Water-turbine generator set guide bearing temperature prediction method based on EMD-LSTM neural network

Abstract

The invention provides a water-turbine generator set guide bearing temperature prediction method based on an EMD-LSTM neural network, and the method comprises the following steps: carrying out the calculation of the collected historical data: the cooling water flow of a generator guide bearing, the cooling oil flow, the inlet and outlet temperatures of cooling water and oil, the vibration data of a generator, and the temperature of a generator pit; data cleaning is carried out on active power, reactive power and rotating speed of the generator, wrong data are removed, and data filling is carried out on missing data; processing the obtained historical data by using empirical mode decomposition; inputting the processed data into the built long-short-term memory neural network for training; and predicting the temperature of the generator by using the trained neural network. According to the method, the temperature of the guide bearing of the water-turbine generator set can be predicted in advance, operation and maintenance personnel can conveniently find and process abnormal temperature points in advance, and stable operation of the water-turbine generator set is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic monitoring of hydroelectric generators, in particular to a temperature prediction method for guide bearings of a hydroelectric generator set based on an EMD-LSTM neural network. BACKGROUND

[0002] In a hydroelectric generator set, the guide bearing and the thrust bearing of the water turbine are important mechanical structures of the hydroelectric generator set. The guide bearing prevents mechanical imbalance and electromagnetic imbalance of the rotating part of the unit by bearing the radial force transmitted by the main shaft, and maintains the stable operation of the main shaft within the bearing gap. The thrust bearing bears the weight of the rotating part of the water turbine unit and part of the axial thrust. The working state of the guide bearing and the thrust bearing directly affects the stability of the operation of the hydroelectric generator set, and the most common fault is the overheating of the bearing bush, and in severe cases, the bearing bush may even be burned out. Therefore, monitoring the temperature of the bearing bush of the hydroelectric generator set has very important practical significance. By monitoring the temperature of the bearing bush of the hydroelectric generator set in real time, a response can be sent in time when the temperature of the bearing bush is too high to remind the staff to take corresponding measures, thereby protecting the hydroelectric generator and prolonging its service life.

[0003] The existing temperature prediction method for the guide bearing of the hydroelectric generator set uses a BP neural network to predict the maximum value of the bearing bush temperature of the unit. The method is to obtain the real-time working parameters of the hydroelectric generator set and the actual monitoring value of the maximum bearing bush temperature of the hydroelectric generator set in the current time period t every time period t. After normalizing the real-time working parameters, input them into the trained BP neural network temperature prediction model. Obtain the monitoring prediction value of the maximum bearing bush temperature in the future time period t from the BP neural network temperature prediction model, and give an early warning for the bearing bush temperature of the hydroelectric generator set according to the monitoring prediction value. The disadvantages of this temperature prediction method are:

[0004] (1) This scheme does not use a professional data preprocessing method for input data. In fact, due to the influence of data collection and storage, there are many errors and missing phenomena in the existing historical data, which will affect the accuracy of the prediction model.

[0005] (2) The input data used for training the model of the guide bearing temperature of the hydroelectric generator set has strong time correlation. The BP neural network is weak in predicting time series, mainly because the neurons in the same layer of the BP network are independent of each other and are not connected, while time series generally have a correlation between the time and are mutually influenced. The BP neural network cannot extract the time trend characteristics.

[0006] (3) This method only predicts the maximum value of the temperature and does not output all the temperature prediction values, which is not conducive to the analysis of the distribution of the bearing bush temperature at different positions by the operation and maintenance personnel. SUMMARY

[0007] The present application aims at solving the problems in the prior art and provides a guide bearing temperature prediction method for a hydroelectric generator set based on an EMD-LSTM neural network.

