Multi-step prediction method for degradation trend of hydroelectric generating set

Abstract

The invention provides a hydroelectric generating set degradation trend multi-step prediction method, and relates to the technical field of hydroelectric generating set state monitoring and trend prediction, and the method comprises the steps: processing vibration monitoring signal waveform data of a hydroelectric generating set through employing a feature processing method, and obtaining a set vibration signal degradation index; combining and smoothing the unit vibration signal degradation indexes according to a time sequence to obtain a training set and a test set; inputting the training set into a hydroelectric generating set degradation trend multi-step prediction model, and training by using a loss function to obtain a trained hydroelectric generating set degradation trend multi-step prediction model; the trained hydroelectric generating set degradation trend multi-step prediction model is used for analyzing the test set, a degradation trend multi-step prediction result is obtained, and multi-step prediction of the hydroelectric generating set degradation trend is completed. The problem that the multi-step degradation trend of the hydroelectric generating set is difficult to accurately predict is solved.

Description

Technical Field

[0001] This specification relates to the technical field of condition monitoring and trend prediction of hydropower units, and particularly relates to a multi-step prediction method for the deterioration trend of hydropower units. Background Art

[0002] In recent years, the capacity scale of single units has been increasing continuously, posing higher requirements for the safe and stable operation and maintenance level of hydropower units. Under the trend of digital transformation, the operation and maintenance mode of hydropower units urgently needs to change from the traditional "accident maintenance" and "planned maintenance" to "condition-based maintenance". However, there are still great difficulties and challenges in the process of improving the transformation of hydropower units to the condition-based maintenance mode. First of all, the monitoring data generated during the operation of hydropower units has the characteristics of multi-source, non-linearity and time-variation, and it is difficult for traditional data analysis methods to effectively mine the hidden information in the data. In addition, the system structure of hydropower units is complex, with a large number of equipment, and different components of the system influence and interact with each other, making it difficult to accurately grasp the real-time changes of the unit state and the development of the fault trend.

[0003] Deterioration trend prediction is a key link in equipment health management and predictive maintenance, aiming to predict the performance degradation trend of equipment by analyzing equipment operation data, so as to arrange maintenance in advance and avoid faults. Previous studies often considered predicting the future single-step trend and ignored the importance of future multi-step trend changes. Summary of the Invention

[0004] Aiming at the above deficiencies in the prior art, a multi-step prediction method for the deterioration trend of hydropower units provided by the present invention solves the problem that it is difficult to accurately predict the multi-step deterioration trend of hydropower units.

[0005] In order to achieve the above invention purpose, the technical solution adopted by the present invention is: a multi-step prediction method for the deterioration trend of hydropower units, including:

[0006] S1: Using a feature processing method, process the vibration monitoring signal waveform data of the hydropower unit to obtain the deterioration index of the unit vibration signal;

[0007] S2: Combine and smooth the deterioration indexes of the unit vibration signal according to the time series to obtain a training set and a test set;

[0008] S3: Construct a multi-step prediction model for the deterioration trend of the hydropower unit;

[0009] S4: Input the training set into the multi-step prediction model for the deterioration trend of the hydropower unit, and use the loss function for training to obtain a trained multi-step prediction model for the deterioration trend of the hydropower unit; wherein, the trained multi-step prediction model for the deterioration trend of the hydropower unit is used to analyze the test set to obtain the multi-step prediction result of the deterioration trend, and complete the multi-step prediction of the deterioration trend of the hydropower unit.

[0010] The beneficial effects of the present invention are as follows: a multi-step prediction method for the deterioration trend of a hydropower unit. (1) By extracting and screening features from the waveform data of the vibration monitoring signals of the hydropower unit, low-quality features are eliminated. Then, through dimensionality reduction, feature redundancy is further reduced. Finally, the main element eigenvalues are weighted and fused using contribution degrees to obtain the unit health assessment index. (2) The multi-step prediction method is applied to the long-term prediction of the hydropower unit, and a multi-step prediction model for the deterioration trend of the hydropower unit is proposed. Based on a multi-input multi-output strategy, the present invention realizes multi-step prediction of the deterioration trend of the hydropower unit. On the basis of the Seq2Seq framework, the overall network structure of the multi-step prediction model is constructed by fusing BiGRU and CNN, which can effectively capture the connections between sequence data and mine the internal features of the sequence data. To improve the model training time, the Optuna method is used to optimize the network parameters. Finally, a dynamic analysis strategy for the deterioration trend is proposed based on the interquartile range. Through the above methods, accurate prediction of the future multi-step trend changes of the unit is realized.

