A large turbo-generator excitation current real-time prediction method

Abstract

The application discloses a large-sized steam turbine generator excitation current real-time prediction method, which takes an NSET model as a prediction model and comprises the following steps: calculating the grey correlation degree between each state variable monitored by a DCS system of the steam turbine generator and the excitation current by a grey correlation degree analysis method, and screening out key state variables capable of representing faults; constructing a memory matrix, dividing historical operation data into multiple intervals, carrying out equidistant sampling in each interval with the excitation current as a reference sequence, adding the sampling results to the memory matrix, and thus completing the memory matrix construction of the NSET model; inputting real-time state variables collected by the DCS system of the steam turbine generator into the NSET model, taking Manhattan distance as a nonlinear operator of the model, and completing the prediction of the excitation current; and the application has the advantage of high prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of generator rotor inter-turn insulation fault diagnosis, more particularly to a large steam turbine generator excitation current real-time prediction method. BACKGROUND

[0002] The large steam turbine generator rotor inter-turn insulation fault is a common fault, and the fault frequency has shown an upward trend in recent years. The rotor winding works in a harsh environment, especially when the large-capacity steam turbine generator is in normal operation, it needs to withstand the superposition of strong electric field, strong magnetic field and strong mechanical force, which causes the rotor to rub and extrude each other, easy to appear structure deformation, and then inter-turn short circuit fault occurs. It has been proved that the inter-turn insulation fault is accompanied by significant increase of excitation current, decrease of reactive power, increase of winding temperature, and severe vibration of the generator, and even more serious accidents such as rotor one-point or two-point grounding and large shaft magnetization. If the rotor inter-turn insulation fault occurs, it will lead to unplanned shutdown of the generator set, which brings great pressure to the normal operation of the power plant and causes serious economic loss. Therefore, high-precision prediction of the excitation current of the large steam turbine generator can accurately judge the current health status of the rotor winding, which has important practical significance for improving the operation reliability of the steam turbine generator, reasonably arranging the generator shutdown for maintenance, and ensuring the safety and economy of the power plant operation.

[0003] In the existing large steam turbine generator excitation current prediction method, the diagnosis method based on physical modeling is widely used, for example, Chinese patent publication No. CN105004962A discloses an improved detection method for short circuit fault of turbine generator excitation winding, obtains the generator parameters and characteristic curve, establishes a two-dimensional numerical simulation model of the generator, solves the excitation current value, compares the obtained excitation current value with the measured value, and judges whether a fault occurs. The patent application can improve the shortcomings of the prior art and improve the sensitivity of the excitation current method diagnosis, but the prediction accuracy is low and many generator parameters need to be obtained, which is difficult to realize. The excitation current prediction method based on data-driven is an effective prediction method, for example, Chinese patent publication No. CN103926506A discloses a turbine generator rotor winding short circuit fault diagnosis method based on function construction. This method is based on historical operation data to establish a prediction model of each key state variable when the rotor inter-turn insulation is in normal state, and uses the model to predict the excitation current of the rotor. Due to its clear principle and low modeling difficulty, it has been widely used. Although the data-driven method has obtained good prediction results, in actual application, the training process of the model may appear overfitting and underfitting, which affects the prediction accuracy.

[0004] Yanshan University in May 2016 master degree thesis "based on NSET model of wind turbine fault diagnosis research", it discloses the process of wind turbine fault diagnosis based on NSET model, first select training data, data normalization processing, then based on the method of mahalanobis distance optimization process memory matrix, test data normalization observation vector, finally fault prediction and determine the fault type. But its NSET model adopts the Euclidean distance as a nonlinear operator, the prediction accuracy is not high and its equidistant sampling to construct memory matrix, when the sample number is less, the model prediction accuracy is not enough. SUMMARY

[0005] The technical problem to be solved by the present application is the low prediction accuracy of the existing generator excitation current prediction method.

