Hydroelectric generating set lifting amount prediction method based on LSTM and SVM combined prediction

Abstract

The invention discloses a hydroelectric generating set lifting amount prediction method based on LSTM and SVM combined prediction. The method comprises the steps of collecting working condition data of a hydroelectric generating set; preprocessing the obtained working condition data to form a working condition data set; constructing an LSTM (Long Short Term Memory) machine lifting prediction model and an SVM (Support Vector Machine) machine lifting prediction model respectively by adopting a long short-term memory neural network model and a support vector machine model, and training and testing the LSTM machine lifting prediction model and the SVM machine lifting prediction model respectively by utilizing the working condition data set; respectively determining the weight coefficients of the LSTM machine lifting prediction model and the SVM machine lifting prediction model for the combined prediction result of the unit machine lifting amount; and according to the weight coefficient obtained by combining the real-time hydroelectric generating set working condition data, a final unit lifting amount prediction result is obtained. According to the method, the LSTM and the SVM are combined to predict the unit lifting amount, the adaptability to different operation states is effectively improved, the simulation accuracy of the machine lifting phenomenon is improved, the problem of mathematical model construction is solved, and research scientificity and reliability are improved.

Description

Technical Field

[0001] The invention belongs to the field of hydropower unit fault diagnosis and discloses a method for predicting the lifting capacity of a hydropower unit based on LSTM and SVM combined prediction. Background Art

[0002] In the context of today's energy structure accelerating its transformation to green and low-carbon, hydropower, as a clean, renewable energy with good peak-shaving and frequency-regulating capabilities, occupies an important position in the global power supply system. As the core equipment in the hydropower field, the unit capacity of large hydropower units continues to rise, and their operational stability and reliability are becoming more and more critical, which is directly related to the safe and stable operation of the power system and the efficient supply of clean energy.

[0003] In recent years, with the rapid development of hydropower technology, large-scale hydropower units with a capacity of one million kilowatts have gradually been put into actual operation, providing stronger support for the power system. However, during the operation of such units, the lifting problem has become an important hidden danger affecting their safe and stable operation. The lifting phenomenon is caused by the axial force caused by the water hammer phenomenon in the tailwater pipe and the pump lift generated by the rotation of the runner blades, which causes the unit to lift upward under certain working conditions. Once the lifting failure occurs, it will not only cause serious damage to the key components of the unit, such as thrust bearings, guide bearings, sealing devices, etc., greatly shorten the service life of the unit, increase maintenance costs and downtime, but also may trigger a chain reaction, threatening the safe operation of the entire power system, leading to large-scale power outages, and bringing immeasurable economic losses and social impacts.

[0004] In the past, research on the problem of lifting the machine mainly focused on theoretical analysis and physical model construction. The lifting of the machine can be prevented by replenishing air in the turbine chamber and tailwater pipe of the unit main shaft to improve the water flow conditions, but this method relies on manual experience or delayed pressure pulsation control, and it is difficult to respond effectively in time at the early stage of the risk of lifting the machine. In terms of flow channel mechanics analysis, although the use of CFD meshing and simulation calculations can study the stress conditions of the lifting process, there are many complex and difficult to accurately simulate factors in actual operation, resulting in a large deviation between theory and practice. In addition, measures such as optimizing the segmented closing law of the guide vanes, the PID parameters of the speed regulator, and the design and calculation of the adjustment performance and protection strategy in the unit design stage are all based on ideal conditions. It is impossible to fully consider various accidental factors during the operation and installation of the unit, making it difficult to accurately prevent the occurrence of lifting failures in actual production. At the same time, the lifting indicators are affected by the interaction of multiple complex factors, and it is difficult to establish an accurate mathematical model. The early signs of failure are not obvious, which further increases the difficulty of fault prediction and prevention. Summary of the invention

[0005] The objective of the present invention is to construct a prediction model for the upward displacement of the turbine using a combination of LSTM and SVM, and to accurately train and predict the upward displacement during key transition processes such as the startup, shutdown, and load rejection of the unit. By fully leveraging the advantages of the LSTM network in processing time series data and the excellent performance of the SVM algorithm in data classification and anti-overfitting, the limitations of traditional BiLSTM and LSTM algorithms are overcome, and a prediction model for the upward displacement with higher prediction accuracy and capable of accurately fitting the true value under different operating states is constructed, significantly improving the adaptability and reliability of the model, and comprehensively and effectively solving the technical problems related to the mathematical model construction, prediction accuracy and adaptability, and pre-fault monitoring and early warning of the upward displacement of large hydropower units.

[0006] The technical solution provided by the present invention is a method for predicting the upward displacement of a hydropower unit based on combined prediction of LSTM and SVM, including the following steps:

[0007] Step 1: Collect the operating condition data of the hydropower unit, conduct a correlation analysis, and determine the characteristic parameters for predicting the upward displacement.

[0008] Step 2: Preprocess the operating condition data obtained in Step 1 to form an operating condition data set, which is divided into a training set, a validation set, and a test set.