[0011] Further, S1 includes:

[0012] S110: Divide the waveform data of the vibration monitoring signals of the hydropower unit to obtain sample sets of different periods;

[0013] S120: Use the feature processing method to preprocess the sample sets of different periods to obtain a feature set after dimensionality reduction;

[0014] S130: Combine the main element contribution degrees to perform weighted fusion on the feature set after dimensionality reduction to obtain the deterioration index of the unit vibration signal.

[0015] By constructing a comprehensive evaluation index for feature performance, feature parameters with low fault sensitivity can be effectively eliminated.

[0016] Further, S120 includes:

[0017] Extract features from the sample sets of different periods to obtain time-domain features and frequency-domain features;

[0018] Calculate the feature performance of the time-domain features and frequency-domain features to obtain the corresponding comprehensive evaluation index for feature performance;

[0019] Combine the features with the comprehensive evaluation index of feature performance greater than the evaluation threshold to obtain the original feature set; wherein, the feature extraction, the feature performance calculation, and the feature combination belong to the feature processing method;

[0020] Use the kernel principal component analysis method to perform dimensionality reduction processing on the original feature set to obtain a feature set after dimensionality reduction.

[0021] Furthermore, the expression of the comprehensive evaluation index for characteristic performance is as follows:

[0022] M(l) = a1Mon(l) + a2Cor(l) + a3Rob(l)], a1 + a2 + a3 = 1, 1 > a j > 0;

[0023]

[0024] Among them, M(l) represents the comprehensive evaluation index for characteristic performance, l represents the l-th characteristic, a1 represents the coefficient of the monotonicity index, Mon represents the monotonicity index, a2 represents the coefficient of the correlation index, Cor represents the correlation index, a3 represents the coefficient of the robustness index, Rob represents the robustness index, a j represents the coefficient of the j-th index, d i represents the difference in the rank values of the i-th feature pair, n represents the total number of observed samples, N represents the number of time points, t i represents the i-th vector in the time vector, represents t i 's mean value, y i represents the i-th element in the smoothed data sequence, represents y i 's mean value, x i represents the eigenvalue corresponding to the i-th time point, represents the average trend obtained after smoothing the i-th time point, exp represents the exponential form of the natural constant.

[0025] The contribution level of the signal can reflect the change in the signal state. By introducing the principal component contribution degree, more comprehensive and in-depth information about the unit can be mined, thus obtaining an index that reflects the true change in the unit state.

[0026] Furthermore, the expression of the deterioration index of the unit vibration signal is as follows:

[0027] HAI = χ1X1 + χ2X2 +,..., + χ k X k ;

[0028] Among them, HAI represents the deterioration index of the unit vibration signal, χ1 represents the first principal component eigenvalue, X1 represents the first principal component contribution degree value, χ2 represents the second principal component eigenvalue, X2 represents the second principal component contribution degree value, χ k represents the k-th principal component eigenvalue, X k represents the k-th principal component contribution degree value.

[0029] Perform a trend decomposition on the predicted index data. The decomposed time series data can better reflect the true change law and trend of the unit, facilitating the training effect and prediction accuracy of subsequent prediction models.

[0030] Further, S2 includes:

[0031] Using an exponentially weighted smoothing method, calculate the deterioration index of the unit vibration signal to obtain an exponentially weighted moving average:

[0032] EWMA_t = (1 - α) × x_t + α × EWMA_t-1;

[0033] where EWMA_t represents the exponentially weighted moving average at time point t, α represents the smoothing coefficient, x_t represents the observed value at time point t, and EWMA_t-1 represents the exponentially weighted moving average at time point t - 1;

[0034] Combine the exponentially weighted moving averages according to the time series and divide them proportionally to obtain a training set and a test set.

