problem processing capability of support vector regression (SVR) and the model optimization step, it is ensured that the model can accurately capture the complex relationship between the input parameters and the vibration response, and the accuracy and generalization ability of the prediction result are guaranteed.
[0018] Further, in some embodiments, the input data is refined and dimensionally reduced before model training, the purpose of which is to filter out the core features that are most critical and most informative to predict the structure vibration from the original large number of input parameters. First, by eliminating redundant or irrelevant features, the computational complexity of the model can be greatly reduced, and the efficiency of training and prediction can be significantly improved. Second, it can effectively avoid the curse of dimensionality and reduce the risk of overfitting, because the model will focus on learning the real physical law that works, rather than the noise in the data, thereby enhancing the robustness and generalization ability of the model, making it more reliable when facing new unknown working conditions. Finally, by providing multiple alternative solutions such as mutual information, RFE or PCA, this method has flexibility in handling different data features, ensuring that an efficient and accurate prediction model can always be built in the best way.
[0019] Further, in some embodiments, the parameter configuration that allows the model to achieve the optimal prediction performance is obtained. First, the wide range grid search is combined with the Bayesian optimization based on Gaussian process to predict and guide the search direction, so as to converge to the optimal solution faster with fewer evaluation times. Second, when evaluating each candidate parameter combination, this method uses stratified k-fold cross-validation, which is a recognized evaluation standard that can effectively avoid randomness and overfitting, ensuring that each step of the optimization process is robust, and the optimal parameter combination selected finally has strong generalization ability. Ensures that the SVR prediction model finally built has the best penalty factor, kernel parameter and insensitivity coefficient, thereby maximizing the accuracy and reliability of the prediction of the structure vibration response.
[0020] Further, in some embodiments, the prediction uncertainty quantification is introduced, which greatly improves the reliability and practical value of the prediction model. First, it provides a key basis for decision-making, which no longer gives a single, possibly biased prediction point, but provides a prediction interval with a clear probability. This allows engineers and decision-makers to understand the potential fluctuation range of the prediction result, thereby making more reliable risk assessment and safety redundancy design. Second, it enhances the credibility of the model. When the prediction interval is wide, it can alert the user that the prediction result under this working condition is highly uncertain and should be used with caution. BRIEF DESCRIPTION OF DRAWINGS
[0021] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which a number of embodiments of the application are shown by way of example, and in which like reference numerals refer to like elements throughout. In the drawings: Figure 1 An exemplary flowchart of a support vector machine-based pumped storage power station structural vibration prediction method according to an embodiment of the present application is shown; Figure 2 An exemplary flowchart of a process of evaluating prediction uncertainty of an SVR model according to an embodiment of the present application is shown; Figure 3 An exemplary flowchart of optimizing an SVR model according to an embodiment of the present application is shown; Figure 4 An exemplary structural block diagram of a support vector machine-based pumped storage power station structural vibration prediction system according to an embodiment of the present application is shown. DETAILED DESCRIPTION
[0022] The technical solutions in the embodiments of the present application will be apparently and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.
[0023] It should be understood that the terms “include” and “contain” used in the specification and claims of the present application indicate the presence of the described features, whole, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, whole, steps, operations, elements, components and / or sets thereof.
[0024] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.
[0025] Figure 1 An exemplary flowchart of a support vector machine-based method for predicting structural vibration of a pumped storage power station according to an embodiment of this application is shown.
[0026] like Figure 1 As shown, in step S110, a training sample containing input parameters and structural vibration response is obtained by parametric simulation of the pipeline-powerhouse-mountain coupling model of the pumped storage power station.
[0027] In the embodiments of this application, initial samples are generated in batches by transient simulation of the pipeline-powerhouse-mountain coupling model of the pumped storage power station. Input parameters are sampled within a preset range, and structural vibration responses at key locations are extracted to obtain training samples containing input parameters and structural vibration responses.