[0008] To achieve the above technical features, the present application is implemented in the following manner: a guide bearing temperature prediction method for a hydroelectric generator set based on an EMD-LSTM neural network, comprising the following steps:

[0009] S1, collecting historical data that may cause the temperature of the guide bearing of the hydroelectric generator set to change;

[0010] S2, performing data cleaning on the collected historical data, removing erroneous data and filling in missing data;

[0011] S3, performing standardization processing on the cleaned historical data;

[0012] S4, processing the historical data obtained after the standardization processing using empirical mode decomposition (EMD) to obtain a series of IMFs and a linear superposition of the remaining part;

[0013] S5, inputting each IMF and the remaining part into a long short-term memory neural network (LSTM) to be trained;

[0014] S6, collecting data that may cause the temperature to change at the current time in real time and inputting the data into the trained model in S5 to predict the temperature of the guide bearing of the hydroelectric generator set.

[0015] Preferably, the historical data in S1 includes the cooling water flow of the generator guide bearing, the cooling oil flow, the inlet and outlet temperatures of the cooling water and oil, the vibration data of the generator, the generator pit temperature, and the active power, reactive power and speed of the generator.

[0016] Preferably, the K-nearest neighbor algorithm is used to find a replacement value for the missing value in S2, that is, the average value of K samples is used to fill in the missing value.

[0017] Preferably, the K-nearest neighbor algorithm in S2 has the following specific steps:

[0018] 1) Selecting m candidate samples, calculating the Euclidean distance between the missing sample and all samples in the candidate sample set, and the formula is as follows:

[0019]

[0020] Wherein, L represents the Euclidean distance, x and X represent the features of missing samples and candidate samples respectively, i represents the number of features;

[0021] 2) Calculate m L values [L1, L2,..., Lm] from 1), arrange these values in order from small to large;

[0022] 3) Select K value, take out the temperature values of the candidate samples arranged in the first K L values, and calculate the temperature average value:

[0023]

[0024] Wherein, y is the replacement value to be found, Y is the temperature value of the candidate sample.

[0025] Preferably, the S3 will include the historical journal bearing temperature values and the historical data causing the change thereof are standardized, and the standardization processing specifically adopts z-score standardization processing.

[0026] Preferably, the z-score standardization processing formula is as follows:

[0027]

[0028] Wherein, x * is the standardized value, x is the original value, μ and σ are the mean and standard deviation of all samples respectively.

[0029] Preferably, the S4 will decompose the historical journal bearing temperature data as the original data sequence S(t) by EMD, and the decomposition steps are as follows:

[0030] 1) Find all the maximum points and minimum points of the original data sequence S(t), and use a cubic spline interpolation function to fit to form the upper envelope line and the lower envelope line of the original data;

[0031] 2) Calculate the mean of the upper envelope line and the lower envelope line to obtain the mean envelope line m1;

[0032] 3) Subtract the mean envelope line m1 from the original sequence S(t) to obtain the suspected IMF component h(t);

[0033] 4) Judge whether h(t) satisfies the two constraints of IMF, if yes, h(t) is an IMF component; if not, take h(t) as the original signal, and re-analyze steps 1) to 4) until IMF1 is obtained;

[0034] 5) after IMF1 is obtained through steps 1) to 4), subtract IMF1 from the original signal S(t), take S(t)-IMF1 as a new original signal, and then analyze through steps 1) to 4) again to obtain IMF2, and so on, until the final remaining signal is a single-valued signal or a monotonic signal, at which time the EMD is completed, and n IMF data and a residual data are obtained, and the original signal S(t) is decomposed into a linear superposition of a series of IMF and a residual part through the EMD method:

[0035]

[0036] wherein n is the number of IMF, and r(t) is the residual part.

[0037] Preferably, the IMF component needs to satisfy two constraint conditions:

[0038] Condition 1: the number of extreme points and the number of zero-crossing points must be equal or differ by at most one within the entire IMF;

[0039] Condition 2: at any time, the average value of the upper envelope line formed by the local maximum points and the lower envelope line formed by the local minimum points is zero, i.e., the upper and lower envelope lines are locally symmetric with respect to the time axis.

