Hydroelectric generating set fault prediction method based on EMD and LSTM fusion model

Abstract

The invention relates to a hydroelectric generating set fault prediction method based on an EMD and LSTM fusion model, and the method comprises the steps: collecting a hydroelectric generating set vibration signal data sequence, and carrying out the set preprocessing of the data sequence, and obtaining a data set D1; performing empirical mode decomposition on the data set D1 by using a waveform matching continuation method of parameter traversal optimization to obtain a plurality of corresponding decomposition components; performing high and low frequency signal combination on the plurality of intrinsic mode components obtained by decomposition based on an extreme point division method to generate high and low frequency component reconstruction components; and respectively predicting the high-frequency and low-frequency reconstruction components by using a fusion model based on EMD and LSTM, and carrying out equal weight summation and superposition on prediction results to obtain a final fault prediction result of the hydroelectric generating set. According to the method, the EMD and LSTM fusion model is constructed, the improved empirical mode decomposition is utilized, the potential unit operation state information of the hydroelectric generating set vibration signals is deeply mined, the endpoint effect problem existing in the traditional empirical mode decomposition is inhibited, and the data decomposition quality is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of hydropower unit state trend prediction, and in particular relates to a hydropower unit fault prediction method based on an EMD and LSTM fusion model. Background Art

[0002] As an important equipment for peak load regulation and frequency regulation of power grid, hydropower units are affected by hydraulic, mechanical and electrical factors during operation, and various potential faults seriously affect the safe operation of hydropower stations. The swing signal of the hydropower unit contains a large amount of equipment operation status information, which is a key monitoring quantity for evaluating the safe and stable operation of the unit. Therefore, the trend change of the swing signal of the hydropower unit can timely grasp its operation status and law, avoid potential equipment failure risks, and effectively improve the comprehensive benefits of the hydropower station.

[0003] The existing technical method is to apply the empirical mode decomposition algorithm to the hydropower unit swing sequence data, and then use the signal decomposition algorithm to process the unit swing data to extract the unit operation status information in the data. However, in the empirical mode decomposition, it is impossible to clearly judge the extreme value at the endpoint, and the flying wing phenomenon occurs when the envelope is fitted at the endpoint. Summary of the invention

[0004] The technical problem of the present invention is to improve the accuracy of the swing signal trend of the hydropower unit and reduce the influence of the nonlinear and non-stationary signals of the unit on the prediction results, and to predict the operation trend of the unit by analyzing the normal operation data and fault data of the hydropower unit, so as to timely discover the potential fault hazards of the unit, reduce the losses caused by shutdown and maintenance, and ensure the stable operation of the unit. The purpose of the present invention is to solve the above problems and propose a hydropower unit fault prediction method based on the EMD and LSTM fusion model, comprising the following steps: S1: Collect the data sequence of the hydropower unit vibration signal, and perform set preprocessing on the data sequence to obtain the data set D1; S2: Using the waveform matching extension method with parameter traversal optimization, perform empirical mode decomposition on the data set D1 to obtain the corresponding multiple decomposition components; S3: Based on the extreme point partitioning method, multiple intrinsic mode components decomposed in S2 are combined into high- and low-frequency signals to generate high- and low-frequency component reconstruction components; S4: Use the EMD and LSTM fusion model to predict the high-frequency and low-frequency reconstruction components respectively, and superimpose the prediction results by equal weight summation to obtain the final prediction result of the hydropower unit failure.

[0005] Preferably, in step S1, the data set is preprocessed, including removing outliers, deleting missing values ​​and performing sliding average processing on the original monitoring data; the sliding average processing is used to strengthen the effect of recent data and weaken the influence of long-term data by calculating the moving average of the sequence data.

[0006] Preferably, preprocessing the data set also includes normalizing the data, and the calculation formula is: ; In the formula, represents the original sequence data, is the standardized data sequence, x max Represents the maximum value of the data sequence, x min Represents the minimum value of the data sequence.

[0007] Furthermore, in step S2, the waveform matching extension method of parameter traversal optimization includes extending the endpoint part of the data sequence to suppress the endpoint effect of traditional EMD and improve the quality of data decomposition.

[0008] Preferably, the waveform matching extension method of traversing and optimizing comprises the following sub-steps: 1) Extend the endpoints of the original simulation signal; 2) Mark the maximum and minimum points within the index range of the original signal, and all maximum and minimum points constitute a set respectively; 3) Set the initial value and traversal step of the matching parameter, and calculate the waveform matching comprehensive index of the set under different matching parameter values; 4) Take the signal of the matching wavefront corresponding to the maximum value of the waveform matching index as the extended wave, extend the original signal, record the maximum value coordinates and calculate the coordinates of the left and right endpoints; 5) Perform empirical mode decomposition on the extended data sequence to obtain multiple intrinsic mode components.