[0028] In some embodiments of this application, Latin hypercube sampling (LHS) is used to sample the input parameters within a preset range. In other embodiments of this application, other methods may be used to sample the input parameters, and this application does not impose any limitations on this method.
[0029] In the embodiments of this application, the input parameters include load characteristics, material characteristics, and geometric characteristics.
[0030] Specifically, load characteristics include peak pressure, dominant frequency, and attenuation coefficient.
[0031] Specifically, material characteristics include elastic modulus, Poisson's ratio, damping ratio, etc.
[0032] Specifically, geometric features include the length of the water pipe, the branch radius, the burial depth of the plant, the thickness of the cover, and the configuration of measuring points.
[0033] The following section, in conjunction with Table 1, provides a detailed explanation of the value range and meaning of the above input parameters.
[0034] Table 1. Input Parameter Range and Meaning Parameter Value range / Explanation Pressure peak 0.4-1.4 MPa; derived from upper and lower limits of hydraulic transient condition Dominant frequency 1-25 Hz; can be set according to working condition Attenuation coefficient 0.01-0.20; exponential attenuation approximation Elastic modulus 20-60 GPa; material grade related to lithology Poisson's ratio 0.18-0.28; interval based on engineering experience Damping ratio 1%-5%; equivalent damping of structure Length of penstock 500-2500 m; including straight section and bend Branch radius 1.5-4.5 m; typical value range Plant buried depth 60-300 m; related to cover thickness Cover thickness 80-500 m; constrained by stratum condition Configuration of measuring points Outlet of tailrace pipe, center of floor, mountain amplification area, etc. In the embodiments of the present application, the corresponding statistical quantities extracted from at least one of the tail water pipe outlet, the plant floor and the mountain amplification area are taken as the structural vibration response target values.
[0035] In the embodiments of the present application, the statistical quantities include one or more of the maximum acceleration, the mean value, the standard deviation and the preset quantile.
[0036] In some embodiments of the present application, the preset quantile is the 95% quantile. In other embodiments of the present application, the preset quantile can also be other quantiles, which are not limited in the present application.
[0037] By systematically considering multiple key physical parameters covering loads, materials and geometric characteristics, and extracting rich statistical quantities including maximum values, mean values, quantiles and the like from multiple core positions such as tail water pipes, plants and mountains as response targets, it is ensured that the training data not only has rich dimensions, but also highly matches the real-world physical laws.
[0038] After step S110 is performed, in step S120, the input feature vectors and the structural vibration response target values are constructed based on the training samples, and the input feature vectors and the structural vibration response target values are standardized.
[0039] In the embodiments of the present application, feature selection or feature dimension reduction is performed in the process of standardizing the input feature vectors.
[0040] Specifically, in the process of feature selection, the first M input feature vectors with the highest contribution to the structural vibration response target values are selected by mutual information, or input feature vectors with a contribution less than a contribution threshold to the structural vibration response target values are iteratively eliminated by recursive feature elimination (RFE).
[0041] Through feature selection, redundant or irrelevant features are eliminated, and the input dimension of the model is reduced, which will directly reduce the time and computing resources required for model training and subsequent prediction, making the entire prediction process more lightweight and fast.
[0042] Specifically, in the process of feature dimension reduction, the input feature vectors are dimensionally reduced by principal component analysis (PCA), and the cumulative variance contained in the dimensionally reduced input feature vectors accounts for at least a preset percentage of the total variance contained in the original input feature vectors.
[0043] Through feature dimension reduction, the curse of dimensionality problem can be effectively avoided, and the risk of overfitting of the model can be significantly reduced. Because the model will focus on learning the real physical relationship between key features and vibration responses, rather than being misled by noise or accidental correlations in the data, it is ensured that the model can still make accurate and reliable predictions when facing new, unseen working conditions.