[0009] Step 3: Construct an LSTM upward displacement prediction model and an SVM upward displacement prediction model using a long short-term memory neural network model and a support vector machine model respectively, and use the operating condition data set in Step 2 to train and validate the LSTM upward displacement prediction model and the SVM upward displacement prediction model respectively.

[0010] Step 3.1: Construct an LSTM upward displacement prediction model and an SVM upward displacement prediction model using a long short-term memory neural network model and a support vector machine model respectively, and use the operating condition data set in Step 2 to train the LSTM upward displacement prediction model and the SVM upward displacement prediction model respectively. The input of the LSTM upward displacement prediction model and the SVM upward displacement prediction model is the preprocessed operating condition data set in Step 2, and the output is the predicted upward displacement result.

[0011] Step 3.2: Select the validation set in Step 2 to validate the LSTM upward displacement prediction model and the SVM upward displacement prediction model obtained in Step 3.1 respectively. If the prediction effect of the model on the validation set is unqualified, adjust the hyperparameters of the corresponding model, and execute Step 3.1 again to retrain the current model. If the validation is passed, execute Step 3.3.

[0012] Step 3.3: After passing the validation on the validation set, obtain the optimal LSTM upward displacement prediction model and SVM upward displacement prediction model respectively.

[0013] Step 4: Determine the weight coefficients of the combined prediction results of the LSTM unit lifting prediction model and the SVM unit lifting prediction model for the unit lifting amount respectively;

[0014] Step 5: Respectively use the trained LSTM unit lifting prediction model and SVM unit lifting prediction model to obtain the real-time LSTM unit lifting prediction value and SVM unit lifting prediction value according to the real-time hydropower unit condition data, and perform weighted summation in combination with the weight coefficients obtained in Step 4 to obtain the final unit lifting amount prediction result.

[0015] Further, in the above Step 1, the hydropower unit condition sample data is obtained through acquisition points, the control unit LCU, and on-site measurement. The hydropower unit conditions include starting up, shutting down, and load rejection.

[0016] Preferably, in Step 1, the characteristic parameters for unit lifting amount prediction include the unit head C 1 , active power C 2 , guide vane opening C 3 , unit speed C 4 , +X direction swing at the upper guide bearing C 5 , +Y direction swing at the upper guide bearing C 6 , +X direction swing at the water guide bearing C 7 , +Y direction swing at the water guide bearing C 8 , +X direction swing at the lower guide bearing C 9 , +Y direction swing at the lower guide bearing C 10 , unit volute inlet pressure C 11 , volute end vacuum pressure C 12 , volute flow pressure difference C 13 , draft tube elbow pressure C 14 , draft tube inlet pressure C 15 , pressure between runner and guide vane C 16 and pressure between draft tube C 17 .

[0017] Preferably, in Step 2, the preprocessing includes removing distortion, filling missing values, and noise reduction for historical data, removing untrustworthy data, and forming a training set, a validation set, and a test set according to a ratio of 7:2:1.

[0018] Further, the long short-term memory network LSTM model includes memory units, forget gates, input gates, and output gates. The forget gate determines to forget the historical information with low relevance from the memory state information, the input gate determines the sensitivity of the memory unit state to the input data, and the output gate determines the information output at this moment.

[0019] The specific process of obtaining the prediction result by using the support vector machine SVM model includes:

[0020] 1) Map the input data to a high-dimensional feature space using a kernel function, where the kernel function includes a linear kernel, a polynomial kernel, and a radial basis function kernel;

[0021] 2) Determine the optimal hyperplane to maximize the functional margin of the samples closest to the plane;

[0022] 3) Introduce slack variables to handle noise in the input data and the case where the data is linearly inseparable;

[0023] 4) Solve through the dual problem to reduce the computational complexity;

[0024] 5) Make predictions through the optimal hyperplane, and the prediction result is the projection value of the sample on the hyperplane.

[0025] Step 4: Determine the weight coefficients of the combined prediction results of the LSTM liftoff prediction model and the SVM liftoff prediction model for the unit liftoff amount respectively;

[0026] Furthermore, let the weight of the LSTM liftoff prediction model be Q LSTM , and the error of the LSTM liftoff prediction model be E LSTM , and the weight of the SVM liftoff prediction model be Q SVM , and the error of the SVM liftoff prediction model be E SVM ;

[0027] Determine the weight coefficients through E LSTM and E SVM , and set a threshold θ;

[0028] When , it indicates that the prediction results of the two models have a large gap:

[0029] If E LSTM > E SVM Then let Q LSTM = 0, Q SVM = 1, that is, take the prediction result of the SVM liftoff prediction model as the final prediction result of the unit liftoff amount;

[0030] If E LSTM < E SVM Then let Q SVM = 0, Q LSTM = 1, that is, take the prediction result of the LSTM liftoff prediction model as the final prediction result of the unit liftoff amount;

[0031] Otherwise, determine the weight coefficients according to the error E LSTM of the LSTM liftoff prediction model and the error E SVM of the SVM liftoff prediction model, and the formula is as follows:

[0032]

[0033] Preferably, the calculation formula for the predicted result of the final unit lifting amount is as follows:

[0034] Y = Y LSTM .Q LSTM + Y SVM .Q SVM ; (4)

[0035] In the formula, Y LSTM is the LSTM lifting prediction value Y LSTM , Y SVM is the SVM lifting prediction value, Q LSTM is the weight of the LSTM lifting prediction model, Q SVM is the weight of the SVM lifting prediction model.