[0006] The present application solves the above technical problems by the following technical means: a large-scale steam turbine generator excitation current real-time prediction method, the method takes NSET model as the prediction model, including the following steps:

[0007] Step 1: Calculate the gray correlation degree of each state variable monitored by the DCS system of the steam turbine generator itself and the excitation current by the gray correlation degree analysis method, and select the key state variable that can represent the fault;

[0008] Step 2: Use the normal historical operation data collected by the generator itself under different operating conditions to construct the memory matrix, and divide the historical operation data into multiple intervals, and sample equidistantly in each interval with the excitation current as the reference sequence, and add the sampling results to the memory matrix, thereby completing the construction of the memory matrix of the NSET model;

[0009] Step 3: Input the real-time state variables collected by the DCS system of the steam turbine generator into the NSET model, and use Manhattan distance as the nonlinear operator of the model to complete the prediction of the excitation current.

[0010] The present application replaces the nonlinear operator in the traditional NSET model with Manhattan distance instead of Euclidean distance, improves the prediction accuracy of the model, and adds the interval sampling method on the basis of equidistant sampling in the construction of the memory matrix of the traditional NSET model, which can improve the model accuracy when the sample number is small without affecting the calculation efficiency of the model.

[0011] Further, the step 1 further includes the process of normalizing each state variable monitored by the DCS system.

[0012] Further, the normalization process is:

[0013] Based on the normal historical operating data of various state variables collected by the generator's own DCS system, the formula is used to... Normalization is performed, where x represents the actual state data. max x is the maximum value of this set of state variables. min Let x' be the minimum value of the set of state variables, and x′ be the normalized data.

[0014] Furthermore, the state variables include timestamp, stator current, stator voltage, excitation current, excitation voltage, active power, reactive power, and vibration displacement.

[0015] Further, step 1 includes:

[0016] Through formula Calculate the grey relational coefficients between the corresponding elements of each state variable sequence and the excitation current sequence, where i is the state variable sequence number, and ξ is the state variable sequence number. i (k) is the state variable sequence x i The correlation coefficient of the reference sequence x0 at time k; ρ is the resolution coefficient, taken as 0.5;

[0017] Through formula Calculate the grey relational degree.

[0018] Select state variable sequences whose gray correlation exceeds a preset value as key state variables.

[0019] Furthermore, in step 2, a memory matrix is ​​constructed using normal historical operating data collected by the generator itself under different operating conditions during operation, including:

[0020] 1) Obtain historical operating data under normal conditions to construct a training matrix

[0021] A generator has n distinct state variables. At any time k, the observation of these variables is represented by the observation vector matrix.

[0022] X(k)=[x 1k x 2k … x nk ] T

[0023] The training matrix K represents the operational data of each state variable collected by the generator under normal conditions. It is essential to ensure that the training matrix can completely represent the normal operating space of the entire generator. Therefore, the training matrix K is represented as K = [X(1), X(2), ..., X(n)].

[0024] 2) Constructing the memory matrix

[0025] Within the training matrix K, m distinct historical observation vectors under different operating states of the generator are extracted to construct a memory matrix D, where...

[0026]

[0027] Further, the different operating states include high load, low load, start-stop machine.

[0028] Further, in step 2, the historical operation data is divided into multiple intervals, and in each interval, the excitation current is taken as a reference sequence for equidistant sampling, and the sampling results are added to the memory matrix, thereby completing the memory matrix construction of the NSET model, including:

[0029] The historical operation data is arranged in ascending order of the normalized value of the excitation current, and is divided into five intervals;

[0030] According to the screening results of the key state variables, in each interval, the excitation current is taken as a reference sequence for equidistant sampling, and the calculation formula is as follows:

[0031] |I f (i)-A|<δ

[0032] Where, I f (i) is the excitation current of the i-th group of data, A is the step, and δ is the sampling point interval;

[0033] The sampling results satisfying the above formula are added to the memory matrix, thereby completing the memory matrix construction of the NSET model.

[0034] Further, in step 3, the Manhattan distance formula is d(i, j) = |X i -X j |+|Y i -Y j |, where d(i, j) is the sum of the horizontal and vertical displacements between two points, |X i -X j | is the horizontal displacement between two points, and |Y i -Y j | is the vertical displacement between two points.