[0040] Preferably, the LSTM modeling process in S5 is as follows:

[0041] 1) Set the hyperparameters of the LSTM neural network, including the input dimension, the output dimension, the hidden layer dimension, and the cell state dimension, wherein the input dimension is the number of features that may cause temperature changes, and the output dimension is the unit shaft bearing temperature value at a certain time corresponding to the input dimension;

[0042] 2) initialize the weight matrix and bias matrix required for forward calculation of the LSTM, i.e., initialize W f , W i , W o , W c , b f , b i , b o , and b c , which respectively represent the forgetting gate weight matrix, the input gate weight matrix, the output gate weight matrix, the cell state weight matrix, the forgetting gate bias item matrix, the input gate bias item matrix, the output gate bias item matrix, and the cell state bias item matrix of the LSTM neural network;

[0043] 3) bring the selected data that may cause temperature changes into the forgetting gate, the input gate, the output gate, and the cell state of the LSTM at the current time for training of the neural network;

[0044] f t The formula for calculating the forget gate is:

[0045] f t = σ(W f · [h t-1 , x t ] + b f ) ;

[0046] where σ is an activation function of the neural network, and h t-1 is the output matrix of the neural network at the previous moment;

[0047] i t The formula for calculating the input gate is:

[0048] i t = σ(W i · [h t-1 , x t ] + b i ) ;

[0049] c t The formula for calculating the current moment cell state is:

[0050]

[0051] where The calculation method is:

[0052]

[0053] In the formula, the symbol represents element-wise multiplication, and tanh is another activation function,

[0054] o t The formula for calculating the output gate is:

[0055] o t = σ(W o · [h t-1 , x t ] + b o ) ;

[0056] The final output h t of the neural network is determined by the output gate and the cell state:

[0057]

[0058] 4) After the training is completed, the error term of the neural network prediction value and the true value is reversely calculated, the error propagation of the LSTM neural network has two directions, which are respectively along time reverse propagation and propagation to the previous layer of the neural network;

[0059] The output value obtained by the neural network at t time is h t , and the error term at this time is The formula for the reverse propagation along time to any k time is:

[0060]

[0061] Wherein, W fh , W fx , W ih , W ix , W oh , W ox , W ch , W cx are weight matrices W f , W i , s, W c are branches of input dimensions and output dimensions at the previous time;

[0062] The error term of the neural network propagated to the previous layer is:

[0063]

[0064] Wherein, l is a current layer, f l-1 is an activation function of the l-1 layer, is a weighted input value of the t time of the l-1 layer network;

[0065] 5) The gradient of each neuron connection weight W and the bias term b is calculated as follows:

[0066]

[0067] Wherein, ▽ is a gradient operator, and δ is the error obtained in the previous step;

[0068] 6) The connection weight W and the bias term b are updated according to the gradient descent method, and then steps 2) to 5) are repeatedly executed until the error reaches the minimum, that is, the model training is completed.

[0069] Preferably, the S6 specifically comprises: selecting the parameters that may cause the temperature of the guide bearing to change at the current time, and inputting the normalized parameters into the trained LSTM model to obtain normalized data of the predicted value of the guide bearing temperature in the future period of time, and the true value of the predicted data can be obtained after the predicted data is de-normalized.

[0070] The present application has the following beneficial effects:

[0071] 1. The prediction method of this invention can predict the temperature of the guide bearing of a hydro-generator unit in advance, which makes it convenient for operation and maintenance personnel to detect abnormal temperature points in advance and take measures to ensure the stable operation of the hydro-generator unit.

[0072] 2. This invention improves the performance of the prediction model by cleaning the data and then reassigning erroneous or missing data.

[0073] 3. This invention ensures the accuracy of subsequent predictions by performing standardized preprocessing on the data.

[0074] 4. This invention employs a long short-term memory neural network prediction model with time series, thereby enhancing the correlation between neurons and effectively extracting time trend features, thus ensuring the reliability of the prediction model.