[0036] Preferably, the lifting amount prediction method further includes evaluating the combined prediction result of the unit lifting amount by selecting the root mean square error RMSE, mean absolute error MAE, mean square error MSE, and goodness of fit R2.

[0037] Compared with the prior art, the beneficial effects of the present invention include:

[0038] (1) The present invention predicts the unit lifting amount through the combination of LSTM and SVM, giving full play to the advantages of the LSTM network in processing time series data and the performance of the SVM algorithm in data classification and anti-overfitting. It can accurately predict the lifting amount during the key transition processes of unit startup, shutdown, and load rejection, and can issue early warnings at the initial stage of the lifting risk, solving the problem of untimely response of traditional methods and achieving early accurate warning. It provides strong technical support for ensuring the safe and stable operation of the unit, reduces the operation risk caused by inaccurate lifting prediction, and ensures that the prediction performance meets the high requirements of technical standards.

[0039] (2) The present invention uses real operation data for model training and prediction, and obtains associated characteristic parameters through correlation analysis. It can comprehensively consider various complex factors in actual operation, providing a practical new approach for the prevention and control of lifting faults of large hydropower units, significantly improving the intelligent operation and maintenance level of the hydropower industry, and ensuring the safe and stable operation of the power system. It greatly improves the simulation accuracy of the lifting phenomenon, provides a solid and reliable basis for subsequent in-depth analysis and prediction, effectively solves the long-term problem of mathematical model construction, significantly improves the scientificity and reliability of the research, and makes the model construction process more standardized and accurate.

[0040] (3) The present invention separately constructs an LSTM unit-lifting prediction model and an SVM unit-lifting prediction model, and also innovatively determines the weight coefficients of the combined prediction results of the two models for the unit-lifting amount of the unit. In this way, the powerful advantage of the LSTM network in processing time series data and the excellent performance of the SVM algorithm in data classification and anti-overfitting can be fully integrated. Compared with a single model or a traditional simple combination method, the weights can be dynamically adjusted according to different operating conditions and data characteristics, so that the final predicted result of the unit-lifting amount can be closer to the true value in various complex situations, effectively improving the reliability of the prediction and the adaptability to different operating states, and providing more powerful support for the accurate prevention of unit-lifting faults of large hydropower units.

[0041] (4) The combined prediction model of LSTM and SVM proposed by the present invention shows higher prediction accuracy compared with traditional algorithms such as BiLSTM and LSTM. It can accurately predict the unit-lifting amount under different operating states, with a small gap from the true value, effectively overcoming the problem of the disconnection between theory and practice of traditional methods and improving the accuracy of unit-lifting amount prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The present invention will be further described below in conjunction with the drawings and embodiments.

[0043] Figure 1 It is a schematic flow chart of the method for predicting the unit-lifting amount of a hydropower unit according to an embodiment of the present invention.

[0044] Figure 2 It is a logical structure diagram of the method for predicting the unit-lifting amount of a hydropower unit according to an embodiment of the present invention.

[0045] Figure 3 It is a schematic diagram of the layout of the unit condition data acquisition points of a hydropower unit according to an embodiment of the present invention.

[0046] Figure 4 It is a change curve graph of unit-lifting, guide vane opening and rotational speed during the startup process according to an embodiment of the present invention.

[0047] Figure 5 It is a change curve graph of unit-lifting, guide vane opening and rotational speed during the shutdown process according to an embodiment of the present invention.

[0048] Figure 6 It is a change curve graph of unit-lifting, guide vane opening and rotational speed during the load rejection process according to an embodiment of the present invention.

[0049] Figure 7 It is a comparison curve graph of real-time prediction evaluation indexes during the startup process according to an embodiment of the present invention.

[0050] Figure 8 It is a comparison curve graph of real-time prediction evaluation indexes during the shutdown process according to an embodiment of the present invention.

[0051] Figure 9 The contrast curve graph of real-time prediction evaluation indexes during the load rejection process according to the embodiment of the present invention.

[0052] Figure 10 The contrast curve graph of real-time prediction evaluation indexes during the starting process according to the embodiment of the present invention.

[0053] Figure 11 The contrast curve graph of real-time prediction evaluation indexes during the shutdown process according to the embodiment of the present invention.