[0035] Further, the multi-step prediction model for the deterioration trend of a hydropower unit includes:

[0036] A bidirectional gated recurrent unit layer for obtaining the bidirectional dependency relationship in the training set, performing encoding processing, and obtaining encoded data;

[0037] A convolutional neural network layer for analyzing the context relationship of the encoded data, performing decoding processing, and obtaining decoded data;

[0038] A fully connected layer for activating the decoded data using an activation function to obtain the multi-step prediction result of the deterioration trend.

[0039] By setting an alarm threshold as the basis for dynamic adjustment of model training, the accuracy of the prediction result can be effectively improved.

[0040] Further, S4 includes:

[0041] S410: Input the training set into the multi-step prediction model for the deterioration trend of the hydropower unit to obtain deterioration trend data, and use a loss function for training to obtain a trained multi-step prediction model for the deterioration trend of the hydropower unit;

[0042] S420: Based on the alarm threshold, analyze the deterioration trend data to obtain the number of occurrences of abnormal points;

[0043] S430: When the number of consecutive occurrences of abnormal points is greater than the qualified threshold, return to S410 for retraining; otherwise, obtain a trained multi-step prediction model for the deterioration trend of the hydropower unit;

[0044] S440: Use the trained multi-step prediction model for the deterioration trend of the hydropower unit to analyze the test set, obtain the multi-step prediction results of the deterioration trend, and complete the multi-step prediction of the deterioration trend of the hydropower unit.

[0045] Furthermore, the expression of the alarm threshold is:

[0046] θ = QU + 1.5IQR;

[0047] Where θ represents the alarm threshold, QU represents the 75th percentile, and IQR represents the interquartile range, that is, the difference between QU and the 25th percentile. Description of the Drawings

[0048] This specification will be further described in the form of exemplary embodiments, and these exemplary embodiments will be described in detail through the drawings. These embodiments are not restrictive. In these embodiments, the same numbers represent the same structures, where:

[0049] Figure 1 is an exemplary flowchart of a multi-step prediction method for the deterioration trend of a hydropower unit according to some embodiments of this specification. Detailed Embodiments

[0050] The following describes the detailed embodiments of the present invention to facilitate those skilled in the art of this technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the detailed embodiments. For those of ordinary skill in the art of this technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

[0051] Embodiment

[0052] Figure 1 is an exemplary flowchart of a multi-step prediction method for the deterioration trend of a hydropower unit according to some embodiments of this specification. As Figure 1 shown, the process includes the following steps. In some embodiments, the process can be executed by a processor.

[0053] S1: Use the feature processing method to process the vibration monitoring signal waveform data of the hydropower unit to obtain the deterioration index of the unit vibration signal.

[0054] The feature processing method is a method for extracting time-domain features and frequency-domain features. For example, the feature processing method can include methods such as feature extraction, feature screening, and feature dimensionality reduction.

[0055] The vibration monitoring signal waveform data of the hydropower unit is the axial vibration A waveform data of the hydropower unit.

[0056] In some embodiments, the processor may collect the axial vibration A waveform data of the hydro-generator unit vibration monitoring system to obtain the vibration monitoring signal waveform data of the hydro-generator unit.

[0057] The unit vibration signal deterioration index is a set of time-domain features and frequency-domain features related to the deterioration trend of the hydro-generator unit.

[0058] In some embodiments, the processor may implement S1 based on the following steps.

[0059] S110: Divide the vibration monitoring signal waveform data of the hydro-generator unit to obtain sample sets of different periods.

[0060] The sample sets of different periods are the vibration monitoring signal waveform data of the hydro-generator unit divided based on the time sequence and the same number of samples. For example, the sample sets of different periods may include 400 * 4096 groups of samples divided by a sampling point of 4096 in chronological order.

[0061] S120: Use a feature processing method to preprocess the sample sets of different periods to obtain a feature set with reduced dimensions.

[0062] The feature set with reduced dimensions is a sample set with reduced feature dimensions.