[0035] Further, in step 3, the prediction of the excitation current includes:

[0036] The real-time state variables collected by the DCS system of the steam turbine generator are input into the NSET model, and the observation vector matrix collected by n different state variables is saved as X obs ; the output vector input into the NSET model is the prediction vector matrix X est at the same time, and the calculation formula is:

[0037] W=(D T ·D) -1 ·(DT ·X obs )

[0038] X est =D·W=D·(D T ·D) -1 ·(D T ·X obs )

[0039] Wherein, W is a weight vector matrix.

[0040] The present application has the advantages that: the present application improves the prediction accuracy of the model by replacing the nonlinear operator in the traditional NSET model with the Manhattan distance instead of the Euclidean distance, and on the basis of using equidistant sampling in the memory matrix construction of the traditional NSET model, the interval sampling method is added, which can improve the model accuracy when the sample quantity is small, and does not affect the calculation efficiency of the model. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 A flow chart of a large steam turbine generator excitation current real-time prediction method provided by the embodiment of the present application;

[0042] Figure 2 The excitation current actual value and prediction value curve of the large steam turbine generator excitation current real-time prediction method provided by the embodiment of the present application under the sharp fluctuation working condition;

[0043] Figure 3 The excitation current prediction residual curve of the large steam turbine generator excitation current real-time prediction method provided by the embodiment of the present application under the sharp fluctuation working condition;

[0044] Figure 4 The excitation current actual value and prediction value curve of the large steam turbine generator excitation current real-time prediction method provided by the embodiment of the present application under the smooth running working condition;

[0045] Figure 5 The excitation current prediction residual curve of the large steam turbine generator excitation current real-time prediction method provided by the embodiment of the present application under the smooth running working condition;

[0046] Figure 6 The excitation current prediction residual schematic diagram in the large steam turbine generator excitation current real-time prediction method provided by the embodiment of the present application based on different nonlinear operators;

[0047] Figure 7 The excitation current prediction residual schematic diagram in the large steam turbine generator excitation current real-time prediction method provided by the embodiment of the present application based on different memory matrix construction methods. DETAILED DESCRIPTION

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. 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.

[0049] like Figure 1 As shown, a real-time prediction method for the excitation current of a large steam turbine generator is presented. The method uses the NSET model as the prediction model and includes the following steps:

[0050] Step 1: Calculate the grey relational analysis method to determine the grey relational degree between the state variables monitored by the DCS system of the turbine generator and the excitation current, and screen out the key state variables that can characterize the fault. In this embodiment, the generator used is a 50WT23E-138 turbine generator from the Dabieshan Power Plant in Hubei Province, and its operating parameters are shown in Table 1.

[0051] Table 1 Generator Parameters

[0052]

[0053] The turbine generator's own DCS system records real-time operating data of various state variables measured by each sensor every 5 minutes, confirming that the generator was in normal operating condition during the period of collecting normal historical operating data. Important state variables include, but are not limited to, timestamps, stator current, stator voltage, excitation current, excitation voltage, active power, reactive power, vibration displacement, and other parameters. The collected state variable data is normalized using the following formula: Where x represents the actual state data, x max x is the maximum value of this set of state variables. min Let x' be the minimum value of the set of state variables, and x′ be the normalized data.

[0054] Then, through the formula Calculate the grey relational coefficients between the corresponding elements of each state variable sequence and the excitation current sequence, where i is the state variable sequence number, and ξ is the state variable sequence number. i (k) is the state variable sequence x i The correlation coefficient of the reference sequence x0 at time k; ρ is the resolution coefficient, taken as 0.5;

[0055] Next, through the formula Calculate the grey relational degree.

[0056] Finally, state variable sequences with a gray correlation degree exceeding a preset value are selected as key state variables.

[0057] In this embodiment, the excitation current I f The grey correlation degrees of the stator current I, the stator voltage U, the excitation voltage U f , the active power P, the reactive power Q, the rotor x-direction vibration displacement V x and the rotor y-direction vibration displacement V y are calculated, and the calculation results are shown in Table 2.