[0044] In some embodiments of the present application, the preset percentage is 95%. In other embodiments of the present application, the preset percentage can also be set according to actual needs and historical experience, which is not limited in the present application.
[0045] After step S120 is performed, in step S130, the SVR model is trained based on the standardized input feature vector and the structural vibration response target value, and the SVR model is optimized.
[0046] In an embodiment of the present application, the SVR model adopts a radial basis kernel function.
[0047] In an embodiment of the present application, in the process of training the SVR model, the data set composed of the standardized input feature vector and the structural vibration response target value is divided into a training set, a validation set and a test set according to 6:2:2, and the working condition distribution is kept consistent.
[0048] Specifically, the performance of the SVR model in the training process is verified by the mean absolute error (MAE), the root mean square error (RMSE) and the determination coefficient (R 2 ) of the SVR model on the validation set.
[0049] In an embodiment of the present application, in the process of training the SVR model, the prediction uncertainty of the SVR model is evaluated to improve the reliability of the prediction result.
[0050] In an embodiment of the present application, in the process of evaluating the prediction uncertainty of the SVR model, the corresponding SVR model is trained at each preset quantile point to output the corresponding prediction confidence interval and the median prediction value.
[0051] In some embodiments of the present application, the corresponding SVR model is trained at τ=0.05, 0.50 and 0.95 quantile points to directly output the prediction confidence interval and the median prediction . In other embodiments of the present application, other quantile points can also be taken according to actual needs and historical experience, which is not limited in the present application.
[0052] By directly training the model at a specific quantile point (such as 5% and 95%), the confidence interval and the median of the prediction result can be directly and accurately output. This provides the most intuitive and most reliable upper and lower limit boundary for risk assessment, and the result is strong in interpretability.
[0053] In another embodiment of the present application, in the process of evaluating the prediction uncertainty of the SVR model, the specific process can be referred to in Figure 2 .
[0054] Figure 2 An exemplary flowchart of a process of evaluating the prediction uncertainty of the SVR model according to an embodiment of the present application is shown.
[0055] As shown in FIG. 2, in step S210, k SVR sub-models are trained by k-fold cross-validation. In step S220, k prediction results are obtained by inputting the corresponding standardized input feature vectors into the k SVR sub-models respectively. In step S230, the variance of the k prediction results is calculated, and the prediction confidence interval is obtained based on the variance of the k prediction results. Figure 2
[0056] The uncertainty is estimated by comparing the difference (variance) of the prediction results of the k sub-models generated in the model training process. This obtains the quantitative index of the uncertainty while completing the model verification, which is extremely efficient and practical.
[0057] In the embodiments of the present application, during the optimization processing of the SVR model, the SVR model is optimized by optimizing the penalty factor, the kernel parameter and the insensitive coefficient of the SVR model.
[0058] In the embodiments of the present application, the specific process of optimizing the penalty factor, the kernel parameter and the insensitive coefficient of the SVR model to optimize the SVR model can be referred to in the description of Figure 3 .
[0059] Figure 3 An exemplary flowchart of an optimization process of the SVR model according to an embodiment of the present application is shown.
[0060] As shown in FIG. 3, in step S310, a wide-range grid search is performed on the parameter combination composed of the penalty factor, the kernel parameter and the insensitive coefficient to perform preliminary exploration, to obtain an initial evaluation sample and locate an effective optimization interval. In step S320, the next parameter combination to be evaluated in the effective optimization interval is selected based on the Gaussian process regression surrogate model. In step S330, the selected parameter combination is evaluated by hierarchical k-fold cross-validation to obtain the performance index of the selected parameter combination. In step S340, the current optimal parameter combination and its confidence interval are updated based on the performance index of the selected parameter combination. In step S350, it is determined whether the optimization stop condition is reached. In response to reaching the optimization stop condition, in step S360, the current optimal parameter combination is taken as the optimal parameter combination of the SVR model. In response to not reaching the optimization stop condition, it is returned to step S320, i.e., it is returned to the step of selecting the next parameter combination to be evaluated in the effective optimization interval based on the Gaussian process regression surrogate model. Figure 3
[0061] Based on step S310, preliminary exploration is carried out through extensive grid search instead of blind start. This is not the inefficient repetition of traditional grid search, but a warm-up link of Bayesian optimization, which is used to quickly obtain preliminary evaluation samples and quickly frame an effective optimization interval. This greatly reduces the range of subsequent refined search and avoids the waste of a large amount of computing resources in the Bayesian optimization in a too wide unknown space, thereby significantly improving the overall efficiency.
