Pumped storage power station underground powerhouse side wall deformation early warning credibility analysis method and device

Abstract

The application discloses a credibility analysis method and device for deformation early warning of a side wall of a pumped storage power station underground powerhouse, which uses deformation displacement monitoring data of the side wall of the pumped storage power station underground powerhouse as a training sample to train an early warning model, inputs real-time monitoring displacement data to obtain predicted displacement data, and then combines the average prediction variance e of the training sample set to adopt a credibility analysis index of fitting degree. A collection unit uses a high-precision three-way displacement sensor to collect three-direction displacement change data of the side wall of the pumped storage power station underground powerhouse in real time. A transmission unit is used to send the collected data to a processing unit. The processing unit is used to apply the trained early warning model to the displacement change data to predict the next moment, and combine the predicted deformation displacement data with the average prediction variance of the training sample set to perform fitting degree analysis. An evaluation unit is used to evaluate the credibility of the early warning according to the fitting degree analysis result. The application is favorable for improving the accuracy of the deformation early warning of the side wall of the pumped storage power station underground powerhouse.

Description

Technical Field

[0001] This invention belongs to the field of engineering geological disaster early warning technology, and relates to a reliability analysis method and device for early warning of deformation of the sidewall of the underground powerhouse of a pumped storage power station. Background Technology

[0002] The underground powerhouse of a pumped-storage power station is its "heart," resembling a vast underground palace. It houses reversible turbines, generator units, and various auxiliary equipment, serving as a complex of hydraulic structures, mechanical and electrical equipment, and the operational area for personnel. Pumped-storage power station underground powerhouse caverns are characterized by large spans, high sidewalls, numerous cavern intersections, and complex underground cavern structures. Excavation of these caverns is highly susceptible to surrounding rock deformation, posing a constant risk of collapse during construction and operation. Therefore, strengthening the control of sidewall deformation is crucial for the safety of the pumped-storage power station's underground powerhouse.

[0003] Displacement is an important observation indicator in the deformation process of the sidewall of the underground powerhouse of pumped storage power station, and the prediction of sidewall deformation displacement is of great significance in engineering. By studying the laws contained in the monitored data of the deformation displacement of the sidewall of the underground powerhouse of pumped storage power station, and using displacement time series analysis method to predict the future deformation displacement of the sidewall of the underground powerhouse of pumped storage power station, it is one of the important ways to predict the deformation of the sidewall of the underground powerhouse of pumped storage power station. At present, the methods of using displacement time series analysis method to predict the deformation of the sidewall of the underground powerhouse of pumped storage power station include multiple regression, ARMA, grey system theory, artificial neural network (ANN), support vector machine (SVM), etc. However, the existing research on the prediction results of the deformation of the sidewall of the underground powerhouse of pumped storage power station still has some shortcomings in terms of credibility analysis, mainly due to (1) the defects of the established prediction model, including the difficulty in determining the parameters and the inaccuracy of the prediction results. (2) the indicators selected for credibility analysis are not representative. (3) the credibility prediction needs to be determined based on the measured value, and the measured value is a lagging "future" value, so the credibility analysis is uncertain. (4) the equipment used is not convenient and fast enough. Summary of the Invention

[0004] The reliability analysis method and device for early warning of deformation of underground powerhouse sidewalls based on information vector machine adopted in this invention can effectively solve the above problems.

[0005] The Informative Vector Machine (IVM) is an approximation algorithm based on the Gaussian process. By employing Bayesian statistical learning theory and kernel methods, IVM offers numerous advantages, including adaptive hyperparameter acquisition, strong adaptability to high-dimensional and complex nonlinear problems, and probabilistically meaningful predicted outputs. Furthermore, IVM preserves the process variance implicit in the kernel function, which can be tracked and, through the selection of active subsets, provides a sparse representation for the model, significantly reducing the time and space complexity of learning. Compared to the Gaussian process, the IVM algorithm reduces computational complexity and memory usage without sacrificing accuracy. Using the IVM learning algorithm, a nonlinear mapping relationship between the deformation displacement time series of the underground powerhouse sidewalls in pumped-storage power stations can be effectively established, thereby enabling the prediction of the deformation of the underground powerhouse sidewalls.