[0058] Table 2 Grey correlation degrees

[0059]

[0060] The state quantities with the correlation degrees ∈ [0.75, 1] are screened as the key state quantities, so I f , U f , U, I, P, and Q are selected as the key state quantities.

[0061] Step 2: The memory matrix is constructed using the normal historical operation data collected by the generator itself under different working conditions during operation, and the historical operation data is divided into multiple intervals, and the excitation current is used as the reference sequence for equidistant sampling in each interval. The sampling results are added to the memory matrix, so as to complete the construction of the memory matrix of the NSET model. The specific process is as follows:

[0062] 1) Obtain the historical operation data under the normal state to construct the training matrix

[0063] There are n different state variables in a generator, and at any time i, the observation thereof is represented as an observation vector matrix

[0064] X(k) = [x 1k x 2k … x nk ] T

[0065] The training matrix K is the running data of each state variable collected by the generator under the normal state, and it must be ensured that the training matrix can completely represent the normal working space of the entire generator. Therefore, the training matrix K is represented as K = [X(1), X(2), …, X(n)]

[0066] 2) Construct the memory matrix

[0067] m different historical observation vectors of the generator under different running states (including high load, low load, start-stop machine, etc.) are extracted from the training matrix K to construct the memory matrix D, wherein

[0068]

[0069] The historical operation data is arranged in ascending order of the normalized value of the field current, and is divided into five intervals;

[0070] According to the screening result of the key state quantity, the field current is taken as the reference sequence for equidistant sampling in each interval, and the calculation formula is as follows:

[0071] |I f (i)-A|<δ

[0072] Wherein, I f (i) is the field current of the i th group of data, A is the step, and δ is the sampling point interval;

[0073] The sampling results meeting the above formula are added to the memory matrix, so as to complete the memory matrix construction of the NSET model.

[0074] Step 3: input the real-time state variables collected by the DCS system of the steam turbine generator into the NSET model, take Manhattan distance as the nonlinear operator of the model, complete the prediction of the field current, and the specific process is as follows:

[0075] Manhattan distance is also called city block distance, which represents the sum of horizontal and vertical displacement between two points, and does not involve the displacement in diagonal direction. The Manhattan distance formula is d (i, j) = |X i -X j |+|Y i -Y j |, wherein d (i, j) is the sum of horizontal and vertical displacement between two points, |X i -X j | is the horizontal displacement between two points, and |Y i -Y j | is the vertical displacement between two points.

[0076] The real-time state variables collected by the DCS system of the steam turbine generator are input into the NSET model, and the observation vector matrix collected by n different state quantities is saved as X obs ; The output vector of the input NSET model is the prediction vector matrix X est at the same time, and the calculation formula is as follows:

[0077] W=(D T ·D) -1 ·(D T ·X obs )

[0078] X est =D·W=D·(D T ·D) -1 ·(D T ·X obs )

[0079] wherein, W is a weight vector matrix.

[0080] The simulation verification of the embodiment of the application proves the prediction effect of the model under the sharp fluctuation working condition and the smooth running working condition, and the excitation current analytical prediction model and the BP neural network excitation current prediction model are selected for comparison to further prove the effectiveness of the application. As shown in Fig. 2, it is the predicted value of each model obtained by inputting the test set under the sharp fluctuation working condition; as shown in Fig. 3, it is the excitation current prediction residual of each model under the working condition. Since the predicted value of the excitation current analytical prediction model is too large from the actual value of the excitation current, it is far from reaching the standard of accurate prediction of the excitation current, and has no comparability, so it is not embodied in Fig. 3. RMSE, MAPE and MAE are selected as the evaluation indexes of the model, and the evaluation results are shown in Table 3. Figure 2 Figure 3 Figure 3

[0081] Table 3 Evaluation results of each prediction model under the sharp fluctuation working condition

[0082]