[0075] 5. The temperature data predicted by this invention includes the bearing temperature at different locations, thereby improving the comprehensiveness of the prediction. Attached Figure Description

[0076] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0077] Figure 1 This is a flowchart of a method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network. Detailed Implementation

[0078] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0079] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] Example 1:

[0081] A method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network, the method comprising:

[0082] S1, collect historical data that may cause temperature changes in the guide bearings of the hydro-generator unit;

[0083] The historical data collected in this embodiment includes the cooling water flow rate of the generator guide bearing, the cooling oil flow rate, the inlet and outlet temperatures of the cooling water and oil, the generator's vibration data, the generator pit temperature, and the generator's active power, reactive power, and speed.

[0084] S2, data cleaning is performed on the collected historical data, and error data is removed and missing data is filled in;

[0085] Since the LSTM memory network is sensitive to time sequence, if a sudden mutation value or null value appears in the time sequence, it will weaken the time feature grabbing ability of the memory network and affect the prediction accuracy of the model. Therefore, it is necessary to reassign the error data and missing data to improve the performance of the prediction model. Historical data may have some missing data or data errors. In this embodiment, K-Nearest Neighbor (KNN) algorithm is used to find the substitute value of the missing value, that is, the average value of K samples is used to fill in the missing value.

[0086] The specific steps of K-Nearest Neighbor algorithm are as follows:

[0087] 1) Select m candidate samples, calculate the Euclidean distance between the missing sample and all samples in the candidate sample set, and the formula is as follows:

[0088]

[0089] Wherein, L represents the Euclidean distance, x and X represent the features of the missing sample and the candidate sample respectively, in this application, the features causing the temperature change of the guide bearing are selected: guide bearing cooling water flow, cooling oil flow, cooling water and oil inlet and outlet temperature, generator vibration data, machine pit temperature, active power, reactive power and rotating speed, i represents the number of features, in this application, i = 9;

[0090] 2) The m L values [L1, L2,..., Lm] calculated in 1) are arranged in order from small to large;

[0091] 3) Select K value, take out the temperature values of the candidate samples arranged in the first K L values, and calculate the temperature average value:

[0092]

[0093] Wherein, y is the substitute value to be found, and Y is the temperature value of the candidate sample.

[0094] S3, standardizing the historical data after cleaning;

[0095] The guide bearing temperature value and the guide bearing cooling water flow, cooling oil flow, cooling water and oil inlet and outlet temperature, generator vibration data, machine pit temperature, active power, reactive power and rotating speed causing the change thereof are standardized, and the standardization is specifically z-score standardization.

[0096] The formula of z-score standardization is as follows:

[0097]

[0098] where x * is the normalized value, x is the original value, μ and σ are the mean and standard deviation of all samples, respectively.

[0099] S4, the historical data obtained after the normalization is processed using empirical mode decomposition (EMD), and a series of IMFs and a linear superposition of the residual part are obtained by decomposition;

[0100] The historical journal bearing temperature data is taken as the original data sequence S(t) for EMD decomposition, and the decomposition steps are as follows:

[0101] 1) Find all the maximum points and minimum points of the original data sequence S(t), and use a cubic spline interpolation function to fit the upper envelope line and the lower envelope line of the original data;

[0102] 2) Calculate the mean of the upper envelope line and the lower envelope line to obtain the mean envelope line m1;

[0103] 3) Subtract the mean envelope line m1 from the original sequence S(t) to obtain the suspected IMF component h(t);

[0104] 4) Determine whether h(t) satisfies the two constraints of IMF, if it does, h(t) is an IMF component; if not, take h(t) as the original signal and repeat steps 1) to 4) until IMF1 is obtained;

[0105] The two constraints that the IMF component needs to satisfy are:

[0106] Condition 1: Within the entire IMF, the number of extreme points and the number of zero-crossing points must be equal or differ by at most one;

[0107] Condition 2: At any time, the average value of the upper envelope line formed by the local maximum points and the lower envelope line formed by the local minimum points is zero, i.e. the upper and lower envelope lines are locally symmetric with respect to the time axis.