[0063] In some embodiments, the processor may implement S120 based on the following steps: Extract features from the sample sets of different periods to obtain time-domain features and frequency-domain features; Calculate the feature performance comprehensive evaluation index for the time-domain features and frequency-domain features; Combine the features with a feature performance comprehensive evaluation index greater than the evaluation threshold to obtain an original feature set; wherein, the feature extraction, the feature performance calculation, and the feature combination belong to the feature processing method; Use the kernel principal component analysis method to perform dimensionality reduction processing on the original feature set to obtain a feature set with reduced dimensions.

[0064] Time-domain features are features related to the time domain. For example, time-domain features may include features such as maximum value, minimum value, maximum absolute value, mean value, peak-to-peak value, kurtosis, standard deviation, root mean square amplitude, root mean square value, absolute mean value, skewness, margin index, waveform index, pulse index, peak index, skewness factor, kurtosis index, etc.

[0065] Frequency-domain features are features related to the frequency. For example, frequency-domain features may include features such as frequency mean value, frequency variance, frequency skewness, frequency kurtosis, center frequency, standard deviation frequency, root mean square frequency, weighted frequency, frequency dispersion ratio, standard skewness, standard kurtosis, frequency ratio, square root frequency dispersion ratio, etc.

[0066] The feature performance comprehensive evaluation index is an index used to evaluate the comprehensive performance of features.

[0067] In some embodiments, the expression of the comprehensive evaluation index of characteristic performance can be:

[0068] M(l) = a1Mon(l) + a2Cor(l) + a3Rob(l)], a1 + a2 + a3 = 1, 1 > a j > 0;

[0069]

[0070] Among them, M(l) represents the comprehensive evaluation index of characteristic performance, l represents the l-th feature, a1 represents the coefficient of the monotonicity index, Mon represents the monotonicity index, a2 represents the coefficient of the correlation index, Cor represents the correlation index, a3 represents the coefficient of the robustness index, Rob represents the robustness index, a j represents the coefficient of the j-th index, d i represents the difference in the rank values of the i-th feature pair, n represents the total number of observed samples, N represents the number of time points, t i represents the i-th vector in the time vector, represents t i 's mean value, y i represents the i-th element in the smoothed data sequence, represents y i 's mean value, x i represents the eigenvalue corresponding to the i-th time point, represents the average trend obtained after smoothing the i-th time point, exp represents the exponential form of the natural constant.

[0071] The original feature set is a feature set related to the deterioration trend.

[0072] In some embodiments, the processor can combine the time-domain features and frequency-domain features with a comprehensive evaluation index of characteristic performance greater than 0.3 to obtain the original feature set.

[0073] S130: Combine the principal component contribution degree to perform weighted fusion on the dimension-reduced feature set to obtain the deterioration index of the unit vibration signal.

[0074] The principal component contribution degree is the degree of explanation of each principal component to the total variation of the data in the kernel principal component analysis method, and it is also the basis for performing dimension reduction processing.

[0075] In some embodiments, the expression of the deterioration index of the unit vibration signal can be:

[0076] HAI = χ1X1 + χ2X2 +,..., + χ k X k ;

[0077] Among them, HAI represents the deterioration index of the unit vibration signal, χ1 represents the first principal component eigenvalue, X1 represents the first principal component contribution value, χ2 represents the second principal component eigenvalue, X2 represents the second principal component contribution value, and χ k represents the k-th principal component eigenvalue, and X k represents the k-th principal component contribution value.

[0078] In some embodiments, the principal components and contribution degrees are shown in Table 1.

[0079] Table 1 Principal components and corresponding contribution degree table

[0080]

[0081] S2: Combine and smooth the deterioration index of the unit vibration signal according to the time series to obtain a training set and a test set.

[0082] The training set is a data set used to train the multi-step prediction model for the deterioration trend of the hydropower unit.

[0083] The test set is a data set used to test the performance of the multi-step prediction model for the deterioration trend of the hydropower unit.

[0084] In some embodiments, the processor can divide the deterioration index of the unit vibration signal in a ratio of 7:3 to obtain a training set and a test set.

[0085] In some embodiments, the processor can implement S2 based on the following steps: Use the exponentially weighted smoothing processing method to calculate the deterioration index of the unit vibration signal to obtain the exponentially weighted moving average value; Combine the exponentially weighted moving average values according to the time series and divide them according to a ratio to obtain a training set and a test set.