[0062] In the embodiments of the present application, in the process of selecting the next parameter combination to be evaluated in the effective optimization interval based on the Gaussian process regression surrogate model, the Gaussian process regression surrogate model is used to carry out Bayesian optimization in the effective optimization interval, and the cross-validation mean and variance weighted acquisition function is used to sample the parameter combination to be evaluated.
[0063] In the effective interval of positioning, the Bayesian optimization strategy combining the Gaussian process regression surrogate model and the cross-validation mean and variance weighted acquisition function is adopted. This makes each selection of parameter combination highly intelligent, meaning that it neither only turns in the area with known good performance (avoiding local optimum) nor actively explores unknown areas with high model prediction uncertainty but may contain higher performance, thereby finding the globally optimal hyperparameter combination to the greatest extent.
[0064] In the embodiments of the present application, in the process of evaluating the selected parameter combination by stratified k-fold cross-validation, k=5-fold cross-validation is carried out in the range of C∈[10 −1 ,10 3 ]、γ∈[10 −4 ,10 −1 ]、ε∈[10 −4 ,10 −1 ], wherein C is a penalty factor, γ is a kernel parameter, and ε is an insensitive coefficient.
[0065] By using stratified k-fold cross-validation to evaluate the performance of the selected parameter combination each time, the data set is divided into k parts, and training and verification are carried out in turn. This ensures that the sample distribution of each verification set remains consistent with the original data set, avoids evaluation deviation caused by the randomness of data division, and makes the obtained performance indicators (such as cross-validation mean and variance) highly reliable, thereby ensuring that the finally selected hyperparameter combination has good generalization ability in actual application.
[0066] In the embodiments of the present application, the optimization stop condition is that the number of iterations reaches a threshold or the performance indicator of the selected parameter combination reaches a preset value.
[0067] By setting the optimization stop condition, the optimization process can be terminated in time after obtaining satisfactory results, avoiding unnecessary waste of computing resources, making the entire tuning process more efficient and controllable.
[0068] In the embodiments of the present application, after obtaining the optimal parameter combination, the optimized SVR model and the corresponding optimal parameter combination are stored persistently, ensuring that the same verified optimized model and parameter combination are used for prediction each time. This eliminates the fluctuations in prediction results caused by inconsistent model or parameter configurations, ensuring the stability and high reliability of the prediction results, which is crucial for the operation decision of pumped storage power stations that require high precision and reliable prediction.
[0069] After step S130 is executed, in step S140, the input parameters of the working condition to be predicted are received, and after input feature vector construction and standardization processing, they are input into the optimized SVR model to output the predicted structural vibration response.
[0070] In the embodiments of the present application, during the process of receiving the input parameters of the working condition to be predicted, a command line interface (CLI) and / or an application programming interface (HTTP API) are provided to receive new working condition input parameters and return the predicted structural vibration response. The above interface supports batch prediction function and can export the prediction results in CSV format or JSON format.
