Intelligent operation method based on statistics and analysis of operation data of large hydropower station unit

Abstract

The invention discloses an intelligent operation method based on statistics and analysis of operation data of a large hydropower station unit, and the method comprises the steps: firstly achieving the cleaning of feature vector values through a GWO-SDAE model, carrying out the reconstruction and restoration of abnormal point data, and guaranteeing the quality of the feature vector data; carrying out preliminary analysis on the cleaned feature vector values by utilizing a descriptive statistical model; further learning feature vector values and extracting internal rules by using an MVO-BiLSTM hybrid model so as to realize automatic and intelligent prediction and analysis of feature values, comparing collected feature data values with predicted values, sending out an alarm signal in time if errors exceed a specified range, and performing manual inspection so as to realize automatic and intelligent prediction and analysis of the feature values. And finally generating an operation data analysis report of the large hydropower station unit.

Description

Technical Field

[0001] The present invention belongs to the technical field of engineering BIM modeling, and in particular relates to an intelligent operation method based on statistics and analysis of operation data of large-scale hydropower station units. Background Art

[0002] In the context of energy interconnection in today's society, the safe and stable operation of hydropower plant power generation equipment plays an increasingly important role in the interconnection of power systems and energy supply. As one of the daily tasks of operators, timely completion of statistics and analysis of hydropower station equipment operation data plays an important reference role in accurately judging the health of equipment, timely discovering potential equipment failure information, and further formulating equipment maintenance plans, providing strong support for the stable operation of the power system.

[0003] At present, the statistics and analysis of the operation data of hydropower station units mostly consume a lot of human resources. Effective data information is extracted from the scattered raw data, and the data is sorted and analyzed, mostly relying on manual experience to finally draw conclusions. Obviously, with the rapid development of the power system, this method can no longer meet the needs. It is urgent to use digital, information and intelligent technology as a carrier to realize the intelligent statistics and analysis of the operation data of large hydropower station equipment.

[0004] The existing technology has the following shortcomings: large-scale hydropower stations have many equipment and complex systems. The analysis of huge amounts of data increases a lot of manpower and time costs, which will lead to inaccurate equipment analysis results, easily miss important latent equipment failures, and be easily affected by human factors. It is inefficient and has great uncertainty. It cannot detect potential equipment failures in time, affecting the health of the equipment. Therefore, it is necessary to design an intelligent operation method based on statistics and analysis of large hydropower station unit operation data to solve the above problems. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an intelligent operation method based on the statistics and analysis of the operation data of large hydropower station units, and to perform real-time data analysis on hydropower station equipment by means of digitalization, informatization and intelligentization technology, so as to achieve the purpose of real-time visualization of power station equipment data.

[0006] In order to achieve the above technical effects, the technical solution adopted by the present invention is: A smart operation method based on statistics and analysis of operation data of large hydropower station units includes the following steps: S1, determining the statistical analysis modules of the operating data of the large hydropower station units and the feature vector categories corresponding to each statistical analysis module; S2, based on the feature vector categories corresponding to each statistical analysis module, automatically extracts data from the computer monitoring system background database and other various online monitoring systems; S3, using the GWO-SDAE model to clean the extracted feature vectors, reconstruct and repair the abnormal point data, and ensure the quality of the feature vector data; S4, according to the rated limit range of the equipment nameplate and a large amount of operating experience, set the variation range of the eigenvector value corresponding to each statistical analysis module, and use the descriptive statistical model to perform a preliminary analysis on the cleaned eigenvector value. If the analysis result meets the equipment parameter value range, proceed to the next step. If not, issue an alarm for inspection; S5, after descriptive statistical analysis of the eigenvector values, the MVO-BiLSTM hybrid model is used to learn the eigenvector values ​​and extract internal rules, so as to realize automatic and intelligent prediction and analysis of the eigenvalues; the collected eigendata values ​​are compared with the predicted values, and if the error exceeds the specified range, an alarm signal is issued in time and an inspection is carried out, otherwise the next step is carried out; S6, determine the format of the operation data analysis report of the large hydropower station unit, fill in the analysis conclusion of the characteristic value in step S5, further review, revise and finalize, and output the corresponding analysis report.