[0083] As shown in Fig. 4, it is the predicted value of each model obtained by inputting the test set under the smooth running working condition, and as shown in Fig. 5, it is the corresponding curve of the actual value in Fig. 4 and the corresponding curve of the improved NSET model. The curves of the actual value in Fig. 4 and Fig. 5 are highly coincident, so the curve of the actual value in Fig. 5 is not shown; as shown in Fig. 6, it is the excitation current prediction residual of each model under the working condition. Since the predicted value of the excitation current analytical prediction model is too large from the actual value of the excitation current, it is far from reaching the standard of accurate prediction of the excitation current, and has no comparability, so it is not embodied in Fig. 6. The evaluation results are respectively: Figure 4 Figure 2 Figure 4 Figure 2 Figure 4 Figure 5 Figure 5

[0084] Table 4 Evaluation results of each prediction model under the smooth running working condition

[0085]

[0086] From the excitation current prediction results under the above two different working conditions, it can be seen that the large-scale turbine generator excitation current prediction method based on the NSET model proposed in the application has good prediction accuracy, thereby proving the feasibility of the application. From the table, it can also be known that the prediction effect of the application is better than that of the excitation current analytical prediction model and the BP neural network.

[0087] ​​​​​​​​​​In the present application, since it involves multi-dimensional variables, the units of each variable are different, and when using Euclidean distance, the differences in the units of the variables cannot be considered, leading to possible distortion of the measurement; while the Manhattan distance calculation process has a standardization process, and thus is independent of the units of the variables, and can fully consider the relationship between the variables. Therefore, the Manhattan distance can be understood as a modification of the Euclidean distance. Based on this theoretical analysis, the inventors also conducted a comparative study, as follows:

[0088] (1) Nonlinear operator optimization

[0089] Randomly extract single-day historical operation time series data on March 2015, select three different nonlinear operators for field current prediction, and compare the prediction results. The prediction residual of each nonlinear operator is shown in Figure 6 .

[0090] The model evaluation results based on different nonlinear operators are shown in Table 5. From the evaluation results, it can be seen that the root mean square error of the field current prediction model based on Manhattan distance is 0.9676 A, the maximum absolute error is 5.9514 A, and the average absolute percentage error is 2.22%, which are all lower than the prediction results obtained by the other two nonlinear operators. In summary, Manhattan distance is preferred as the nonlinear operator of the field current prediction model.

[0091] Table 5 Evaluation results of field current prediction model based on different nonlinear operators

[0092]

[0093] (2) Memory matrix construction method improvement

[0094] Through the optimization of the nonlinear operator, the root mean square error and the maximum absolute error of the prediction model established using Manhattan distance have been greatly improved compared to the traditional NSET model using Euclidean distance as the nonlinear operator. However, to achieve early warning of rotor turn insulation faults, the prediction accuracy of the prediction model still needs to be further improved. Therefore, the present application proposes a method of segmenting memory matrix data, which increases the number of samples while not affecting the overall calculation efficiency of the model. The specific implementation is to divide the sampling interval of the memory matrix equally, and to sample each sub-sampling interval equally, i.e. each memory matrix will have 2000 historical data records under normal operating conditions. By comparing the effects of the field current prediction models constructed by different memory matrix segmentation methods, the number of memory matrix segments used by the model can be determined.

[0095] The inventors compared the prediction effects of the memory matrix without segmentation (D1), segmented into two (D2), segmented into four (D4), segmented into five (D5), segmented into eight (D8) and segmented into ten (D10), the nonlinear operator adopts Manhattan distance, the data for prediction is the same as the data selected when performing nonlinear operator optimization in the foregoing, and the prediction residual is as shown in Table 1. Figure 7

[0096] The model evaluation results based on different memory matrix construction methods are shown in Table 6.

[0097] Table 6: Evaluation results of the excitation current prediction model based on different memory matrix construction methods

[0098]

[0099] Through the above technical solutions, the present application replaces the nonlinear operator in the traditional NSET model from Euclidean distance to Manhattan distance, improves the prediction accuracy of the model, and adds the interval sampling method on the basis of equal interval sampling in the construction of the memory matrix of the traditional NSET model, which can improve the model accuracy when the number of samples is small, and does not affect the calculation efficiency of the model.