[0086] In some embodiments, the expression of the exponentially weighted moving average value can be:

[0087] EWMA_t = (1 - α) × x_t + α × EWMA_t-1;

[0088] Among them, EWMA_t represents the exponentially weighted moving average value at time point t, α represents the smoothing coefficient, x_t represents the observed value at time point t, and EWMA_t-1 represents the exponentially weighted moving average value at time point t-1.

[0089] S3: Construct a multi-step prediction model for the deterioration trend of the hydropower unit.

[0090] The multi-step prediction model for the deterioration trend of the hydropower unit is a neural network model used to obtain the multi-step prediction result of the deterioration trend. The type of the multi-step prediction model for the deterioration trend of the hydropower unit can be various. For example, the type of the multi-step prediction model for the deterioration trend of the hydropower unit can include convolutional neural network, deep neural network, etc.

[0091] In some embodiments, the input of the multi-step prediction model for the deterioration trend of a hydroelectric generating unit may be a training set, and the output of the multi-step prediction model for the deterioration trend of a hydroelectric generating unit may be the multi-step prediction result of the deterioration trend.

[0092] In some embodiments, the structure of the multi-step prediction model for the deterioration trend of a hydroelectric generating unit is as follows:

[0093] The multi-step prediction model for the deterioration trend of a hydroelectric generating unit includes a bidirectional gated recurrent unit layer, a convolutional neural network layer, and a fully connected layer. The output of the bidirectional gated recurrent unit layer serves as the input of the convolutional neural network layer, the output of the convolutional neural network layer serves as the input of the fully connected layer, and the output of the fully connected layer serves as the final output of the multi-step prediction model for the deterioration trend of a hydroelectric generating unit.

[0094] The bidirectional gated recurrent unit layer is used to obtain the bidirectional dependencies in the training set, perform encoding processing, and obtain encoded data. The input of the bidirectional gated recurrent unit layer may include the training set, and the output may include the encoded data.

[0095] The encoded data is the data obtained by encoding the training set.

[0096] In some embodiments, the bidirectional gated recurrent unit layer may perform forward calculation on the input sequence X = x1, x2, …, x T to obtain the hidden states from left to right and from right to left; concatenate the hidden states in the two directions to obtain the final hidden state; and pass the hidden state to a fully connected layer to obtain the output encoded data.

[0097] In some embodiments, the expressions for the hidden states from left to right and from right to left may be respectively:

[0098]

[0099] where and respectively represent the hidden states from left to right and from right to left, represents the hidden state at time t - 1, represents that at time t + 1, GRU represents a GRU cell, and xt represents the t-th element in the input sequence.

[0100] In some embodiments, the expression for the final hidden state may be:

[0101]

[0102] where [·; ·] represents the vector concatenation operation.

[0103] In some embodiments, the expression for the encoded data may be:

[0104] y t = softmax(Wh t + b);

[0105] where y t represents the encoded data, softmax represents the softmax function, W represents the weight term, h t represents the final hidden state, and b represents the bias.

[0106] In some embodiments, the processor may utilize a BiGRU to better capture the bidirectional dependencies of the sequence data in the training set, so as to realize the connection between the input time series during the encoding process by the model and obtain the encoded data.

[0107] The convolutional neural network layer is used to analyze the context relationship of the encoded data and perform decoding processing to obtain the decoded data. The input of the fully connected layer may include the encoded data, and the output may include the decoded data.

[0108] The decoded data is the data obtained after decoding the encoded data.

[0109] In some embodiments, the processor may utilize the convolutional neural network layer, based on the context vector of the encoded data, and then output a new hidden state to obtain the decoded data. For example, the processor may utilize the convolutional neural network layer to transform the time series data into the form of a supervised learning problem, that is, divide the data into multiple samples with input and output components; for time series data, regard the input data as a one-dimensional structure, use one-dimensional convolution for feature extraction, and different sizes of convolutional kernels can be selected to capture feature information of different scales, and use an activation function to perform non-linear processing on the convolution result; use a pooling sub-layer for downsampling to reduce the size of the feature map and retain key features; after the convolutional and pooling sub-layers, connect the obtained feature map to the output sub-layer through a fully connected sub-layer; the fully connected sub-layer can map the feature map to the space of the target prediction value to obtain the decoded data.