[0007] Preferably, in step S2, automatically extracting data from the computer monitoring system backend database and other various online monitoring systems includes: S201, according to the time difference of extracting characteristic information of different equipment in large hydropower stations, set extraction conditions corresponding to characteristic information of different equipment, the conditions include: N hours after the turbine generator set is shut down, N hours before the turbine generator set is started and connected to the grid, the turbine generator set is connected to the grid and generates electricity, the active power of the unit is greater than the preset value, the delay is Q hours, and N hours before the turbine generator set is shut down and completely stopped; The grid-connected power generation status of a hydro-generator unit refers to: The unit speed is greater than 95% of the rated speed, the main synchronous switch is in the closed position, the stator three-phase current is greater than 5% of the rated current, and the three-phase terminal voltage is greater than 90% of the rated voltage. The grid-connected shutdown state of a hydro-generator unit refers to: The unit speed is less than 1%, the guide vane is in the fully closed position, the main synchronous switch is in the open position, the automatic lock spindle is in operation, and the water system is shut down. S202, setting multiple sensor data collection methods, stipulating that data information is recorded once every 1 second, and each sensor collects 86,400 data information every day; all data information collected by all sensors of the hydropower station is transmitted to the computer monitoring database; S203, extracting effective feature information according to the time differentiation requirement for extracting feature information of different devices in step S201.

[0008] Preferably, in step S3, the method for constructing the GWO-SDAE model includes constructing a GWO model part: The parameter optimization process of the GWO model is divided into surrounding prey, chasing prey and attacking prey. First, the wolf pack is divided into four levels: α, β, δ and ω. The roles of the wolf packs at levels ω, δ, β, and α are: the main force of the encirclement target, determining the direction of the encirclement target, finding the encirclement target and the leading role in the encirclement process. The process of encircling the prey is: ; In the formula, , Represents the current position of the gray wolf and the position after iteration, is the prey position, D is the distance between the wolf and the prey, A and C are random vectors, is a linearly decreasing value between [2,0], , is a random number between [0, 1]; By iteratively updating the fitness value of the gray wolf, the optimal solution, excellent solution and suboptimal solution are obtained respectively: α, β, δ, ω The hunting direction and position update of the wolf are jointly determined by α, β, and δ. The specific formula is as follows: ; ; ; Where D1, D2 and D3 are α, β and δ respectively, the distance between the wolf and the prey; A1, A2, A 3、 C1, C2 and C3 are the random vectors of α, β and δ wolves respectively; X1, X2 and X3 are the current positions of α, β and δ wolves respectively. , and These are the prey locations of α, β, and δ wolves respectively.

[0009] Preferably, the method for constructing the GWO-SDAE model in step S3 includes constructing a SDAE data cleaning model: The stacked denoising autoencoder SDAE is an unsupervised learning model. It is a deep learning architecture composed of multiple DAEs stacked together. By randomly adding noise to the feature data, the rules between the feature data are further learned, and the potential connections between the initial feature sequences are mastered. The SDAE adopts a layer-by-layer training method. The hidden layer features extracted from the encoding part of the previous DAE are used as the input of the DAE encoder of the next layer. This process is repeated until all DAEs are trained. The specific method is: (1) In the initial feature sequence Randomly add noise data to get the latest feature sequence ; (2) For new feature sequences Encode and obtain the DAE features of the first layer : ; In the formula, , are the weight matrix and bias vector, is the sigmoid activation function; (3) Decode the DAE features of the first layer Perform reconstruction and repair to obtain the reconstruction result: ; In the formula, , are the weight matrix and bias vector of the decoder respectively; (4) Determine the reconstruction results With the initial feature sequence The error between MAPE Indicates the quality of model performance: ; The micro gradient descent method is used to further reduce and Error between: ; ; In the formula, x ij , z ij Respectively represent the values ​​before and after data cleaning using the SDAE model, a and b is the type and dimension of the feature vector set, and Respectively represent i Layer neurons pass nThe weights and bias vectors after the residual iteration are calculated backwards. r and m Indicates the current training times and training number of samples, l r is the learning rate of the model; (5) According to the principle of layer-by-layer training of the SDAE model, continue to train and learn each layer of DAE separately, determine the size of the key parameters of each layer, and finally form a complete SDAE model.