[0054] Figure 12 The contrast curve graph of real-time prediction evaluation indexes during the load rejection process according to the embodiment of the present invention. Specific implementation manners

[0055] As Figure 1 and Figure 2 shown, the prediction method for the lifting amount of a hydropower unit based on the combined prediction of LSTM and SVM includes the following steps:

[0056] Step 1: Collect the working condition data of the hydropower unit and obtain the characteristic parameters for predicting the lifting amount by using correlation analysis;

[0057] In this embodiment, a large hydropower unit in the Jinsha River Basin is taken as an example, and the parameters of the unit are shown in Table 1;

[0058] Table 1 Parameters of a certain million-kilowatt hydropower unit

[0059] Parameter Name Value Parameter Name Value Maximum Water Ice 243.1m Rated Output 1000MW Minimum Water Head 163.9m Rated Speed 111.1r / min Rated Water Head 202m Runaway Speed 202r / min Suction Lift -13m Runner Diameter 8.482m

[0060] As Figure 3 shown, the working condition data is obtained through the acquisition points, the control unit LCU and on-site measurement, and the working conditions of the hydropower unit include starting, shutdown and load rejection.

[0061] As Figure 4 shown, in the on-site measurement curve during starting, first the guide vane opening rapidly rises and then remains unchanged. After the unit speed approaches the rated speed, the guide vane closes. This is to improve the starting speed. First, the guide vane opening is rapidly increased, and when the unit speed approaches the rated speed, to prevent the unit from overspeeding and the "reverse regulation phenomenon", the guide vane is rapidly closed. It is found in the figure that during the starting process, at the moment when the guide vane opens, the lifting amount of the unit rapidly rises from zero. After the guide vane closes, the lifting amount continues to reach the peak. When the unit continues to increase the load, the lifting amount slightly callbacks and remains. This process shows that although the reason for the generation of the lifting is very complex, its change process can be characterized by the guide vane opening.

[0062] As Figure 5As shown, during the shutdown process, the guide vane closes rapidly, and the jacking-up amount has obvious hysteresis and fluctuations with the unit speed. First, the unit speed drops in two stages. In the first stage, the unit speed decreases from the rated value with the decrease of the guide vane opening until it reaches about 10%. In the second stage, braking is applied when the speed is 10%, and the speed accelerates to drop to 0. During this whole process, the change of the jacking-up amount is divided into three stages. In the first stage, starting with the closing action of the guide vane, the jacking-up amount quickly jumps to a platform and remains. In the second stage, when the speed drops to about 30%, the jacking-up amount increases in the reverse direction and remains. In the third stage, after braking is applied, the jacking-up amount jumps to 0.

[0063] As Figure 6 shown, when the load is shed, the unit is disconnected, and the guide vane closes quickly to avoid the unit from running away. After the load is shed, the unit speed first rises and then drops to the rated value. The jacking-up amount drops rapidly with the closing of the guide vane opening, and after the guide vane is fully closed, it does not drop to zero. Subsequently, as the unit speed stabilizes, the guide vane opening gradually opens to the no-load opening, and the jacking-up amount also increases and remains at a low value. Through comparison, it is found that both the guide vane opening and the speed affect the jacking-up during the load shedding process.

[0064] The characteristic parameters for predicting the jacking-up amount are obtained through the Pearson correlation coefficient r value in the correlation analysis for predicting the jacking-up amount;

[0065] The calculation formula for the Pearson correlation coefficient r value is:

[0066]

[0067] In the formula, x t represents the value of a characteristic parameter, and y t represents the value of the jacking-up amount, represents the average value of a characteristic parameter, the average value of the jacking-up amount.

[0068] There are many factors affecting the jacking-up. When the model prediction is carried out in the present invention, 17 characteristic parameters associated with the jacking-up are excavated. The meanings and representative symbols of the characteristic parameters are summarized in Table 2.

[0069] Table 2 Characteristic parameters associated with jacking-up

[0070] Symbol Related Factor Name Symbol Related Factor <![CDATA[C 1 > Unit Water Head <![CDATA[C 9 > Lower Guide + X-directional Swing <![CDATA[C 2 > Active Power <![CDATA[C 10 > Lower Guide + Y-directional Swing <![CDATA[C 3 > Guide Vane Opening <![CDATA[C 11 > Unit Spiral Case Inlet Pressure <![CDATA[C 4 > Unit Speed <![CDATA[C 12 > Vacuum Pressure at the End of the Spiral Case <![CDATA[C 5 > Upper Guide + X-directional Swing <![CDATA[C 13 > Spiral Case Flow Differential Pressure <![CDATA[C 6 > Upper Guide + Y-directional Swing <![CDATA[C 14 > Bend Elbow Pressure of the Draft Tube <![CDATA[C 7 > Water Guide + -directional Swing <![CDATA[C 15 > Draft Tube Inlet Pressure <![CDATA[C 8 > Water Guide + Y-directional Swing <![CDATA[C 16 > Pressure between the Runner and the Guide Vane Y Lifting Amount <![CDATA[C 17 > Pressure between the Runner and the Draft Tube