[0110] The fully connected layer is used to activate the decoded data by using an activation function to obtain the multi-step prediction result of the degradation trend. The input of the fully connected layer may include the decoded data, and the output may include the multi-step prediction result of the degradation trend.

[0111] In some embodiments, the processor may use linear as the activation function of the fully connected layer, and use relu as the activation function for the remaining layers; use MSE as the loss function index; determine the model parameters through the Optuna algorithm to avoid the huge waste of computing resources caused by traditional methods such as grid search, random search, and Bayesian optimization.

[0112] S4: Input the training set into the multi-step prediction model for the deterioration trend of the hydropower unit, and train it using the loss function to obtain a trained multi-step prediction model for the deterioration trend of the hydropower unit. Among them, the trained multi-step prediction model for the deterioration trend of the hydropower unit is used to analyze the test set to obtain the multi-step prediction result of the deterioration trend, thereby completing the multi-step prediction of the deterioration trend of the hydropower unit.

[0113] The multi-step prediction result of the deterioration trend is the prediction of the deterioration trend of the hydropower unit.

[0114] In some embodiments, the processor may implement S4 based on the following steps.

[0115] S410: Input the training set into the multi-step prediction model for the deterioration trend of the hydropower unit to obtain the deterioration trend data, and train it using the loss function to obtain a trained multi-step prediction model for the deterioration trend of the hydropower unit.

[0116] In some embodiments, the multi-step prediction model for the deterioration trend of the hydropower unit can be trained by multiple labeled training samples. For example, multiple labeled training samples can be input into the initial multi-step prediction model for the deterioration trend of the hydropower unit, and the loss function is constructed based on the labels and the results of the initial multi-step prediction model for the deterioration trend of the hydropower unit. The parameters of the initial multi-step prediction model for the deterioration trend of the hydropower unit are iteratively updated based on the loss function by gradient descent or other methods. When the preset conditions are met, the model training is completed to obtain a trained multi-step prediction model for the deterioration trend of the hydropower unit. Among them, the preset conditions can be that the loss function converges, the number of iterations reaches the threshold, etc.

[0117] In some embodiments, the training samples may include the training set. The label can be the corresponding deterioration trend result. The label can be manually marked.

[0118] S420: Analyze the deterioration trend data based on the alarm threshold to obtain the number of occurrences of abnormal points.

[0119] The alarm threshold is the threshold used to determine whether the data is abnormal and requires an alarm.

[0120] In some embodiments, the expression of the alarm threshold can be:

[0121] θ = QU + 1.5IQR;

[0122] Among them, θ represents the alarm threshold, QU represents the 75th percentile, and IQR represents the interquartile range, that is, the difference between QU and the 25th percentile.

[0123] S430: When the number of consecutive occurrences of abnormal points is greater than the qualified threshold, return to S410 for retraining; otherwise, obtain a trained multi-step prediction model for the deterioration trend of the hydropower unit.

[0124] In some embodiments, the performance evaluation metrics of the multi-step prediction model for the degradation trend of a hydro-generating unit may include:

[0125]

[0126] where MSE represents the mean squared error, MAE represents the mean absolute error, RMSE represents the root mean squared error, m represents the number of predicted values, T represents the corresponding time point, and y T represents the predicted value of the multi-step prediction model for the degradation trend of the hydro-generating unit, and represents the actual value.

[0127] In some embodiments, the training effect of the multi-step prediction model for the degradation trend of the hydro-generating unit is shown in Table 2.