[0071] Through the command line interface (CLI) or the application programming interface (HTTP API), the accessibility and integration flexibility of the model are greatly improved. Whether it is the need for automated script processing, local rapid testing (CLI), or the requirement for cross-platform, remote calling of Web service integration (HTTP API), the most suitable interface method can be found to meet the diverse business needs and technical environment. On this basis, the batch prediction function significantly improves the efficiency and throughput of processing a large amount of new working condition data, effectively reducing communication overhead and time-consuming of repeated model loading. At the same time, the prediction results can be exported in CSV or JSON format, ensuring the high interoperability and ease of use of the data, so that the prediction results can be seamlessly consumed, analyzed and visualized by various downstream systems (such as report tools, data analysis platforms), greatly facilitating the user's access and utilization of the structural vibration response.
[0072] In summary, the application provides a pumped storage power station structure vibration prediction system based on a support vector machine. The system adopts the pumped storage power station structure vibration prediction method based on a support vector machine as described in any one of the first aspect embodiments to perform pumped storage power station structure vibration prediction based on a support vector machine. The system comprises: a data generation module configured to obtain training samples containing input parameters and structure vibration responses by parameterizing simulation of a pipeline-pow erhouse-mountain coupling model of a pumped storage power station; a feature processing module configured to construct input feature vectors and structure vibration response target values based on the training samples, and perform standardization processing on the input feature vectors and structure vibration response target values; a model training module configured to train an SVR model based on the standardized input feature vectors and structure vibration response target values, and perform optimization processing on the SVR model; and a prediction interface module configured to receive input parameters of a working condition to be predicted, perform input feature vector construction and standardization processing on the input parameters, and input the input parameters into the optimized SVR model to output a predicted structure vibration response.
[0073] Through the pumped storage power station structure vibration prediction scheme based on a support vector machine as provided above, the embodiments of the application construct an SVR model to replace the traditional time-consuming pumped storage power station coupling system simulation calculation, shorten the originally time-consuming complex simulation process to seconds, and achieve near real-time prediction of the structure vibration response. The training data is generated through systematic parameterization simulation, effectively overcoming the problem of sparse field measurement data and incomplete working condition coverage. By combining the powerful nonlinear problem processing capability of support vector regression (SVR) and the model optimization step, it is ensured that the model can accurately capture the complex relationship between the input parameters and the vibration response, and the accuracy and generalization ability of the prediction results are guaranteed.
[0074] Further, in some embodiments, the input data is refined and reduced in dimension before model training, which aims to filter out the core features that are most critical and have the most information for predicting structure vibration from the original large number of input parameters. First, by removing redundant or irrelevant features, the computational complexity of the model can be greatly reduced, and the efficiency of training and prediction can be significantly improved. Second, it can effectively avoid the curse of dimensionality and reduce the risk of overfitting, because the model will focus on learning the real physical laws that work, rather than the noise in the data, thereby enhancing the robustness and generalization ability of the model, making it more reliable when facing new unknown working conditions. Finally, by providing multiple alternative solutions such as mutual information, RFE or PCA, this method has flexibility in handling different data features, ensuring that an efficient and accurate prediction model can always be constructed in the best way.
[0075] Further, in some embodiments, the parameters that can make the model prediction performance optimal are obtained. Firstly, the wide range grid search is combined with the Gaussian process-based Bayesian optimization to predict and guide the search direction, so as to converge to the optimal solution faster with fewer evaluation times. Secondly, when each candidate parameter combination is evaluated, the method uses stratified k-fold cross-validation, which is a recognized and effective evaluation standard that can effectively avoid contingency and overfitting, ensuring that each step of the optimization process is robust, and the finally selected optimal parameter combination has strong generalization ability. It is ensured that the finally constructed SVR prediction model has the best penalty factor, kernel parameter and insensitivity coefficient, thereby maximizing the accuracy and reliability of the prediction of the structural vibration response.
[0076] Further, in some embodiments, prediction uncertainty quantification is introduced, which greatly improves the reliability and practical value of the prediction model. First, it provides a key basis for decision-making, which no longer gives a single, possibly biased prediction point, but provides a prediction interval with a clear probability. This allows engineers and decision-makers to understand the potential fluctuation range of the prediction result, thereby making more reliable risk assessment and safety redundancy design. Second, it enhances the credibility of the model. By quantifying the uncertainty, when the prediction interval is wide, it can alert users that the prediction result under this working condition has high uncertainty and should be used with caution.