[0071] 17 characteristic parameters associated with jacking-up, the unit head C 1 and the active power C 2 come from the LCU, and the guide vane opening C 3 and the differential pressure of the volute flow rate C 13 come from on-site measurements. The parameters of the upper guide + X-direction swing C 5 , the upper guide + Y-direction swing C 6, Water guide + X-directional swing C 7 , Water guide + Y-directional swing C 8 , Lower guide + X-directional swing C 9 , Lower guide + Y-directional swing C 10 The measuring points are arranged on the plane of each oil baffle ring, C 5 and C 6 , C 7 and C 8 , C 9 and C 10 Arranged at positions 90 degrees apart in the same plane, the vibration and swing acquisition device is an integrated non-contact sensor; the inlet pressure C of the unit's spiral case 11 and the vacuum pressure at the end of the spiral case C 12 The measuring points are set on the flange of the spiral case wall, and the elbow pressure of the draft tube C 14 The measuring points are on the flange of the elbow tube wall, and the inlet pressure of the draft tube C 15 The measuring points are at the flange of the draft tube wall, and the pressure between the runner and the guide vane C 16 The measuring points are set on the top cover of the turbine chamber, and the pressure between the runner and the draft tube C 17 The measuring points are on the flange at the connection between the runner and the draft tube. The pressure acquisition device is a pressure sensor with an accuracy error ≤ ±0.2% and a response speed ≤ 0.5 ms. The jacking-up amount uses a non-contact eddy current sensor with a sensitivity of 1.6 mA / mm;

[0072] Form the jacking-up related parameter set C with the obtained characteristic parameters j , Classify the collected operating condition data into three types of data: start-up process, shutdown process, and load rejection process according to the operating state of the unit

[0073] Step 2: Preprocess the operating condition data obtained in Step 1 to form a sample data set and divide it into a training set, a validation set, and a test set;

[0074] The preprocessing includes removing distortion, filling in missing values, and VMD noise reduction for historical data, removing untrustworthy data, and forming a training set, a validation set, and a test set according to a ratio of 7:2:1.

[0075] Step 3: Construct an LSTM jacking-up prediction model and an SVM jacking-up prediction model using a long short-term memory neural network model and a support vector machine model respectively, and use the operating condition data set in Step 2 to train and test the LSTM jacking-up prediction model and the SVM jacking-up prediction model respectively;

[0076] In the embodiments of the present invention, the transient process data used for prediction in the hydropower unit is long-time series data with a fixed sampling period, which has good adaptability to LSTM. Secondly, LSTM can continuously retain information and use historical states to discriminate future states. LSTM includes memory cells, forget gates, input gates, and output gates. The forget gate determines to forget the historical information with low relevance from the memory state information, the input gate determines the sensitivity of the memory cell state to the input data, and the output gate determines the information output at that moment.

[0077] Forget gate f t Determine C t-1 The features used for calculation in are used to calculate C t :

[0078] f t =σ(w f [H t-1 ,x t +b f ;(2)

[0079] Input gate i t Used to determine the data saved in the neuron state at time t:

[0080] i t =σ(w i .[H t-1 ,x t +b f );(3)

[0081] Calculate the neuron of the current state:

[0082] C t ′=tanσ(w c .[H t-1 ,x t +b c );(4)

[0083] Calculate the new neuron state C t :

[0084] C t =C t-1 ×f t +i t ×C t ′;(5)

[0085] Output gate O t Used to determine the virtual machine data and transfer it to the next neuron:

[0086] o t =σ(w o [H t-1 ,x t +bo ; (6)

[0087] Calculate the predicted value and generate the input for the next time:

[0088] h t = O t tanh(C t ); (7)

[0089] Through the above steps, the prediction result y 1 , y 2 , …, y m .

[0090] In the formula, the forget gate f t , the input gate i t and the output gate O t respectively represent the state value, input value and output value at time t; H t-1 , h t are the hidden information input and output by the memory module respectively; x t is the input vector at time t; C t-1 , C t are the memory cell states at times t-1 and t respectively; σ represents the sigmod activation function, and tanh represents the hyperbolic tangent activation function; w represents the weight matrix; b represents the bias vector.

[0091] The specific process of obtaining the prediction result by the support vector machine SVM model includes:

[0092] 1) Use the kernel function to map the input data to a high-dimensional feature space. The kernel functions include linear kernel, polynomial kernel, and radial basis function (RBF) kernel.

[0093] 2) Find an optimal hyperplane to maximize the functional margin of the samples (support vectors) closest to the

[0094] hyperplane, and improve the robustness of the function.

[0095] 3) Introduce slack variables to handle noise and inseparable cases.

[0096] 4) Solve through the dual problem to reduce the computational complexity.

[0097] 5) Make predictions through the optimal hyperplane. The prediction result is the projection value of the sample on the hyperplane.