[0128] Table 2 Training Effect of the Multi-step Prediction Model for the Degradation Trend of the Hydro-generating Unit

[0129]

[0130]

[0131] In some embodiments, as shown in Table 2, by constructing a Seq2Seq framework and introducing a deep learning model, it can be seen from the obtained prediction results that this framework has a lower error when dealing with multi-step prediction, demonstrating obvious advantages. Further, it can be seen that compared with the prediction model based on LSTM, the BiGRU-CNN model shows excellent performance when dealing with time series data. After combining with the Optuna optimization algorithm, the error of the model is significantly reduced. For example, the error results obtained by the Seq2Seq-BiGRU-CNN method are 0.088, 0.071, and 0.0078 respectively; while the error results obtained by this method after hyperparameter optimization are 0.041, 0.0385, and 0.0026 respectively.

[0132] By making full use of the unit monitoring signal data, through feature extraction, feature screening, and feature dimensionality reduction of the waveform data, low-quality features are eliminated and features with large amounts of information are retained; then, combined with the weighted fusion of the principal component contribution degree, the unit degradation evaluation index is obtained; finally, by constructing an Optuna-Seq2Seq-BiGRU-CNN model, multi-step prediction of the unit degradation trend is effectively carried out, which helps to carry out the unit fault prevention work.

[0133] S440: Use the trained multi-step prediction model for the degradation trend of the hydro-generating unit to analyze the test set, obtain the multi-step prediction results of the degradation trend, and complete the multi-step prediction of the degradation trend of the hydro-generating unit.

[0134] In some embodiments of this specification, a multi-step prediction method for the deterioration trend of a hydropower unit is proposed. (1) By performing feature extraction and feature screening on the waveform data of the vibration monitoring signals of the hydropower unit, low-quality features are eliminated. Then, through dimensionality reduction, feature redundancy is further reduced. Finally, the main element eigenvalues are weighted and fused using contribution degrees to obtain the unit health assessment index. (2) Applying the multi-step prediction method to the long-term prediction of hydropower units, a multi-step prediction model for the deterioration trend of hydropower units is proposed. Based on a multi-input multi-output strategy, the present invention realizes multi-step prediction of the deterioration trend of hydropower units. On the basis of the Seq2Seq framework, the overall network structure of the multi-step prediction model is constructed by fusing BiGRU and CNN, which can effectively capture the connections between sequence data and mine the internal features of sequence data. To improve the model training time, the Optuna method is used to optimize the network parameters. Finally, a dynamic analysis strategy for the deterioration trend is proposed based on the interquartile range. Through the above methods, accurate prediction of the future multi-step trend changes of the unit is achieved.

Claims

1. A multi-step prediction method for the deterioration trend of a hydropower unit, characterized in that, Including: S1: Using a feature processing method, process the waveform data of the vibration monitoring signals of the hydro-generating unit to obtain the deterioration index of the unit vibration signal; S2: Combine and smooth the deterioration index of the unit vibration signal according to the time series to obtain a training set and a test set; S3: Construct a multi-step prediction model for the deterioration trend of the hydro-generating unit; S4: Input the training set into the multi-step prediction model for the deterioration trend of the hydro-generating unit, and use the loss function for training to obtain a trained multi-step prediction model for the deterioration trend of the hydro-generating unit; wherein, the trained multi-step prediction model for the deterioration trend of the hydro-generating unit is used to analyze the test set to obtain the multi-step prediction result of the deterioration trend, and complete the multi-step prediction of the deterioration trend of the hydro-generating unit.

2. The multi-step prediction method for the deterioration trend of a hydropower unit according to claim 1, wherein, The S1 includes: S110: Divide the waveform data of the vibration monitoring signals of the hydro-generating unit to obtain sample sets of different periods; S120: Using a feature processing method, preprocess the sample sets of different periods to obtain a feature set after dimensionality reduction; S130: Combine and weight-fuse the feature set after dimensionality reduction with the principal component contribution degree to obtain the deterioration index of the unit vibration signal.

3. The multi-step prediction method for the deterioration trend of a hydropower unit according to claim 2, characterized in that, The S120 includes: Extract features from the sample sets of different periods to obtain time-domain features and frequency-domain features; Calculate the feature performance of the time-domain features and frequency-domain features to obtain the corresponding comprehensive evaluation index of feature performance; Combine the features with the comprehensive evaluation index of feature performance greater than the evaluation threshold to obtain an original feature set; wherein, the feature extraction, the feature performance calculation, and the feature combination belong to the feature processing method; Use the kernel principal component analysis method to perform dimensionality reduction processing on the original feature set to obtain a feature set after dimensionality reduction.