[0010] Preferably, in step S3, the specific process of using the GWO-SDAE model to clean the extracted feature vector and reconstruct and repair the abnormal point data is as follows: S301, extracting initial feature data based on the data in the background database of the computer monitoring system; S302, randomly adding noise to the extracted initial feature data, and performing standardization processing, dividing the data into training set data and test set data according to a fixed ratio; S303, setting the maximum number of iterations of the GWO model, the population size, the value range of the number of hidden layer nodes and the noise coverage of the SDAE model, and the search speed range, and randomly initializing the population; S304, inputting the training set data into the SDAE data cleaning model to train each layer of SDAE; S305, calculate the fitness of the calculation particles and determine the reconstruction result With the initial feature sequence The error between MAPE Indicates the performance of the model. The smaller the fitness function value, the better the data cleaning effect. S306, updating the individual optimal position and the group optimal position of the particles of the GWO model, and continuously updating the velocity and displacement of the particles, and judging whether the GWO model has reached the termination condition. If the termination condition is met, the optimal number of hidden layer nodes and noise coverage are obtained, otherwise, the process returns to step S304 to continue iterative updating; the termination condition is set as the fitness value of the particles of the GWO model tends to be stable or the number of iterations reaches the maximum; S307, input the test set data into the reconstructed GWO-SDAE network data cleaning model, reconstruct and repair the abnormal point data, and finally perform denormalization on the results and evaluate them in combination with the evaluation indicators.

[0011] Preferably, in step S4, using a descriptive statistical model to perform a preliminary analysis on the cleaned feature vector values ​​includes: The mean, median, mode, variance and standard deviation corresponding to each eigenvector value are calculated respectively, and the calculation results are compared with the corresponding setting range. If the value range specified by the equipment parameters is exceeded, an alarm is issued for inspection. If all eigenvector values ​​are within the specified range, it is judged that the large hydropower station unit is in good operating condition and the equipment can operate sustainably.

[0012] Preferably, in step S5, the MVO-BiLSTM hybrid model is used to learn the eigenvector value after the preliminary analysis in step S4 and grasp the internal change law, so as to realize automatic and intelligent prediction and analysis of the eigenvalue. The specific process is as follows: The MVO-BiLSTM hybrid model is used to accurately predict the feature vector data, and the collected feature data values ​​are compared with the predicted values. If the error exceeds the specified range, an alarm signal is issued in time and manual inspection is performed. The specific process is as follows: S501, taking each feature vector value after preliminary analysis by the descriptive statistical model in step S4 as the input of the MVO-BiLSTM hybrid prediction model and performing normalization processing; S502, dividing the normalized data into a training set and a test set according to an 8:2 ratio; S503, setting the maximum number of iterations of the multiverse optimization algorithm MVO, the population size, the value range of the number of neurons m in the hidden layer of the BiLSTM model and the learning rate ε, and the search speed range, and initializing the universe population kernel parameters; S504, input the training set data into the BiLSTM network data prediction model and calculate , , calculate the expansion rate of each universe, implement the roulette mechanism to select cosmic objects, and continuously update the expansion rate of the universe; S505, calculate the particle fitness value and determine whether the iteration termination condition is met: If the termination condition is reached, the optimal number of hidden layer neurons is output. m With learning rate ε If the judgment condition is not met, then return to step S504 to iterate and optimize again; the iteration termination condition is that the particle fitness value tends to be flat or the number of iterations reaches the maximum; S506, using the obtained optimal hyperparameters to rebuild the MVO-BiLSTM hybrid prediction model, taking the test set data as input, performing iterative prediction, outputting the prediction results, and evaluating them in combination with the evaluation indicators. The formula is as follows: ; ; In the formula, y MAPE is the mean absolute percentage error,y RMSE is the root mean square error. The smaller the mean absolute percentage error and the root mean square error, the better the model prediction effect. X act ( i )and X pred ( i ) respectively represent i The actual value and predicted value at the moment; S506, compare the collected characteristic data value with the predicted value, and if the error exceeds the specified range, issue an alarm signal in time and conduct manual inspection.