[0009] Preferably, the waveform matching comprehensive index includes using the Pearson correlation coefficient to calculate the linear similarity of the shape between the characteristic wave and the matching wave, and the calculation formula is: ; In the formula, It means to calculate the covariance between two variables. represents the calculated variance, x and y represent the characteristic wave and the matching wave respectively.

[0010] Preferably, the waveform matching comprehensive index includes using the Euclidean distance to calculate the similarity between the characteristic wave and the matching wave in amplitude, and the calculation formula is: ; ; Where d( ) represents the result of calculating the Euclidean distance, and s represents the amplitude similarity, which ranges from [0,1].

[0011] Furthermore, the waveform matching comprehensive index is obtained by integrating the shape linearity and amplitude similarity between the characteristic wave and the matching wave. The calculation formula of the waveform matching comprehensive index is: ; In the formula, The calculation result of the comprehensive index representing the waveform matching degree, represents the matching parameter, represents the Pearson correlation coefficient.

[0012] Preferably, step S3 includes using an extreme point partitioning method to count the number of maximum and minimum points of each eigenmode component, and selecting the number of extreme points as a threshold for distinguishing high- and low-frequency signal components, thereby reconstructing the eigenmode components into high-frequency components and low-frequency components.

[0013] Compared with the prior art, the beneficial effects of the present invention include: 1) The present invention proposes a hydropower unit fault prediction method based on the EMD and LSTM fusion model, and uses the improved empirical mode decomposition to deeply mine the potential unit operation status information of the hydropower unit swing signal, suppresses the endpoint effect problem existing in the traditional empirical mode decomposition, and improves the data decomposition quality.

[0014] 2) The present invention utilizes the extreme point partitioning method to reconstruct the decomposed components, and the obtained signal components reduce the prediction calculation amount and the complexity of the prediction model, thereby improving the prediction efficiency, prediction stability and generalization ability of the model.

[0015] 3) The present invention constructs an EMD and LSTM fusion model to analyze and predict the trend change of the unit swing signal, which can discover potential fault hazards of the unit in advance and has better prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0017] Figure 1 A flow chart of a method for predicting hydropower unit faults using a fusion model according to an embodiment of the present invention; Figure 2 It is an original time series diagram of the oscillation signal of the hydropower unit according to an embodiment of the present invention; Figure 3 A schematic diagram of a waveform matching and extension process according to an embodiment of the present invention; Figure 4 A curve diagram of the decomposition components of the oscillation signal of a hydropower unit according to an embodiment of the present invention; Figure 5 A characteristic component curve diagram reconstructed from the decomposed components of an embodiment of the present invention; Figure 6 This is a structural diagram of an EMD and LSTM fusion neural network according to an embodiment of the present invention; Figure 7 This is a diagram showing the trend prediction results of the oscillation signal of a hydropower unit according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0019] The present invention is aimed at the 16F unit of a domestic hydropower station from January 1, 2023 to July 30, 2023. X The monitoring data of the upper guide swing was experimentally studied and verified. It was found that the unit had been in a stable operating state from January 1 to July 16. However, on July 17, the upper guide rotor of the unit suddenly jumped up, causing the upper guide swing to trigger an alarm.

[0020] like Figure 1 As shown, the present invention proposes a hydropower unit fault prediction method based on the EMD and LSTM fusion model, comprising the following steps: S1: Collect the data sequence of the hydropower unit vibration signal, and perform set preprocessing on the data sequence to obtain the data set D1; like Figure 2 As shown, the data set is preprocessed, including outlier removal, missing value deletion and sliding average processing of the original monitoring data; the sliding average processing is used to strengthen the effect of recent data and weaken the impact of long-term data by calculating the moving average of the sequence data.

[0021] Preprocessing the data set also includes normalizing the data. The calculation formula is: ; In the formula, represents the original sequence data, is the standardized data sequence, x max Represents the maximum value of the data sequence, x min Represents the minimum value of the data sequence.

[0022] S2: Using the waveform matching extension method of parameter traversal optimization, the data set D1 Perform empirical mode decomposition to obtain corresponding multiple decomposition components; In step S2, the waveform matching extension method of parameter traversal optimization includes extending the endpoint part of the data sequence to suppress the traditional EMD endpoint effect and improve the quality of data decomposition.