[0098] The LSTM-SVM lifter prediction method establishes the mapping relationship between the lifter correlation feature parameter C j and the lifter amount Y through Equation (10), and obtains the optimal model and the lifter amount that meets the proposed indicators through training, validation, and testing.

[0099]

[0100] In the formula, τ is the sampling time, q is the number of selected correlation feature quantities, and q = 17, is the sample data, is the identification coefficient of the j-th correlation feature parameter at the τ-th moment. Y is the predicted jacking-up quantity set at the corresponding moment, and F is the mapping.

[0101] Step 3.1: Build an LSTM jacking-up prediction model and an SVM jacking-up prediction model respectively using a long short-term memory neural network model and a support vector machine model, and use the working condition data set in Step 2 to train the LSTM jacking-up prediction model and the SVM jacking-up prediction model respectively. The input of the LSTM jacking-up prediction model and the SVM jacking-up prediction model is the preprocessed working condition data set in Step 2, and the output is the predicted jacking-up quantity result;

[0102] Use the working condition data set in Step 2 to train the LSTM jacking-up prediction model and the SVM jacking-up prediction model respectively to obtain the jacking-up prediction model f LSTM of LSTM and the jacking-up prediction model f SVM of SVM;

[0103] Step 3.2: Select the validation set in Step 2 to verify the LSTM jacking-up prediction model and the SVM jacking-up prediction model obtained in Step 3.1 respectively. If the prediction effect of the model on the validation set is unqualified, adjust the hyperparameters of the corresponding model, and then execute Step 3.1 again to retrain the current model; if the verification passes, execute Step 3.3;

[0104] Verify the two models f LSTM , f SVM on the validation set, and use the evaluation index mean square error MSE as the judgment condition to verify whether the trained model is the optimal model. If MSE < 0.1 is not satisfied, update the hyperparameters of the corresponding model, and then execute Step 3.1 again to retrain the current model; if the verification passes, execute Step 3.3;

[0105] Step 3.3: After passing the verification on the validation set, obtain the optimized LSTM jacking-up prediction model and SVM jacking-up prediction model respectively;

[0106] After passing the verification on the validation set, obtain the optimized LSTM jacking-up prediction model M LSTM and the optimized SVM jacking-up prediction model M SVM , and compare the predicted jacking-up quantity value of the validation set with the true value of the working condition data to obtain the validation set error E LSTM and E SVM .

[0107] Step 4: Determine the weight coefficients of the optimized LSTM jacking prediction model and the SVM jacking prediction model obtained in Step 3.3 for the combined prediction results of the unit jacking volume respectively;

[0108] Let the weight of the LSTM jacking prediction model be Q LSTM , and the weight of the SVM jacking prediction model be Q SVM ;

[0109] Determine the weight coefficients through E LSTM and E SVM , and set the threshold θ;

[0110] When , it indicates that the prediction results of the two models have a large difference:

[0111] If E LSTM > E SVM then Q LSTM = 0, discard the prediction result of the LSTM jacking prediction model, Q SVM = 1, take the prediction result of the SVM jacking prediction model;

[0112] If E LSTM < E SVM then Q SVM = 0, discard the prediction result of the SVM jacking prediction model, Q LSTM = 1, take the prediction result of the LSTM jacking prediction model;

[0113] Otherwise, the results of the two algorithms are determined by E LSTM and E SVM to determine the weight coefficients, and the formula is as follows:

[0114]

[0115] Step 5: Respectively use the trained LSTM jacking prediction model and the SVM jacking prediction model to obtain the real-time LSTM jacking prediction value and the SVM jacking prediction value according to the real-time hydropower unit condition data, and perform weighted summation in combination with the weight coefficients obtained in Step 4 to obtain the final prediction result of the unit jacking volume.

[0116] Model M LSTM and M SVM are tested in the test set, and the prediction results, the LSTM jacking prediction value Y LSTM and the SVM jacking prediction value Y SVM , are combined according to the weight coefficients determined by the judgment logic in Step 4 for the prediction results Y LSTM and Y SVM to obtain the final prediction result Y, and the expression is:

[0117] Y = Y LSTM .QLSTM +Y SVM .Q SVM ;(11)

[0118] To verify the prediction effect of the combined prediction of LSTM and SVM, 785 groups of measured data during the startup process, 966 groups of data during the shutdown process, and 589 groups of data during the load rejection process were adopted. The acquisition period was 1 second. The acquisition time during the startup process was from the receipt of the startup command to the time when the unit was connected to the grid and carried the rated load, which was 20 minutes; the shutdown process included the process from the initial state of the unit carrying the rated load to the reduction of the load until the unit was completely shut down, and the shutdown time was 25 minutes; the load rejection process was the disconnection of the unit from the power grid while carrying the rated load, and the time was 7 minutes. The prediction results are as Figure 7 shown

[0119] As Figure 7 and Figure 8 and Figure 9 shown, the prediction curves of the combined prediction of LSTM and SVM under the operating states of the startup process, shutdown process, and load rejection process of the three working condition data are in the highest degree of coincidence with the true value curves, thus verifying the effectiveness of the combined prediction of LSTM and SVM for the prediction of the jacking-up amount.