[0013] The beneficial effects of the present invention are as follows: The present invention proposes an intelligent operation method based on the statistics and analysis of the operation data of large-scale hydropower station units. It uses digital, information and intelligent technologies as means to regularly analyze the operation data of hydropower station equipment and automatically generate a monthly report on the unit operation. It has the advantages of liberating human resources, high accuracy of analysis reports, and highlighting abnormal data, which has greatly promoted the development of smart power stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flow chart of the present invention; Figure 2 This is a diagram of the statistical analysis module architecture for the operating data of a large hydropower station unit of the present invention; Figure 3 Schematic diagram of the internal structure of SDAE in the present invention; Figure 4 Schematic diagram of the SDAE network model training method in the present invention; Figure 5 Schematic diagram of the data cleaning method based on the GWO-SDAE model in the present invention; Figure 6 Schematic diagram of the BiLSTM network model in the present invention; Figure 7 This is a data prediction flow chart based on the MVO-BiLSTM hybrid model in the present invention; Figure 8 The results of the 26F hydro-generator vibration swing statistics and analysis module are shown in the embodiment of the present invention. Figure 1 ; Fig. 9 The results of the 26F hydro-generator vibration swing statistics and analysis module are shown in the embodiment of the present invention. Figure 2 ; Fig.10 The results of the 26F hydro-generator vibration swing statistics and analysis module are shown in the embodiment of the present invention. Figure 3 ; Fig.11 The results of the 26F hydro-generator vibration swing statistics and analysis module are shown in the embodiment of the present invention. Figure 4 ; Fig.12 The results of the vibration swing statistics and analysis module of all units in the power station according to the embodiment of the present invention are displayed. Figure 1 ; Fig.13 The results of the vibration swing statistics and analysis module of all units in the power station according to the embodiment of the present invention are displayed. Figure 2 ; Fig.14 The results of the vibration swing statistics and analysis module of all units in the power station according to the embodiment of the present invention are displayed. Figure 3 . DETAILED DESCRIPTION

[0015] Embodiment 1: like Figure 1 As shown, a smart operation method based on statistics and analysis of operation data of large hydropower station units includes the following steps: S1, determining the statistical analysis modules of the operating data of the large hydropower station units and the feature vector categories corresponding to each statistical analysis module; S2, based on the feature vector categories corresponding to each statistical analysis module, automatically extracts data from the computer monitoring system background database and other various online monitoring systems; S3, using the GWO-SDAE model to clean the extracted feature vectors, reconstruct and repair the abnormal point data, and ensure the quality of the feature vector data; S4, according to the rated limit range of the equipment nameplate and a large amount of operating experience, set the variation range of the eigenvector value corresponding to each statistical analysis module, and use the descriptive statistical model to perform a preliminary analysis on the cleaned eigenvector value. If the analysis result meets the equipment parameter value range, proceed to the next step. If not, issue an alarm for inspection; S5, after descriptive statistical analysis of the eigenvector values, the MVO-BiLSTM hybrid model is used to learn the eigenvector values ​​and extract internal rules, so as to realize automatic and intelligent prediction and analysis of the eigenvalues; the collected eigendata values ​​are compared with the predicted values, and if the error exceeds the specified range, an alarm signal is issued in time and an inspection is carried out, otherwise the next step is carried out; S6, determine the format of the operation data analysis report of the large hydropower station unit, fill in the analysis conclusion of the characteristic value in step S5, further review, revise and finalize, and output the corresponding analysis report.

[0016] In this embodiment, the statistical analysis module of the operation data of the large hydropower station units is as follows: Figure 2As shown; in step S1, the statistical analysis module of the operating data of the large hydropower station unit and the corresponding feature vector category are determined, taking the vibration swing statistical analysis module of the hydro-generator unit as an example: The corresponding eigenvector categories are determined as follows: unit head, active power, reactive power, upper guide X-direction swing, upper guide Y-direction swing, lower guide X-direction swing, lower guide Y-direction swing, water guide X-direction swing, water guide Y-direction swing, upper frame X-horizontal vibration, upper frame Y-horizontal vibration, upper frame X-vertical vibration, upper frame Y-vertical vibration, stator frame X-horizontal vibration, stator frame Y-horizontal vibration, stator frame X-vertical vibration, stator core horizontal vibration, lower frame X-horizontal vibration, lower frame Y-horizontal vibration, lower frame X-vertical vibration, lower frame Y-vertical vibration, top cover X-horizontal vibration, top cover Y-horizontal vibration, top cover X-vertical vibration, pressure pulsation under the top cover, pressure pulsation between the runner and the guide vane, volute inlet pressure pulsation, draft tube upstream pressure pulsation, draft tube downstream pressure pulsation. Each of the above eigenvector categories is divided into maximum value, minimum value and average value. The unit of unit head is: m, the unit of active power is: MW, the unit of reactive power is: Mvar, the unit of unit swing is: mm, the unit vibration unit is: mm, and the unit pressure pulsation unit is: %.