4. The multi-step prediction method for the deterioration trend of a hydropower unit according to claim 3, wherein The expression of the comprehensive evaluation index of feature performance is: M(l) = a1Mon(l) + a2Cor(l) + a3Rob(l)], a1 + a2 + a3 = 1, 1 > a j > 0; Among them, M(l) represents the comprehensive evaluation index of characteristic performance, l represents the l-th feature, a1 represents the coefficient of the monotonicity index, Mon represents the monotonicity index, a2 represents the coefficient of the correlation index, Cor represents the correlation index, a3 represents the coefficient of the robustness index, Rob represents the robustness index, a j represents the coefficient of the j-th index, d i represents the difference in the rank values of the i-th feature pair, n represents the total number of observed samples, N represents the number of time points, t i represents the i-th vector in the time vector, represents t i the mean value of, y i represents the i-th element in the smoothed data sequence, represents y i the mean value of, x i represents the eigenvalue corresponding to the i-th time point, represents the average trend obtained after smoothing the i-th time point, exp represents the exponential form of the natural constant.

5. The multi-step prediction method for the deterioration trend of a hydropower unit according to claim 2, wherein The expression of the deterioration index of the unit vibration signal is: HAI = χ1X1 + χ2X2 +,..., + χ k X k ; Among them, HAI represents the deterioration index of the unit vibration signal, χ1 represents the first principal component eigenvalue, X1 represents the first principal component contribution value, χ2 represents the second principal component eigenvalue, X2 represents the second principal component contribution value, χ k represents the k-th principal component eigenvalue, X k represents the k-th principal component contribution value.

6. The multi-step prediction method for the deterioration trend of a hydropower unit according to claim 1, characterized in that The S2 includes: Using the exponentially weighted smoothing processing method, calculate the deterioration index of the unit vibration signal to obtain the exponentially weighted moving average value: EWMA_t = (1 - α) × x_t + α × EWMA_t-1; wherein, EWMA_t represents the exponentially weighted moving average value at time point t, α represents the smoothing coefficient, x_t represents the observed value at time point t, and EWMA_t-1 represents the exponentially weighted moving average value at time point t - 1; Combine the exponentially weighted moving average values according to the time series and divide them proportionally to obtain a training set and a test set.

7. The multi-step prediction method for the deterioration trend of a hydropower unit according to claim 1, characterized in that The multi-step prediction model for the deterioration trend of the hydro-generating unit includes: A bidirectional gated recurrent unit layer, which is used to obtain the bidirectional dependence relationship in the training set, perform encoding processing, and obtain encoded data; A convolutional neural network layer, which is used to analyze the context relationship of the encoded data, perform decoding processing, and obtain decoded data; A fully connected layer, which is used to activate the decoded data using an activation function to obtain the multi-step prediction result of the deterioration trend.

8. The multi-step prediction method for the deterioration trend of a hydropower unit according to claim 1, characterized in that, The S4 includes: S410: Input the training set into the multi-step prediction model for the deterioration trend of the hydro-generating unit to obtain the deterioration trend data, and use the loss function for training to obtain a trained multi-step prediction model for the deterioration trend of the hydro-generating unit; S420: Analyze the deterioration trend data based on the alarm threshold to obtain the number of occurrences of abnormal points; S430: When the number of consecutive occurrences of abnormal points is greater than the qualified threshold, return to S410 for retraining; otherwise, obtain the trained multi-step prediction model for the deterioration trend of the hydropower unit; S440: Use the trained multi-step prediction model for the deterioration trend of the hydropower unit to analyze the test set, obtain the multi-step prediction results of the deterioration trend, and complete the multi-step prediction of the deterioration trend of the hydropower unit.

9. The multi-step prediction method for the deterioration trend of a hydropower unit according to claim 8, characterized in that, The expression of the alarm threshold is: θ = QU + 1.5IQR; where θ represents the alarm threshold, QU represents the 75th percentile, and IQR represents the interquartile range, that is, the difference between QU and the 25th percentile.

Images