[0108] 5) After obtaining IMF1 through steps 1) to 4), subtract IMF1 from the original signal S(t), S(t)-IMF1 as the new original signal, and repeat steps 1) to 4) to obtain IMF2, and so on. When the final remaining signal is a single-valued signal or a monotonic signal, stop repeating the above steps; at this time, the EMD is complete, and n IMFs and a residual data are obtained. The original signal S(t) is decomposed into a series of IMFs and a linear superposition of the residual part by the EMD method:

[0109]

[0110] Wherein, n is the number of IMF, r(t) is the residual part.

[0111] S5, each IMF and the residual part is input into the built long short-term memory neural network (LSTM) for training.

[0112] The rolling prediction is adopted in the application, and the historical data is used to predict future values, so the prepared sample data needs to be divided into two parts, the input of the training process is the parameters of the first part of the sample data that may cause temperature change, and the output is the bearing temperature value of the same time length of the second part of the sample data. Because the change frequency and nonlinearity degree of each IMF component are different, in order to accurately and completely extract the characteristics of each IMF, the structure of each LSTM model is different, and the LSTM modeling process is as follows:

[0113] 1) The hyperparameters of the LSTM neural network are set, including input dimension, output dimension, hidden layer dimension, and unit state dimension, wherein the hidden layer dimension is equal to the unit state dimension, wherein the input dimension is the number of parameters that may cause temperature change, and the application selects 9, including the cooling water flow of the guide bearing, the cooling oil flow, the inlet and outlet temperatures of the cooling water and oil, the generator vibration data, the machine pit temperature, the active power, the reactive power and the rotating speed, and the output dimension is the unit guide bearing temperature value at a certain time corresponding to the input dimension.

[0114] 2) The weight matrix and the bias matrix required for forward calculation of the LSTM are initialized, that is, W f , W i , W o , W c , b f , b i , b o and b c are initialized, which respectively represent the forgetting gate weight matrix, the input gate weight matrix, the output gate weight matrix, the calculation unit state weight matrix, the forgetting gate bias item matrix, the input gate bias item matrix, the output gate bias item matrix, and the calculation unit state bias item matrix of the LSTM neural network.

[0115] 3) The selected parameters that may cause temperature change, including the cooling water flow of the guide bearing, the cooling oil flow, the inlet and outlet temperatures of the cooling water and oil, the generator vibration data, the machine pit temperature, the active power, the reactive power and the rotating speed, are brought into the forgetting gate, the input gate, the output gate and the current unit state of the LSTM for neural network training.

[0116] f t is the forgetting gate calculation formula:

[0117] f t =σ(W f ·[h t-1 ,xt ]+b f ;

[0118] where σ is an activation function of the neural network, and h t-1 is the output matrix of the neural network at the last time;

[0119] i t is the input gate calculation formula:

[0120] i t = σ(W i · [h t-1 , x t ]+b i );

[0121] c t is the current time unit state calculation formula:

[0122]

[0123] where The calculation method is:

[0124]

[0125] The symbol in the formula represents element-wise multiplication, and tanh is another activation function,

[0126] o t is the output gate calculation formula:

[0127] o t = σ(W o · [h t-1 , x t ]+b o );

[0128] The final output h t of the neural network is determined by the output gate and the unit state:

[0129]

[0130] 4) After training, the error term of the neural network prediction value and the true value is calculated in reverse. The propagation of the LSTM neural network error has two directions, which are respectively the time reverse propagation and the propagation to the last layer of the neural network;

[0131] The output value obtained by the neural network at time t is h t , and the error term at this time is δ t T , which is propagated in time to any k time formula:

[0132]

[0133] wherein W fh , W fx , W ih , W ix , W oh , W ox , W ch , W cx is a weight matrix W f , W i , s, W c is the branch of the input dimension and the output dimension of the last time;