[0006] The reliability analysis of information vector machine for early warning of deformation of the sidewalls of underground powerhouses in pumped-storage power stations relies on the goodness of fit R of the predicted values. The goodness of fit R is calculated by combining the prediction results of the early warning model with the average prediction variance e of the training set. A higher goodness of fit R indicates higher reliability. The average prediction variance e is updated in a timely manner; that is, the variance between the measured value and the predicted value is calculated each time a new measured value is obtained, and then the average prediction variance is recalculated. Using the average prediction variance of existing data instead of the variance between the measured and predicted values ​​in the calculation of the goodness of fit R improves the reliability of the early warning indicator because this indicator does not depend on "future" measured values, resulting in higher certainty in the analysis results.

[0007] The information vector machine reliability analysis device for early warning of deformation of the underground powerhouse sidewalls in pumped-storage power stations serves as the carrier of the aforementioned early warning model. This device utilizes high-precision triaxial displacement sensors to ensure the accuracy of the displacement data. Simultaneously, the use of a wireless bridge for data transmission not only facilitates convenient and quick equipment deployment but also offers long-distance and high-speed transmission.

[0008] This invention provides a new method and apparatus for improving the accuracy of early warning of deformation of the sidewalls of underground powerhouses in pumped storage power stations. It solves, to a certain extent, the practical engineering problems of predicting deformation of the sidewalls of underground powerhouses and provides a reference for evaluating the overall stability of underground caverns and selecting support measures.

[0009] The purpose of this invention is to provide an information vector machine reliability analysis method and device for early warning of deformation of the sidewalls of the underground powerhouse in pumped storage power stations, which is caused by deformation and damage of the rock mass between underground caverns due to excavation of underground cavern groups.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] The first part of this invention provides a reliability analysis method for early warning of deformation of the sidewalls of the underground powerhouse of a pumped storage power station, comprising the following steps:

[0012] Step 1: Construct an early warning model for the deformation of the sidewalls of the underground powerhouse of a pumped-storage power station based on an information vector machine (IVM), referred to as the early warning model. Historical measured data on the deformation of the sidewalls of the underground powerhouse of the pumped-storage power station are compiled and divided into two sets: a training set and a test set. The training set is used to train the early warning model. The test set is used to test the trained early warning model, and the parameters of the early warning model are modified and adjusted based on the test results.

[0013] Step 2: Establish an early warning model based on Information Vector Machine (IVM). This invention employs the IVM machine learning method, which is similar to the Gaussian process machine learning method. The construction of the Gaussian process model mainly includes the following four steps:

[0014] (1) Define the marginal approximation function

[0015] Based on Bayes' theorem, the joint distribution of the latent variable set f and the output observation y is expressed as:

[0016]

[0017] In the formula, p(y n |f n The model (11) is a noise model that shows the relationship between the latent variable f and the output observation y. Integrating equation (11) yields the marginal likelihood function:

[0018]

[0019] In the formula, B is a diagonal matrix, and its nth diagonal element is β. n ,β n The value is

[0020] (2) Calculate the posterior distribution

[0021] Based on Bayes' theorem and the multivariate Gaussian distribution, the posterior distribution of f can be obtained by combining equations (4) and (5):

[0022]

[0023] A key assumption in Gaussian process models is that the noise model must follow a Gaussian distribution. Information Vector Machines (ADF) construct an approximate function to replace the non-Gaussian posterior distribution to maintain the model's applicability. In the ADF algorithm, the approximation process of the true posterior distribution involves using the training data n in J... i If added to I, the posterior distribution (f) can be updated as follows:

[0024]

[0025] Minimizing the KL divergence using moment matching yields a new approximation q. i (f)=N(f|μ i ,∑ i ), where μ i It is q i The mean vector of (f), ∑ i It is q i The covariance matrices of (f) are updated using the following formulas:

[0026]

[0027] in,

[0028] In summary, the ADF algorithm can be used to approximate any noise model using a Gaussian noise model.

[0029] (3) Kernel parameter θ learning

[0030] Kernel parameter θ is estimated by maximizing the marginal likelihood function of equation (5). The objective function containing θ is as follows:

[0031] θ ML =argmax p(y|X,θ) (9)

[0032] The training process of a Gaussian process involves directly selecting the kernel parameters from the training samples. By determining the kernel parameters, classification prediction samples can be generated.

[0033] (4) Data Prediction

[0034] The optimal kernel parameter θ is calculated. ML Then, based on p(f|X,θ)=N(f|0,K) and the properties of the multivariate Gaussian distribution, the joint distribution of [ff(x)] can be obtained. This is achieved by integrating the joint distribution with respect to f using the posterior distribution, as follows:

[0035]

[0036] in, It is the posterior covariance function;

[0037] μ(x)=k T K -1 ∑By represents the posterior mean function.