[0077] The embodiments of the present application also provide a support vector machine-based structural vibration prediction system of a pumped storage power station, which can use the support vector machine-based structural vibration prediction method 100 of the pumped storage power station described above to perform support vector machine-based structural vibration prediction of the pumped storage power station, or use other methods to perform support vector machine-based structural vibration prediction of the pumped storage power station, which is not limited herein.
[0078] Figure 4 An exemplary structural block diagram of the support vector machine-based structural vibration prediction system of the pumped storage power station according to the embodiments of the present application is shown.
[0079] As shown in Figure 4 The system 400 includes a data generation module 410, a feature processing module 420, a model training module 430 and a prediction interface module 440.
[0080] Specifically, the data generation module 410 is configured to obtain training samples containing input parameters and structural vibration responses by parameterizing simulation of a pipeline-power house-mountain coupling model of the pumped storage power station.
[0081] Specifically, the feature processing module 420 is configured to construct input feature vectors and structural vibration response target values based on the training samples, and perform standardization processing on the input feature vectors and the structural vibration response target values.
[0082] Specifically, the model training module 430 is configured to train the SVR model based on the standardized input feature vector and the structural vibration response target value, and to perform optimization processing on the SVR model.
[0083] Specifically, the prediction interface module 440 is configured to receive input parameters of a working condition to be predicted, perform input feature vector construction and standardization processing on the input parameters, and input the processed parameters into the SVR model after optimization processing, and output a predicted structural vibration response.
[0084] When the system 400 performs the support vector machine-based structural vibration prediction of the pumped storage power station according to the support vector machine-based structural vibration prediction method 100 of the pumped storage power station described above, the data generation module 410 performs the step S110 described above, the feature processing module 420 performs the step S120 described above, the model training module 430 performs the step S130, and the prediction interface module 440 performs the step S140. The specific execution process can be referred to the foregoing, and will not be described here again.
[0085] Although the embodiments of the present application have been shown and described herein, it should be apparent to those skilled in the art that such embodiments are merely illustrative of the present application. Many changes, modifications and alternative ways of implementing the present application can be suggested to one skilled in the art and it is intended to encompass such changes, modifications and alternative ways within the scope of the present application. It should be appreciated that various alternatives to the embodiments of the present application described herein can be employed in practicing the present application. The appended claims are intended to cover such alternatives and equivalents.
Claims
1. A pumped storage power plant structure vibration prediction method based on a support vector machine, characterized by, The method comprises the following steps: obtaining training samples containing input parameters and structural vibration responses by parameterized simulation on a pipeline-powerhouse-mountain coupling model of a pumped storage power station; constructing input feature vectors and structural vibration response target values based on the training samples, and performing standardization processing on the input feature vectors and structural vibration response target values; training an SVR model based on the standardized input feature vectors and structural vibration response target values, and performing optimization processing on the SVR model; receiving input parameters of a working condition to be predicted, performing input feature vector construction and standardization processing on the input parameters, inputting the input parameters into the optimized SVR model, and outputting a predicted structural vibration response.
2. The support vector machine-based pumped storage power station structural vibration prediction method according to claim 1, characterized by, The input parameters include load characteristics, material characteristics, and geometric characteristics. The load characteristics include pressure peak value, main frequency, and attenuation coefficient. The material characteristics include elastic modulus, Poisson's ratio, and damping ratio. The geometric characteristics include water diversion pipe length, branch radius, powerhouse burial depth, and cover thickness.