[0134] The error term of the neural network transmitted to the last layer is:

[0135]

[0136] wherein l is the current layer, f l-1 is the activation function of the l-1 layer, is the weighted input value of the l-1 layer network at time t;

[0137] 5) Calculate the gradient of each neuron connection weight W and bias term b, respectively as follows:

[0138]

[0139] wherein▽ is the gradient operator, and δ is the error obtained by the last step calculation;

[0140] 6) Update the connection weight W and the bias term b according to the gradient descent method, and then repeat steps 2) to 5) until the error reaches the minimum, that is, the model training is completed.

[0141] S6, select the parameters that may cause the temperature of the guide bearing to change at the current time, and input the normalized data into the trained LSTM model to obtain the normalized data of the predicted value of the guide bearing temperature in the future period of time. The true value of the predicted data can be obtained after the predicted data is denormalized.

[0142] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and any equivalent embodiments with equivalent changes are equivalent. Any modification, change and modification of the above embodiments according to the technical essence of the present application are still within the scope of the technical solution of the present application.

Claims

1. A method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network, characterized in that, The method includes: S1, collect historical data that may cause temperature changes in the guide bearings of the hydro-generator unit; S2, perform data cleaning on the collected historical data, remove erroneous data and fill in missing data; S3 standardizes the cleaned historical data. S4. The historical data obtained after standardization is processed using Empirical Mode Decomposition (EMD) to decompose it into a series of IMFs and a linear superposition of the remaining parts. S5, input each IMF and the remaining part into the constructed Long Short-Term Memory Neural Network (LSTM) for training; S6 collects data that may cause temperature changes in real time and inputs it into the model trained in S5 to predict the temperature of the guide bearing of the hydro-generator unit.

2. The method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network according to claim 1, characterized in that, The historical data in S1 includes the cooling water flow rate of the generator guide bearing, the cooling oil flow rate, the inlet and outlet temperatures of the cooling water and oil, the generator's vibration data, the generator pit temperature, and the generator's active, reactive, and speed data.

3. The method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network according to claim 1, characterized in that, In S2, the K nearest neighbor algorithm is used to find replacement values ​​for missing values, that is, the average of K samples is used to fill in the missing values.

4. The method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network according to claim 3, characterized in that, The specific steps of the K nearest neighbor algorithm in S2 are as follows: 1) Select m candidate samples and calculate the Euclidean distance between the missing sample and all samples in the candidate sample set, using the following formula: Where L represents the Euclidean distance, x and X represent the features of the missing sample and the candidate sample, respectively, and i represents the number of features; 2) Calculate m L values ​​[L1, L2, ..., Lm] from 1), and arrange these values ​​in ascending order; 3) Select a K value, extract the temperature values ​​of the top K candidate L values, and calculate the average temperature: Where y is the replacement value to be found, and Y is the temperature value of the candidate sample.

5. The method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network according to claim 4, characterized in that, In step S3, historical guide bearing temperature values ​​and historical data that caused their changes will be standardized. Specifically, z-score standardization will be used for the standardization process.

6. The method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network according to claim 5, characterized in that, The z-score standardization formula is as follows: Where, x * σ is the standardized value, x is the original value, and μ and σ are the mean and standard deviation of all samples, respectively.

7. The method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network according to claim 1, characterized in that, In step S4, the historical guide bearing temperature data is used as the original data sequence S(t) for EMD decomposition. The decomposition steps are as follows: 1) Find all the maximum and minimum points of the original data sequence S(t), and use cubic spline interpolation function to fit and form the upper and lower envelopes of the original data; 2) Calculate the mean of the upper and lower envelopes to obtain the mean envelope m1; 3) Subtract the mean envelope m1 from the original sequence S(t) to obtain the component h(t) of the suspected IMF; 4) Determine whether h(t) satisfies the two constraints of IMF. If it does, h(t) is an IMF component. If not, use h(t) as the original signal and repeat steps 1) to 4) until IMF1 is obtained. 5) After obtaining IMF1 through steps 1) to 4), subtract IMF1 from the original signal S(t). S(t) - IMF1 becomes the new original signal. Then, through the analysis in steps 1) to 4), obtain IMF2. Continue this process until the final remaining signal is a single-valued or monotonic signal. At this point, EMD is complete, yielding n IMF data points and one residual data point. The EMD method decomposes the original signal S(t) into a series of IMFs and a linear superposition of the remaining components. Where n is the number of IMFs and r(t) is the remaining part.