[0023] like Figure 3 As shown, the waveform matching extension method of traversal optimization includes the following sub-steps: 1) At the original simulation signal endpoint Extend the left endpoint and define the characteristic wave S 1, assuming that the characteristic wave has a maximum point first and a minimum point later. The left end point of the characteristic wave is S 0, the first maximum point is M 0, the first minimum point is N 0; the length of the characteristic wave is N , left endpoint S 0 and the first maximum point M The distance between 0 is L ,During actual analysis, the characteristic wave length is defined according to the ,characteristics of the measured signal.

[0024] 2) In the original simulation signal Mark the maximum and minimum points within the index range of , and the set of all maximum points is M max , the set of all minimum points is M min .

[0025] 3) Set the initial value and traversal step of the matching parameter, and calculate the waveform matching comprehensive index of the set under different matching parameter values; set the matching parameter The initial value and traversal step length are based on the first maximum point M 0 is the reference point, and the length of the matching wave is the same as the length of the characteristic wave. Different matching degree parameters are calculated. Comprehensive index of waveform matching under the condition of value selection P .

[0026] 4) Set the waveform matching index to the maximum value P max The signal corresponding to the matching wavefront is used as the extension wave to extend the original signal and record the M The maximum coordinates corresponding to 0 are M p , S The coordinates corresponding to 0 are X p = M p - L .

[0027] 5) FromX p Starting from the previous data point, extend the signal to S 0. In the actual data extension process, the number of points to be extended is set in advance. If the number of signal points to be extended is greater than X p If the number of previous data points is equal, the extended waveform is repeatedly extended.

[0028] Perform empirical mode decomposition on the extended data sequence to obtain multiple intrinsic mode components.

[0029] The waveform matching comprehensive index includes the use of Euclidean distance to calculate the similarity between the characteristic wave and the matching wave in amplitude. The calculation formula is: ; ; In the formula, d( ) Indicates the result of calculating the Euclidean distance, s indicates the amplitude similarity, the value range is [0,1], and the distance similarity The bigger the The smaller it is, the closer the distance is, the greater the similarity is.

[0030] The waveform matching comprehensive index is obtained by integrating the similarity of shape and amplitude between the characteristic wave and the matching wave. The calculation formula of the waveform matching comprehensive index is: ; In the formula, The calculation result of the comprehensive index representing the waveform matching degree, represents the matching parameter, represents the Pearson correlation coefficient.

[0031] The waveform matching comprehensive index includes the use of the Pearson correlation coefficient to calculate the linear similarity of the shape between the characteristic wave and the matching wave. The calculation formula is: ; In the formula, It means to calculate the covariance between two variables. represents the calculation of variance, Indicates that the signal EMD After decomposition i eigenmode components; x i is the actual component of the corresponding original signal.

[0032] As shown in Table 1, the parameter traversal optimization extension method effectively suppresses EMD Endpoint effect; Table 1 Comparison of similarity coefficients of simulation signals before and after extension

[0033] The continuation method is used to X Extend the monitoring data of the swing degree. EMD The decomposition results are as follows Figure 4 shown.

[0034] S3: Based on the extreme point partitioning method, multiple intrinsic mode components decomposed in S2 are combined into high- and low-frequency signals to generate high- and low-frequency component reconstruction components; As shown in Table 2, select the appropriate number of extreme points E =200 as the threshold for distinguishing high- and low-frequency signal components; Table 2 Number of extreme points of each IMF component after EMD decomposition

[0035] Step S3 obtains the IMF component reconstruction result as follows: Figure 5 shown.

[0036] S4: Use the EMD and LSTM fusion model to predict the high- and low-frequency reconstruction components respectively, and superimpose the prediction results by equal-weighted summation to obtain the final prediction result of the hydropower unit swing trend.

[0037] like Figure 6 The structure diagram of the EMD and LSTM fusion neural network is shown in the figure. The hidden layer of the LSTM network mainly contains three gate structures: forget gate, input gate and output gate. The forget gate is used to filter and retain t – Neurons at 1 moment c t−1 Historical state information, input gate decision t New input information to neurons at each moment x t The output gate is used to control the output value of the hidden layer. h t Information output. The forward propagation algorithm of the LSTM neural network is as follows: Forget Gate:

[0038] Input Gate:

[0039]

[0040] Neuron state update expression:

[0041] Output Gate:

[0042]

[0043] In the formula, x t for t Always input information from the network. h t-1 for t -1 time network hidden layer state output value, is the candidate state value of the input gate, c t-1 , c t are the states of neurons at different times, is the activation function, W F 、W I 、W C 、W O is the corresponding weight matrix, b f 、b i 、b c 、b o is the corresponding threshold vector.