[0038] Once the distribution of f(x) is determined, the function value of f(x) at x can be predicted, which is the prediction process.

[0039] For the binary classification problem of Gaussian processes, the classification result is a probability value. Therefore, the cumulative Gaussian function in the sigmoid function can be used for transformation to obtain a probabilistic noise model for Gaussian process classification:

[0040]

[0041] Normalization constant Z i :

[0042]

[0043] Therefore, it approximately approximates q i The parameter g in the update formula for the mean vector and covariance matrix of (f) in With γ in They can be represented as:

[0044]

[0045] Among them, g in G represents i The nth element, γ in Then it represents Γ i The nth diagonal element, then The nth diagonal element υ in for:

[0046]

[0047] The noise model has an approximate Gaussian distribution. m n With β n The calculation formula is as follows:

[0048]

[0049] In summary, the approximate formula for the marginal likelihood function in a classification problem is as follows:

[0050] p(y)≈N(m|0,K+B -1 (18)

[0051] When the number of training samples is N, the computational complexity of calculating the gradient of the marginal likelihood function with respect to θ during the θ learning process is O(N). 3 ), and occupy O(N 2 The memory usage of N is limited. When N is large, computational efficiency is significantly affected, limiting the model's applicability. The Information Vector Machine (IVM) machine learning algorithm approximates a Gaussian process by introducing active subset selection and parameter optimization.

[0052] (1) Selection of activity subsets:

[0053] In the information vector machine, the posterior differential entropy of each data point in J is calculated, and the data point with the maximum posterior differential entropy is selected and added to I. When the i-th information vector is selected for I, the posterior differential entropy of the n-th data point in J is:

[0054]

[0055] For Σ in equation (10) i-1 To reduce memory usage, a sparse table is constructed; therefore, Σ can be obtained by solving the continuous vector product of the original variance matrix Σ0=K. i :

[0056]

[0057] In the formula, M i Let be an i×n matrix, with the k-th row being... n k This represents the k-th information vector included in I. Since ζ i-1,n For covariance matrix Σ i-1 The nth diagonal element in the matrix. Then the posterior covariance matrix ζ. i The update formula is expressed in diagonal form:

[0058]

[0059] The formula for updating the posterior mean output vector is:

[0060]

[0061] In summary, the active subset can be determined using the Information Vector Machine (IVM) active subset selection algorithm:

[0062] 1. Initialization. Set the number of information vectors in the active subset to d; m = 0; let ζ0 = diag(K); μ = 0; the inactive subset is all training sample data J; the active subset I is an empty set, and s0 is an empty matrix.

[0063] 2. When i = 1, iterate through all training samples n ∈ J, and calculate g according to equations (13), (14), and (15). in γ in and υ in ΔH is calculated according to formula (19). in .

[0064] 3. Calculate the data point n corresponding to the maximum a posteriori differential entropy. i =argmax n∈J ΔH in .

[0065] 4. Update m according to equations (16) and (17) n and βn And ζ is calculated using equations (21) and (22). i and μ i .

[0066] 5. Extension To M i-1 Update M again i .

[0067] 6. The nth... i Add one data point to I and remove it from J.

[0068] 7. Repeat steps 1 through 6 until d data points are selected and added to I.

[0069] By following the steps above, we can obtain the activity subset I used for training.

[0070] (2) Parameter optimization

[0071] The Information Vector Machine (IVM) algorithm uses training samples from the active subset to perform a Gaussian approximation instead of all training samples, yielding the marginal likelihood function as follows:

[0072]

[0073] The kernel function parameter θ is contained in K I In the process, the optimal θ can be obtained by maximizing the marginal likelihood function of equation (23) using the scale conjugate gradient method.

[0074] Step 3: Train the early warning model based on information vector machine. Divide the training set of historical measured data into multiple displacement time series, and input them into the initially established early warning model as input vectors. Use k-fold cross-validation (K-CV) to randomly divide the samples into 10 parts (k=10). Select one part as the validation sample and the other 9 parts as the training sample, and perform 10 iterations. Finally, estimate the algorithm using the average accuracy of the 10 validation results. A reliable classification model with optimal denoising parameters based on information vector machine is obtained when a 95% accuracy is achieved. Continuously modify the relevant parameters of the early warning model based on the good fit between the predicted and measured values, and determine the optimal displacement time series length L and the average prediction variance e.