8. The method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network according to claim 7, characterized in that, The IMF components need to satisfy two constraints: Condition 1: Within the entire IMF, the number of extreme points and the number of zero-crossing points must be equal or differ by no more than one. Condition 2: At any given time, the average value of the upper envelope formed by the local maxima and the lower envelope formed by the local minima is zero, meaning that the upper and lower envelopes are locally symmetrical with respect to the time axis.

9. The method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network according to claim 7, characterized in that, The LSTM modeling process in S5 is as follows: 1) Set the hyperparameters of the LSTM neural network to include input dimension, output dimension, hidden layer dimension, and cell state dimension, where the cell state dimension is equal to the hidden layer dimension. The input dimension is the number of features that may cause temperature changes, and the output dimension is the temperature value of the unit guide bearing at a certain moment corresponding to the input dimension. 2) Initialize the weight matrix and bias matrix required for the LSTM forward computation, i.e., initialize W. f W i W o W c b f b i b o and b c Initialize them, which respectively represent the forget gate weight matrix, input gate weight matrix, output gate weight matrix, computation unit state weight matrix, forget gate bias term matrix, input gate bias term matrix, output gate bias term matrix, and computation unit state bias term matrix of the LSTM neural network; 3) The selected data that may cause temperature changes are fed into the LSTM's forget gate, input gate, output gate, and the current cell state for training the neural network; f t Formula for calculating the forget gate: f t =σ(W f ·[h t-1 ,x t ]+b f ); Where σ is the activation function of the neural network, and h t-1 This is the output matrix of the neural network at the previous time step; i t The formula for calculating the input gate is: c t Formula for calculating the cell state at the current moment: in The calculation method is as follows: In the formula, the symbol ○ indicates element-wise multiplication, and tanh is another activation function. o t The formula for calculating the output gate is as follows: the t =σ(W o ·[h t-1 ,x t ]+b o ); The final output h of the neural network t It is determined by both the output gate and the cell state: 4) After training, the error term between the predicted value and the true value of the neural network is calculated in reverse. The error of the LSTM neural network propagates in two directions: backward propagation along time and propagation to the next layer of the neural network. The output value of the neural network at time t is h. t Let the error term at this moment be... Its backward propagation to any time k is given by the following formula: Among them W fh W fx W ih W ix W oh W ox W ch W cx The weight matrix W f W i ,s,W c Branches in the input dimension and the output dimension of the previous time step; The error term passed from the neural network to the previous layer is: Where l represents the current layer, and f l-1 The activation function for layer l-1, The weighted input value at time t is the value of the (l-1)th layer network. 5) Calculate the gradients of the connection weights W and bias term b for each neuron, as follows: Where ▽ is the gradient operator, and δ is the error calculated in the previous step; 6) Update the connection weights W and bias term b according to the gradient descent method, and then repeat steps 2) to 5) until the error reaches the minimum, which means the model training is complete.

10. The method for predicting the temperature of the guide bearing of a hydro-generator unit based on an EMD-LSTM neural network according to claim 7, characterized in that, S6 specifically includes: selecting parameters that may cause changes in the bearing temperature at the current moment, normalizing them, and inputting them into the trained LSTM model to obtain normalized data of the predicted bearing temperature over a future period of time. After inverse normalizing the predicted data, the true value of the predicted data can be obtained.

Images