[0120] This example also includes evaluating the combined prediction results of the jacking-up amount of the unit by selecting the root mean square error RMSE, mean absolute error MAE, mean square error MSE, and goodness of fit R2.

[0121] The meanings of each index are expressed as in formulas (12)-(15). The first three indexes each have their own advantages and complement each other. When their values approach 0, it indicates that the prediction result has higher accuracy; the goodness of fit R2 represents the degree of coincidence between the observed data and the expected value of the statistical model and the prediction ability of the model for new data. When this index is close to 1, it indicates that the observed value is close to the expected value of the model.

[0122]

[0123] In the formula, y t is the true value of the jacking-up amount, is the predicted value of the jacking-up amount, is the average value of the jacking-up amount;

[0124] As Figure 10 and Figure 11 and Figure 12 shown, the comparison charts of the evaluation indexes of the three algorithms under the three working conditions are given, and Tables 3-5 are the corresponding index data. Figure 10 and Figure 11 and Figure 12The evaluation results of root mean square error (RMSE), mean absolute error (MAE), mean square error (MSE), and coefficient of determination (R2) are given respectively. When the first three indicators approach 0, it indicates higher prediction accuracy; when the coefficient of determination R2 is close to 1, it shows that the observed values are close to the expected values of the model, and the prediction results are better. Under the three working conditions, the four evaluation indicators of LSTM-SVM are better than those of BiLSTM and LSTM algorithms. This result is consistent with that obtained in Figure 7 and Figure 8 and Figure 9 . At the same time, it is found that the prediction accuracy of the bidirectional BiLSTM model is better than that of the LSTM model under the three working conditions.

[0125] Further comparing the prediction results of the three transition conditions, the coefficients of determination predicted by LSTM-SVM are 99.9%, 99.6%, and 98.1% under the three conditions of starting up, shutting down, and load rejection respectively, and the prediction accuracy is relatively high. Based on the above analysis, the LSTM-SVM adopted in the embodiment of the present invention is reasonable and effective for predicting the jacking-up amount.

[0126] Table 3 Comparison of real-time prediction evaluations during the starting-up process

[0127]

[0128] Table 4 Comparison table of real-time prediction evaluations during the shutting-down process

[0129]

[0130] Table 5 Comparison table of real-time prediction evaluations during the load-rejection process

[0131]

[0132] In this embodiment, the real operation data of large hydropower station units are used, and with the help of the combined intelligent learning algorithm model of LSTM and SVM, the jacking-up amount during the key transition processes of unit starting up, shutting down, and load rejection is trained and predicted. By giving full play to the strong advantages of the LSTM network in processing time series data and the excellent performance of the SVM algorithm in data classification and anti-overfitting, high-precision prediction of the jacking-up phenomenon is achieved. Compared with traditional algorithms such as BiLSTM and LSTM, the combined prediction model of LSTM and SVM proposed in this paper shows higher prediction accuracy, can accurately predict the jacking-up amount under different operating states, has a very small gap with the true value, and has stronger adaptability and reliability. This research result provides a practical new approach for the prevention and control of the jacking-up fault of large hydropower units, significantly improves the intelligent operation and maintenance level of the hydropower industry, and ensures the safe and stable operation of the power system.

Claims

1. A method for predicting the lifting capacity of a hydropower unit based on a combination of LSTM and SVM prediction, characterized in that: The following steps are involved: Step 1: Collect the operating data of the hydropower unit and conduct correlation analysis to determine the characteristic parameters for prediction of the lifting capacity; Step 2: Preprocess the working condition data obtained in step 1 to form a working condition data set, and divide it into a training set, a validation set, and a test set; Step 3: Use the long short-term memory neural network model and the support vector machine model to build the LSTM machine lift prediction model and the SVM machine lift prediction model respectively, and use the working condition data set in step 2 to train and verify the LSTM machine lift prediction model and the SVM machine lift prediction model respectively; Step 4: respectively determine the weight coefficients of the LSTM lift prediction model and the SVM lift prediction model for the combined prediction results of the unit lift quantity; Step 5: Use the trained LSTM machine lift prediction model and SVM machine lift prediction model to obtain real-time LSTM machine lift prediction value and SVM machine lift prediction value according to the real-time hydropower unit operating data, and perform weighted summation with the weight coefficient obtained in step 4 to obtain the final unit lift prediction result.

2. The method for predicting the lifting capacity of a hydropower unit according to claim 1, characterized in that: In the step 1, sample data of the operating conditions of the hydropower unit are obtained through collection points, control unit LCU and on-site measurement. The operating conditions of the hydropower unit include startup, shutdown and load shedding.