[0075] Historical measured data constitutes the training samples for the training model (x) i ,y i The specific method is as follows, where i = 1, 2, ..., n, and the input sample vector is x. i Let y be the first i terms of the historical measured data. i This corresponds to the measured displacement data at the next moment.

[0076] Step 4: Test the early warning model based on information vector machines. Divide the test set from historical measured data into multiple input vectors of optimal time series length L, and input them into the trained early warning model. Determine the feasibility of the early warning model based on the goodness of fit between the prediction results and the measured results. If the prediction results are unsatisfactory, return to step 3 to retrain the early warning model.

[0077] Step 5: Combine the existing measured data on the deformation of the underground powerhouse sidewall of the pumped storage power station to form a displacement time series sample N of length L, which will be used as the input vector.

[0078] Step 6: Input the input vector into the trained early warning model to obtain the predicted displacement data at time A.

[0079] Step 7: Perform a goodness-of-fit analysis between the predicted displacement data at time A and the average prediction variance of the training set to obtain the goodness-of-fit R and determine the credibility of the warning.

[0080] Step 8: Update the average prediction variance e. Using high-precision triaxial displacement sensors deployed on-site, the deformation displacement x of the underground powerhouse sidewalls of the pumped-storage substation is collected in real time. The variance e1 between the measured and predicted values ​​is calculated, and the average prediction variance e is updated. * .

[0081] Furthermore, the goodness of fit R is derived from the statistical concept of goodness of fit, which is used in statistics to analyze regression analysis results. In step 3 of this invention, the goodness of fit is calculated by combining the predicted value and the average prediction variance e of the sample set. Similarly, the goodness of fit R is a value between 0 and 1. The closer the value of R is to 1, the better the goodness of fit, that is, the higher the reliability of the early warning model's prediction.

[0082]

[0083] In the formula: y i denoted as the predicted value of deformation displacement, and e represents the average prediction variance of the sample set.

[0084] The formula for calculating the variance e1 between the measured and predicted values ​​is as follows:

[0085]

[0086] In the formula: This represents the measured value of the deformation displacement.

[0087] Update the average prediction variance e * The method is as follows: Let the number of sample shifts be n. Then the updated sample mean prediction variance is:

[0088]

[0089] Similarly, as the measured deformation displacement values ​​are updated, the sample mean variance is updated in real time based on the variance between the measured and predicted values.

[0090] Table 1. Reliability Judgment Criteria

[0091]

[0092]

[0093] In the table, a, b, and c are threshold standards determined based on the actual engineering situation.

[0094] The steps for constructing the displacement time series in step 3 are as follows:

[0095] (1): Input the monitored deformation and displacement data of the sidewall of the underground powerhouse of the pumped storage power station and construct the displacement time series sample data N according to the optimal length L.

[0096] (2): Let the historical point be l, and use the first 1 to l measured displacement data as the input vector x.

[0097] (3): Input the input vector into the early warning model to obtain the corresponding (l+1)th predicted displacement data. Then, replace the oldest value of the original input vector with the (l+1)th measured displacement data to form a new input sequence with the same time length, and obtain the (l+2)th predicted displacement data. Then, replace the oldest value of the original input vector with the (l+2)th measured displacement data to form a new input sequence with the same time length. In this way, multiple slope displacement time series samples are constructed in a rolling manner.

[0098] Furthermore, the steps for constructing the early warning model in step 1 are as follows:

[0099] S1: Sample Selection. The historical measured displacement data of the underground powerhouse sidewall deformation of the pumped-storage substation are divided into a training set and a test set. The training set is used to train the early warning model based on the information vector machine. The test set is used to test the trained early warning model, and the parameters of the early warning model are modified and adjusted based on the test results.

[0100] S2: Sample Construction. Based on the measured data of deformation and displacement of the underground powerhouse sidewalls, a displacement time series is constructed. Let the number of historical points be l. The first l-1 measured values ​​are used as the input vector, and the (l+1)th predicted value is used as the output value, thus forming the first sample. Then, the (l+1)th measured value is used as the new measured value to replace the oldest value in the original input vector, forming a new input sequence with the same time length. The (l+2)th predicted value is used as the output value, thus forming the second sample. This process is repeated to construct a series of displacement time series samples.