3. The support vector machine-based pumped storage power station structural vibration prediction method according to claim 1 or 2, characterized by, The structural vibration response target values are obtained by extracting corresponding statistics from at least one of the tail water pipe outlet, the powerhouse floor, and the mountain amplification area, wherein the statistics include one or more of maximum acceleration, mean value, standard deviation, and preset quantile. 4.The support vector machine-based pumped storage power station structural vibration prediction method according to claim 1, characterized in that, During the standardization processing of the input feature vectors, feature selection or feature dimension reduction is performed. During the feature selection, the first M input feature vectors with the highest contribution to the structural vibration response target values are selected by mutual information, or input feature vectors with a contribution less than a contribution threshold to the structural vibration response target values are iteratively eliminated by recursive feature elimination. During the feature dimension reduction, principal component analysis is performed on the input feature vectors, and the cumulative variance of the dimension-reduced input feature vectors accounts for at least a preset percentage of the total variance of the original input feature vectors. 5.The support vector machine-based pumped storage power station structural vibration prediction method according to claim 1, wherein, The SVR model uses a radial basis kernel function. During the optimization processing of the SVR model, the penalty factor, kernel parameter, and insensitive coefficient of the SVR model are optimized to optimize the SVR model. 6.The support vector machine-based pumped storage power station structural vibration prediction method according to claim 5, characterized in that, During the optimization of the penalty factor, kernel parameter, and insensitive coefficient of the SVR model, the following steps are performed: a wide-range grid search is performed on the parameter combination of the penalty factor, kernel parameter, and insensitive coefficient to obtain initial evaluation samples and locate an effective optimization interval; the next parameter combination to be evaluated is selected in the effective optimization interval based on a Gaussian process regression proxy model; the selected parameter combination is evaluated by hierarchical k-fold cross-validation to obtain a performance indicator of the selected parameter combination; the current optimal parameter combination and its confidence interval are updated based on the performance indicator of the selected parameter combination; it is determined whether the optimization stopping condition is reached; in response to reaching the optimization stopping condition, the current optimal parameter combination is taken as the optimal parameter combination of the SVR model; in response to not reaching the optimization stopping condition, the step of selecting the next parameter combination to be evaluated in the effective optimization interval based on the Gaussian process regression proxy model is returned to. 7.The support vector machine-based pumped storage power station structural vibration prediction method according to claim 1 or 5, characterized in that, In the process of training the SVR model, the prediction uncertainty of the SVR model is evaluated. 8.The support vector machine-based pumped storage power station structural vibration prediction method according to claim 7, characterized in that, In the process of evaluating the prediction uncertainty of the SVR model, the corresponding prediction confidence interval and median prediction value are output by training the corresponding SVR model at the preset quantile respectively. 9.The support vector machine-based pumped storage power station structural vibration prediction method according to claim 7, characterized in that, In the process of evaluating the prediction uncertainty of the SVR model, the following steps are performed: Train k SVR sub-models through k-fold cross-validation; Obtain k prediction results by inputting the corresponding standardized input feature vectors into the k SVR sub-models respectively; Calculate the variance of the k prediction results, and obtain the prediction confidence interval based on the variance of the k prediction results.
10. A pumped storage power plant structure vibration prediction system based on a support vector machine, characterized by, The support vector machine-based pumped storage power station structure vibration prediction method according to any one of claims 1-9 is used for support vector machine-based pumped storage power station structure vibration prediction, and the system comprises: A data generation module is configured to obtain training samples containing input parameters and structure vibration responses by parameterizing simulation of a pipeline-powerhouse-mountain coupling model of a pumped storage power station; A feature processing module is configured to construct input feature vectors and structure vibration response target values based on the training samples, and to perform standardization processing on the input feature vectors and structure vibration response target values; A model training module is configured to train an SVR model based on the standardized input feature vectors and structure vibration response target values, and to perform optimization processing on the SVR model; A prediction interface module is configured to receive input parameters of a working condition to be predicted, to perform input feature vector construction and standardization processing on the input parameters, and to input the input parameters into the optimized SVR model to output predicted structure vibration responses.