3. The method for predicting the lifting capacity of a hydropower unit according to claim 1, characterized in that: In step 1, the characteristic parameters used for prediction of lift volume are subjected to correlation analysis using the Pearson correlation coefficient; The calculation formula of Pearson correlation coefficient r is: In the formula, x t Represents the value of a characteristic parameter, y t Indicates the value of the lift amount. Represents the average value of a characteristic parameter, the average value of the y lift.

4. The method for predicting the lifting capacity of a hydropower unit according to claim 3 is characterized in that: In step 1, the characteristic parameters used for prediction of the lifting amount include the unit head C 1 , Active power C 2 , guide vane opening C 3 , Unit speed C 4 、Swing C in the +X direction at the upper guide bearing 5 、Upper guide bearing +Y swing C 6 , Water guide bearing +X direction swing C 7 , Water guide bearing +Y swing C 8 , Lower guide bearing +X swing C 9 , Lower guide bearing +Y swing C 10 、Unit volute inlet pressure C 11 、Vacuum pressure at the end of the volute C 12 , Volute flow pressure difference C 13 , tailwater elbow pressure C 14 , tailwater pipe inlet pressure C 15 , Pressure between runner and guide vane C 16 Pressure between tailwater pipe C 17 .

5. The method for predicting the lifting capacity of a hydropower unit according to claim 1, characterized in that: In step 2, the preprocessing includes removing distortion, filling missing values ​​and reducing noise from historical data, removing unreliable data, and forming a training set, a validation set, and a test set in a ratio of 7:2:

1.

6. The method for predicting the lifting capacity of a hydropower unit according to claim 1, characterized in that: The step 3 includes the following sub-steps: Step 3.1: Use the long short-term memory neural network model and the support vector machine model to build the LSTM machine lift prediction model and the SVM machine lift prediction model respectively, and use the working condition data set in step 2 to train the LSTM machine lift prediction model and the SVM machine lift prediction model respectively. The input of the LSTM machine lift prediction model and the SVM machine lift prediction model are both the working condition data set preprocessed in step 2, and the output is the machine lift prediction result; Step 3.2: Select the validation set in step 2 to validate the LSTM lift prediction model and SVM lift prediction model obtained in step 3.

1. If the prediction effect of the LSTM lift prediction model and the SVM lift prediction model on the validation set is unsatisfactory, adjust the hyperparameters of the corresponding model and execute step 3.1 again for retraining; if the validation passes, execute step 3.3; Step 3.3: After verification on the validation set, the optimal LSTM lift prediction model and SVM lift prediction model are obtained respectively.

7. The method for predicting the lifting capacity of a hydropower unit according to claim 1, characterized in that: In step 3, the long short-term memory network model includes a memory unit, a forget gate, an input gate and an output gate. The forget gate determines to forget the historical information with low relevance from the memory state information, the input gate determines the sensitivity of the memory unit state to the input data, and the output gate determines to output the information at that moment.

8. The method for predicting the lifting capacity of a hydropower unit according to claim 1, characterized in that: In step 3, the specific process of using the support vector machine model to obtain the prediction result includes: 1) Mapping the input data to a high-dimensional feature space using a kernel function, wherein the kernel function includes a linear kernel, a polynomial kernel, and a radial basis function kernel; 2) Determine the optimal hyperplane to maximize the function interval of the samples closest to the plane; 3) Introduce slack variables to handle the noise in the input data and the situation where the data is linearly inseparable; 4) Reduce computational complexity by solving the dual problem; 5) Prediction is performed through the optimal hyperplane, and the prediction result is the projection value of the sample on the hyperplane.

9. The method for predicting the lifting capacity of a hydropower unit according to claim 1, characterized in that: In step 4, the weight coefficients of the LSTM lift prediction model and the SVM lift prediction model for the combined prediction results of the unit lift quantity are determined, specifically including: Assume that the weight of the LSTM lift prediction model is Q LSTM , the LSTM lift prediction model error is E LSTM , the weight of the SVM lift prediction model is Q SVM , the error of SVM lift prediction model is E SVM ; By E LSTM and E SVM Determine the weight coefficient and set the threshold θ; when When , it means that the prediction results of the two models are quite different: If E LSTM >E SVM Then let Q LSTM =0,Q SVM =1, that is, the prediction result of the SVM lift prediction model is taken as the final unit lift prediction result; If E LSTM <E SVM Then let Q SVM =0,Q LSTM =1, that is, the prediction result of the LSTM machine lifting prediction model is taken as the final prediction result of the unit lifting quantity; Otherwise, according to the LSTM lift prediction model error E LSTM and SVM lift prediction model error E SVM Determine the weight coefficient, the formula is as follows:

10. The method for predicting the lifting capacity of a hydropower unit according to claim 9, characterized in that: In step 5, the calculation formula for the final unit lifting amount prediction result is: Y=Y LSTM .Q LSTM +Y SVM .Q SVM ; (4) Where Y LSTM is the LSTM lift prediction value, Y SVM is the SVM lift prediction value, Q LSTM is the weight of the LSTM lift prediction model, Q SVM is the weight of the SVM lift prediction model.

Images