[0101] S3: Establish a sidewall deformation early warning model. Train the early warning model, set the initial parameters of the early warning model, and use the classic k-fold cross-validation method to complete an information vector machine sidewall deformation early warning model when the accuracy reaches 95% or higher. This establishes a nonlinear mapping relationship between the displacement value at any time and the displacement value at the previous l times, and determines the optimal displacement time series length L.

[0102] S4: Verify the predictive performance of the early warning model using a test set. If the goodness of fit between the predicted and measured values ​​is high, the established early warning model for sidewall deformation is considered to meet the requirements and is feasible for predicting the sidewall deformation of the underground powerhouse of the pumped storage power station; otherwise, adjust the samples and training parameters of the early warning model and retrain the early warning model.

[0103] In its second part, the present invention also provides a reliability analysis device for early warning of deformation of the sidewalls of underground powerhouses in pumped-storage power stations, including a field device and a terminal device. The field device includes a data acquisition unit and a transmission unit, and the terminal device includes a transmission unit, a processing unit, and an evaluation unit.

[0104] The data acquisition unit includes a high-precision three-dimensional displacement sensor, which is installed at the displacement monitoring point of the underground powerhouse sidewall of the pumped storage power station and connected to wireless equipment to collect real-time data on displacement changes in the three directions of the underground powerhouse sidewall.

[0105] The transmission unit includes a wireless data transmission module. This module employs wireless bridge technology, a store-and-forward device that enables LAN interconnection at the link layer. It uses air as the medium for signal propagation, effectively solving the difficulties of wired deployment. The wireless devices are connected to the displacement sensor and processing unit of the acquisition unit to enable data transmission between field devices and terminal devices.

[0106] The processing unit, including a microprocessor and a read / write memory, is used to predict the deformation and displacement of the underground powerhouse sidewall of a pumped-storage substation at a given moment using a trained early warning model, and to perform goodness-of-fit analysis by combining the predicted data with the average prediction variance e. The microprocessor, a central processing unit composed of a single large-scale integrated circuit, is responsible for retrieving and executing script instructions from the read / write memory, as well as controlling other components. The read / write memory stores various program scripts and generated data required during the computation process, including the information vector machine-based classification model and prediction results.

[0107] The evaluation unit includes a confidence calculation algorithm, which is programmed into a read-write memory to calculate the goodness of fit of the predicted value, obtain the goodness of fit R, and evaluate the confidence of the warning based on the magnitude of the goodness of fit R.

[0108] Compared with the prior art, the advantages of the present invention are:

[0109] (1) The credibility analysis method of early warning of deformation of the sidewall of the underground powerhouse of pumped storage power station is to use the existing measured values ​​of sidewall deformation displacement to input the predicted values ​​obtained by the early warning model, and calculate the fitting degree R by combining the average prediction variance e of the training sample set, and then conduct credibility analysis to weaken the correlation between credibility and the "future" measured values, thereby reducing uncertainty and significantly improving the accuracy of early warning.

[0110] (2) The sample data of deformation of the sidewall of the underground powerhouse of pumped storage power station is generally limited and belongs to the small sample problem. Existing machine learning models are prone to "over-learning" (or under-learning) problems for small sample regression problems, and there is also the problem of difficulty in determining the parameters. However, the information vector machine based on the principle of minimizing structural risk has good adaptability to small sample regression problems.

[0111] (3) The reliability analysis device for early warning of deformation of the sidewall of the underground powerhouse of the pumped storage power station proposed in this invention transmits the high-precision triaxial displacement sensor signal monitoring the deformation of the sidewall of the underground powerhouse of the pumped storage power station to the processing unit through the transmission unit using a wireless bridge. Then, the predicted displacement of the sidewall deformation at a certain moment is output by the trained early warning model based on information vector machine, and the reliability analysis is performed based on the fitting degree analysis results. This device can input and output signals in real time, and has the characteristics of being fast and convenient. At the same time, the high-precision triaxial displacement sensor can be widely installed on different sections of the sidewall of the underground powerhouse for monitoring, which has strong applicability.

[0112] (4) The information vector machine reliability analysis device for early warning of deformation of the underground powerhouse sidewall of the pumped storage power station provided by the present invention adopts wireless bridge technology. Compared with traditional wired technology, wireless bridge technology has the advantages of convenient deployment, long transmission distance and fast transmission speed. As long as it is in the wireless network coverage area, the field equipment can easily access the network. At the same time, the terminal equipment can also achieve zero-configuration access, which is very convenient and fast. Attached Figure Description

[0113] Figure 1 A schematic diagram of the underground powerhouse and displacement monitoring equipment layout for an embodiment;

[0114] Figure 2 This is a schematic diagram of the device structure in the embodiment;

[0115] Figure 3 The flowchart of the analysis method in the embodiment is shown below;

[0116] Figure 4 This is a schematic diagram illustrating the construction process of an early warning model based on an information vector machine, as shown in the example.