[0044] In order to better illustrate the effectiveness of this method, a single LSTM As well as traditional EMD and LSTM The combined model is used as a comparison model, and the comparison results of the prediction evaluation indicators are shown in Table 3; Table 3 Evaluation indicators of each prediction model

[0045] It can be seen from the above data that the method of the present invention can provide more accurate and reliable prediction results with the smallest difference from the true value, and realizes the prediction of the X-direction swing trend of the upper guide of the 16F unit of a domestic hydropower station.

[0046] like Figure 7 As shown in the prediction result diagram of the fusion model of the present invention, the blue solid line in the diagram is the upper derivative X The actual value of the swing is shown in Figure 2. The green solid line is the predicted value of the proposed fusion model. The fusion model successfully analyzes and predicts the upper derivative. X The swing trend changed and potential failure risks of the unit were discovered.

[0047] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A hydropower unit fault prediction method based on the EMD and LSTM fusion model is characterized by: The following steps are involved: S1: Collect the data sequence of the hydropower unit vibration signal, and perform set preprocessing on the data sequence to obtain the data set D1; S2: Using the waveform matching extension method with parameter traversal optimization, perform empirical mode decomposition on the data set D1 to obtain the corresponding multiple decomposition components; S3: Based on the extreme point partitioning method, multiple intrinsic mode components decomposed in S2 are combined into high- and low-frequency signals to generate high- and low-frequency component reconstruction components; S4: Use the EMD and LSTM fusion model to predict the high- and low-frequency reconstruction components respectively, and superimpose the prediction results by equal-weighted summation to obtain the final prediction result of the hydropower unit swing trend.

2. According to the method for predicting hydropower unit faults based on the EMD and LSTM fusion model according to claim 1, it is characterized in that: In step S1, the data set is preprocessed, including removing outliers, deleting missing values ​​and performing sliding average processing on the original monitoring data; the sliding average processing is used to strengthen the effect of recent data and weaken the influence of long-term data by calculating the moving average of the sequence data.

3. According to the hydropower unit fault prediction method based on the EMD and LSTM fusion model of claim 1, it is characterized in that: In step S2, the parameter traversal optimization waveform matching extension method is used to extend the endpoint part of the data sequence to suppress the endpoint effect of the traditional EMD and improve the quality of data decomposition.

4. According to the method for predicting hydropower unit faults based on the EMD and LSTM fusion model according to claim 3, it is characterized in that: The waveform matching and extension method of traversal optimization includes the following sub-steps: 1) Extend the endpoints of the original simulation signal; 2) Mark the maximum and minimum points within the index range of the original signal, and all maximum and minimum points constitute a set respectively; 3) Set the initial value and traversal step of the matching parameter, and calculate the waveform matching comprehensive index of the set under different matching parameter values; 4) Take the signal of the matching wavefront corresponding to the maximum value of the waveform matching index as the extended wave, extend the original signal, record the maximum value coordinates and calculate the coordinates of the left and right endpoints; 5) Perform empirical mode decomposition on the extended data sequence to obtain multiple intrinsic mode components.

5. According to the method for predicting hydropower unit faults based on the EMD and LSTM fusion model according to claim 4, it is characterized in that: The waveform matching comprehensive index includes using the Pearson correlation coefficient to calculate the linear similarity of the shape between the characteristic wave and the matching wave. The calculation formula is: ； In the formula, It means to calculate the covariance between two variables. represents the calculated variance, x and y represent the characteristic wave and the matching wave respectively.

6. The hydropower unit fault prediction method based on the EMD and LSTM fusion model according to claim 4 is characterized in that: The waveform matching comprehensive index includes using the Euclidean distance to calculate the similarity between the characteristic wave and the matching wave in amplitude, and the calculation formula is: ； ； In the formula, Indicates the result of calculating the Euclidean distance. Represents the amplitude similarity, and its value range is [0,1].

7. The hydropower unit fault prediction method based on the EMD and LSTM fusion model according to claim 6 is characterized in that: The waveform matching comprehensive index is obtained by integrating the shape linearity and amplitude similarity between the characteristic wave and the matching wave. The waveform matching comprehensive index calculation formula is: ； In the formula, The calculation result of the comprehensive index representing the waveform matching degree, represents the matching parameter, represents the Pearson correlation coefficient.

8. The hydropower unit fault prediction method based on the EMD and LSTM fusion model according to claim 1 is characterized in that: The step S3 includes using an extreme point partitioning method to count the number of maximum and minimum points of each eigenmode component, and selecting the number of extreme points as a threshold for distinguishing high- and low-frequency signal components, thereby reconstructing the eigenmode components into high-frequency components and low-frequency components.

Images