[0100] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.​

Claims

1. A large turbo-generator excitation current real-time prediction method, characterized by, The method takes the NSET model as a prediction model, and comprises the following steps: Step 1: Calculate the grey correlation degree of each state variable monitored by the DCS system of the steam turbine generator and the excitation current by the grey correlation degree analysis method, and screen out the key state variables that can represent the fault; calculate the grey correlation degree of each state variable sequence and the corresponding elements of the excitation current sequence by the formula The grey correlation degree of each state variable sequence and the corresponding elements of the excitation current sequence is calculated respectively, wherein, i is the state variable sequence number, is the state variable sequence The correlation coefficient of the reference sequence at the moment k ; is the resolution coefficient, which is 0.5; The grey correlation degree is calculated by the formula G=∑|Xi-Xj| Screening state quantity sequences with grey correlation degrees exceeding a preset value as key state quantities; Step 2: constructing a memory matrix using normal historical operation data collected by the generator itself under different operating conditions in the operation process, dividing the historical operation data into multiple intervals, and performing equidistant sampling in each interval with the field current as a reference sequence, and adding the sampling results to the memory matrix, thereby completing the construction of the memory matrix of the NSET model; Step 3: input the real-time state variable collected by the DCS system of the steam turbine generator into the NSET model, take Manhattan distance as the nonlinear operator of the model, and complete the prediction of the field current; the prediction of the field current includes: input the real-time state variable collected by the DCS system of the steam turbine generator into the NSET model, n The observation vector matrix collected by the different state variables is saved as X obs The output vector of the NSET model is the prediction vector matrix at the same time X est , and the calculation formula is: wherein W is a weight vector matrix, and D is a memory matrix.

2. The method of claim 1, wherein the method is characterized by: The step 1 further comprises a process of normalizing each state variable monitored by the DCS system.

3. The method of claim 2, wherein the method is characterized by: The process of normalization comprises: Based on the normal historical operation data of each state variable collected by the DCS system of the generator, the formula is used for normalization, wherein, is the actual state variable data, is the maximum value of the group of state variables, is the minimum value of the group of state variables, is the normalized data.

4. The method of claim 3, wherein the method is characterized by, The state variables comprise a timestamp, a stator current, a stator voltage, a field current, a field voltage, active power, reactive power and vibration displacement.

5. The method of claim 1, wherein the method is characterized by: The step 2 of constructing the memory matrix using the normal historical operation data collected by the generator itself under different operating conditions in the operation process comprises: 1) obtaining historical operation data under a normal state to construct a training matrix There are n different state variables in a generator, the observation of which at any time k is represented as an observation vector matrix training matrix K For the operation data of each state variable collected by the generator in the normal state, it is necessary to ensure that the training matrix can completely represent the normal working space of the entire generator, so the training matrix K is expressed as 2) constructing a memory matrix In training matrix K The different operating states of the generator are extracted from the internal m A memory matrix D is constructed from the different historical observation vectors, wherein, 。 6. The method of claim 5, wherein the method is characterized by: The different operating conditions comprise high load, low load, start and stop.

7. The method of claim 5, wherein the method is characterized by, The step 2 of dividing the historical operation data into multiple intervals, performing equidistant sampling in each interval with the field current as a reference sequence, and adding the sampling results to the memory matrix, thereby completing the construction of the memory matrix of the NSET model, comprises: Arranging the historical operation data in ascending order of the normalized value of the field current as a reference, and dividing the historical operation data into five intervals; According to the screening result of the key state quantities, equidistant sampling is performed in each interval with the field current as a reference sequence, and the calculation formula is as follows: wherein, is the first i group of data, A is the step size, is the sampling point interval; The sampling results satisfying the above formula are added to the memory matrix, thereby completing the construction of the memory matrix of the NSET model.

8. The method of claim 7, wherein the method is characterized by, The Manhattan distance formula in step 3 is wherein, d ( i , j ) is the sum of the lateral and longitudinal displacements between two points, is the lateral displacement between two points, is the longitudinal displacement between two points.

Images