[0117] Figure 5 This is a schematic diagram illustrating the construction of the displacement time series in an embodiment.

[0118] Figure 6This is a comparison chart of the measured and predicted values ​​for an example. Detailed Implementation

[0119] The specific embodiments of the present invention will be further described and illustrated below with reference to the accompanying drawings and examples. It should be noted that the drawings only show the parts relevant to the present invention and not all results. Furthermore, the specific examples are only for explaining the present invention and not for limiting its scope.

[0120] A schematic diagram of the main and auxiliary powerhouses and busbar tunnels of a certain hydropower station is shown below. Figure 1 As shown, there are two monitoring points on the sidewalls of the main and auxiliary powerhouses. The measured displacement data for monitoring point 1 is shown in the table below. This invention provides a reliability analysis device for early warning of deformation of the sidewalls of underground powerhouses in pumped-storage power stations, such as... Figure 2 As shown.

[0121] This example selects 100 displacement sample data points from a monitoring point on the sidewall of an underground powerhouse in a pumped-storage power station to predict the deformation displacement of the sidewall in the next 5 periods. This demonstrates the information vector machine reliability analysis method for early warning of deformation of the sidewall of an underground powerhouse in a pumped-storage power station proposed in this invention. The specific process is as follows: Figure 3 As shown.

[0122] The process of building an early warning model is as follows Figure 4 As shown.

[0123] Step 1: Compile the historical measured data of the deformation of the sidewalls of the underground powerhouse of a pumped storage power station for the previous 100 periods, and divide it into a training sample set and a test sample set. The training sample set is used to train the early warning model based on the information vector machine. The test sample set is used to test the trained model, and the model parameters are modified and adjusted based on the test results.

[0124] The specific method for constructing the training sample set is as follows:

[0125] The following 100 historical test data are divided into several groups of 10 data points each, arranged in chronological order. Since the k-fold cross-validation method (k=10) is used, these groups of data are randomly divided into 10 parts. One part is selected as the validation sample, and the other 9 parts are selected as the training samples. This process is repeated 10 times. Finally, the algorithm is estimated based on the average accuracy of the 10 validation results. A reliable classification model with optimal denoising parameters based on information vector machine is obtained when the accuracy reaches 95%.

[0126] As shown in Tables 2 and 3, the 10 periods of data from January 20, 2023 to January 29, 2023 are divided into one group, and the data from the first nine periods is taken as x. i , (x i ,y i ), where x i =[x i1 ,x i2 ,xi3 ,x i4 ,x i5 ,x i6 ,x i7 ,x i8 ,x i9 The measured data from the tenth period are used as the output value y. i Then, the data from January 30, 2023, was used to replace the data from January 20, 2023, to form a new sample, and so on.

[0127] Table 2 Monitoring data for monitoring point 1 in the previous 100 periods.

[0128]

[0129]

[0130] Table 3 Monitoring data of monitoring point 1 for the last 5 periods.

[0131] 2023 / 4 / 27 18.1 2023 / 4 / 28 18 2023 / 4 / 29 18.3 2023 / 4 / 30 18.45 2023 / 5 / 1 18.84

[0132] Step 2: Establish an early warning model based on information vector machine. Regarding the IVM parameter settings, firstly, the commonly used and effective Gaussian radial basis function (RBF) is selected as the kernel function of the IVM; secondly, the number of information vectors is set to 160; the parameter θ of the RBF kernel function is adaptively obtained by maximizing the marginal likelihood function during the model training process.

[0133] Step 3: Train the early warning model based on information vector machines. Divide the training set of historical measured data into multiple displacement time series, and input them as input vectors into the initially established early warning model. The method for constructing the displacement time series is as follows: Figure 5 As shown.

[0134] The k-fold cross-validation method (k=10) was used to continuously modify the relevant parameters of the early warning model based on the good fit between the predicted and measured values, and to determine the optimal displacement time series length L and the average prediction variance e.

[0135] Step 4: Test the early warning model based on information vector machines. Divide the test set from historical measured data into multiple input vectors of optimal time series length L, and input them into the trained early warning model. Determine the feasibility of the early warning model based on the goodness of fit between the prediction results and the measured results. If the prediction results are unsatisfactory, return to step 3 to retrain the early warning model.

[0136] Based on the training sample set obtained from the above engineering examples, the optimal displacement time series length L is 15 periods, and the average prediction variance e is 0.9.

[0137] Step 5: Combining the existing measured data on the deformation of the underground powerhouse sidewall of the pumped storage power station, a displacement time series sample N with a length L of 15 periods is formed from the last 15 periods and used as the input vector.

[0138] Step 6: Input the input vector into the trained early warning model to obtain the predicted displacement data for the next 5 periods. The prediction is as follows... Figure 6 As shown.

[0139] Table 4. Reliability Analysis of Prediction Results

[0140]

[0141] Step 7: Perform a goodness-of-fit analysis between the predicted displacement data at time A and the sample average prediction variance to obtain the goodness-of-fit R, and determine the credibility of the early warning. The goodness-of-fit R of the deformation of the underground powerhouse sidewall in the last 5 phases of the project obtained from the early warning model is greater than 0.95, and the credibility level is high, as shown in Table 4.

[0142] Step 8: Update the average prediction variance e. Using high-precision triaxial displacement sensors deployed on-site, the deformation displacement x of the underground powerhouse sidewalls of the pumped-storage substation is collected in real time. The variance e1 between the measured and predicted values ​​is calculated, and the average prediction variance e is updated. * .

[0143] It should be noted that the purpose of disclosing the above examples is to help further understand the present invention. However, those skilled in the art will understand that various obvious changes, readjustments, and substitutions to the present invention will not depart from the scope of protection of the present invention. Therefore, the present invention is not limited to the content disclosed in the examples, and the scope of protection of the present invention is defined by the scope of the claims.

Claims

1. A reliability analysis method for early warning of deformation of the sidewall of the underground powerhouse of a pumped storage power station, characterized in that, Includes the following steps: Step 1: Construct an early warning model for the deformation of the sidewall of the underground powerhouse of a pumped storage power station based on information vector machine, referred to as the early warning model; organize the historical measured data of the deformation of the sidewall of the underground powerhouse of the pumped storage power station and divide it into two sets: training set and test set; The training set is used to train the early warning model; the test set is used to test the trained early warning model and modify and adjust the parameters of the early warning model based on the test results. Step 2: Establish an early warning model based on information vector machines, and preliminarily determine the kernel function of the early warning model based on the characteristics of the sample data; Step 3: Train the early warning model based on information vector machine; divide the training set of historical measured data into multiple displacement time series, and input them into the initially established early warning model as input vectors. Use the k-fold cross-validation method to continuously modify the parameters of the early warning model according to the good fit between the predicted value and the measured value, and determine the optimal displacement time series length L and the average prediction variance e. Step 4: Test the early warning model based on information vector machine; assemble multiple sets of input vectors from the test set of historical measured data according to the optimal time series length L, and input them into the trained early warning model. Determine whether the early warning model is feasible based on the good fit between the prediction results and the measured results; if the prediction results are not ideal, return to step 3 to retrain the early warning model. Step 5: Combine the existing measured data on the deformation of the underground powerhouse sidewalls of the pumped storage power station to form a displacement time series sample N of length L, which serves as the input vector; Step 6: Input the input vector into the trained early warning model to obtain the predicted displacement data at time A; Step 7: Perform a goodness-of-fit analysis between the predicted displacement data at time A and the average prediction variance of the training set to obtain the goodness-of-fit R and determine the credibility of the warning. Step 8: Update the average prediction variance e; use high-precision triaxial displacement sensors deployed on-site to collect real-time deformation displacement x of the underground powerhouse sidewalls of the pumped storage power station, calculate the variance e1 between the measured value and the predicted value, and update the average prediction variance e. * .

2. The reliability analysis method for early warning of deformation of the sidewall of the underground powerhouse of a pumped storage power station according to claim 1, characterized in that, In step 3, the goodness of fit is calculated by combining the predicted value and the average prediction variance e of the sample set; the goodness of fit R is a value between 0 and 1. The closer the value of R is to 1, the better the goodness of fit, that is, the higher the reliability of the early warning model's prediction. In the formula: yi represents the predicted value of deformation displacement, and e represents the average prediction variance of the sample set; Variance between measured and predicted values e1 The calculation formula is as follows: In the formula: This represents the measured value of the deformation displacement; The method for updating the average prediction variance e* is as follows: Let the number of sample shifts be n; then the updated sample average prediction variance is: Similarly, as the measured deformation displacement values ​​are updated, the sample mean variance is updated in real time based on the variance between the measured and predicted values.

3. The reliability analysis method for early warning of deformation of the sidewall of the underground powerhouse of a pumped storage power station according to claim 1, characterized in that, The displacement time series mentioned in step 3 is constructed as follows: (1): Input the monitored deformation and displacement data of the underground powerhouse sidewall of the pumped storage power station and construct the displacement time series sample data N according to the optimal length L; (2): Let the historical point be l, and use the first l measured displacement data as the input vector x; (3): Input the input vector into the early warning model to obtain the corresponding (l+1)th predicted displacement data; then replace the oldest value of the original input vector with the (l+1)th measured displacement data to form a new input sequence with the same time length, and obtain the (l+2)th predicted displacement data. Then replace the oldest value of the original input vector with the (l+2)th measured displacement data to form a new input sequence with the same time length. And so on, to build multiple slope displacement time series samples.

4. The reliability analysis method for early warning of deformation of the sidewall of the underground powerhouse of a pumped storage power station according to claim 1, characterized in that, In step 1, the early warning model is constructed as follows: S1: Select samples; Divide the historical measured displacement data of the deformation of the underground powerhouse sidewall of the pumped storage power station into a training set and a test set; The training set is used to train the early warning model based on the information vector machine; The test set is used to test the trained early warning model and modify and adjust the parameters of the early warning model according to the test results; S2: Constructing Samples; Based on the measured data of deformation and displacement of the underground powerhouse sidewalls, construct a displacement time series; Let the number of historical points be l, take the first 1 to l measured values ​​as the input vector, and take the (l+1)th predicted value as the output value, thus forming the first sample; then use the (l+1)th measured value as the new measured value to replace the oldest value in the original input vector, forming a new input sequence with the same time length, and take the (l+2)th predicted value as the output value, thus forming the second sample; and so on, constructing a series of displacement time series samples; S3: Establish a sidewall deformation early warning model; train the early warning model, set the initial parameters of the early warning model, and use the classic k-fold cross-validation method to complete an information vector machine sidewall deformation early warning model when the accuracy rate reaches 95% or higher. In this way, establish a nonlinear mapping relationship between the displacement value at any time and the displacement value at the previous n times, and determine the optimal displacement time series length L. S4: Verify the predictive performance of the early warning model using the test set; if the good fit between the predicted and measured values ​​is high, the established early warning model for sidewall deformation is considered to meet the requirements and is feasible for predicting the sidewall deformation of the underground powerhouse of the pumped storage power station; otherwise, adjust the sample and training parameters of the early warning model and retrain the early warning model.

5. A reliability analysis device for early warning of deformation of the sidewall of the underground powerhouse of a pumped storage power station, characterized in that, It includes field devices and terminal equipment; field devices include acquisition units and transmission units, and terminal equipment includes transmission units, processing units, and evaluation units; The data acquisition unit includes a high-precision three-dimensional displacement sensor, which is installed at the displacement monitoring point of the sidewall of the underground powerhouse of the pumped storage power station and connected to a wireless device to collect real-time data on the displacement changes of the sidewall of the underground powerhouse of the pumped storage power station in three directions. The transmission unit includes a wireless data transmission module. The wireless data transmission module uses wireless bridge technology and is a store-and-forward device that realizes LAN interconnection at the link layer. It uses air as a medium to propagate signals, solving the problem of difficult wired deployment. The wireless devices are connected to the displacement sensor and processing unit of the acquisition unit respectively to realize data transmission between field devices and terminal devices. The processing unit, including a microprocessor and a read / write memory, is used to predict the deformation displacement of the underground powerhouse sidewall of the pumped storage power station at a certain moment by using the displacement input vector with a trained early warning model, and to perform goodness-fit analysis by combining the predicted data with the average prediction variance e. The microprocessor is a central processing unit composed of a large-scale integrated circuit, which is responsible for extracting script instructions from the read / write memory, executing script instructions, and controlling other components. Read / write memory is used to store various program scripts and data generated during the calculation process, classification models based on information vector machines, and prediction results; The evaluation unit includes a confidence calculation algorithm, which is programmed into a read-write memory to calculate the goodness of fit of the predicted value, obtain the goodness of fit R, and evaluate the confidence of the warning based on the magnitude of the goodness of fit R.

Images