[0017] Preferably, in step S2, automatically extracting data from the computer monitoring system backend database and other various online monitoring systems includes: S201, according to the time difference of extracting characteristic information of different equipment in large hydropower stations, set extraction conditions corresponding to characteristic information of different equipment, the conditions include: N hours after the turbine generator set is shut down, N hours before the turbine generator set is started and connected to the grid, the turbine generator set is connected to the grid and generates electricity, the active power of the unit is greater than the preset value, the delay is Q hours, and N hours before the turbine generator set is shut down and completely stopped; The grid-connected power generation status of a hydro-generator unit refers to: The unit speed is greater than 95% of the rated speed, the main synchronous switch is in the closed position, the stator three-phase current is greater than 5% of the rated current, and the three-phase terminal voltage is greater than 90% of the rated voltage. The grid-connected shutdown state of a hydro-generator unit refers to: The unit speed is less than 1%, the guide vane is in the fully closed position, the main synchronous switch is in the open position, the automatic lock spindle is in operation, and the water system is shut down. S202, setting multiple sensor data collection methods, stipulating that data information is recorded once every 1 second, and each sensor collects 86,400 data information every day; all data information collected by all sensors of the hydropower station is transmitted to the computer monitoring database; S203, extracting effective feature information according to the time differentiation requirement for extracting feature information of different devices in step S201.

[0018] Preferably, in step S3, the method for constructing the GWO-SDAE model includes constructing a GWO model part: The parameter optimization process of the GWO model is divided into surrounding prey, chasing prey and attacking prey. First, the wolf pack is divided into four levels: α, β, δ and ω. The roles of the wolf packs at levels ω, δ, β, and α are: the main force of the encirclement target, determining the direction of the encirclement target, finding the encirclement target and the leading role in the encirclement process. The process of encircling the prey is: ; In the formula, , Represents the current position of the gray wolf and the position after iteration, is the prey position, D is the distance between the wolf and the prey, A and C are random vectors, is a linearly decreasing value between [2,0], , is a random number between [0, 1]; By iteratively updating the fitness value of the gray wolf, the optimal solution, excellent solution and suboptimal solution are obtained respectively: α, β, δ, ω The hunting direction and position update of the wolf are jointly determined by α, β, and δ. The specific formula is as follows: ; ; ; Where D1, D2 and D3 are α, β and δ respectively, the distance between the wolf and the prey; A1, A2, A 3、 C1, C2 and C3 are the random vectors of α, β and δ wolves respectively; X1, X2 and X3 are the current positions of α, β and δ wolves respectively. , and These are the prey locations of α, β, and δ wolves respectively.

[0019] Preferably, the method for constructing the GWO-SDAE model in step S3 includes constructing a SDAE data cleaning model: The stacked denoising autoencoder SDAE is an unsupervised learning model. It is a deep learning architecture composed of multiple DAEs stacked together. By randomly adding noise to the feature data, the rules between the feature data are further learned, and the potential connections between the initial feature sequences are mastered. The SDAE adopts a layer-by-layer training method. The hidden layer features extracted from the encoding part of the previous DAE are used as the input of the DAE encoder of the next layer. This process is repeated until all DAEs are trained. The internal structure of SDAE is as follows: Figure 3 As shown, the training method is as follows Figure 4 As shown; the specific method is: (1) In the initial feature sequence Randomly add noise data to get the latest feature sequence ; (2) For new feature sequences Encode and obtain the DAE features of the first layer : ; In the formula, , are the weight matrix and bias vector, is the sigmoid activation function; (3) Decode the DAE features of the first layer Perform reconstruction and repair to obtain the reconstruction result: ; In the formula, , are the weight matrix and bias vector of the decoder respectively; (4) Determine the reconstruction results With the initial feature sequence The error between MAPE Indicates the quality of model performance: ; The micro gradient descent method is used to further reduce and Error between: ; ; In the formula, x ij , z ij Respectively represent the values ​​before and after data cleaning using the SDAE model, a and b is the type and dimension of the feature vector set, and Respectively representi Layer neurons pass n The weights and bias vectors after the residual iteration are calculated backwards. r and m Indicates the current training times and training number of samples, l r is the learning rate of the model; (5) According to the principle of layer-by-layer training of the SDAE model, continue to train and learn each layer of DAE separately, determine the size of the key parameters of each layer, and finally form a complete SDAE model.

[0020] like Figure 5 As shown, in step S3, the GWO-SDAE model is used to clean the extracted feature vectors and the specific process of reconstructing and repairing the abnormal point data is as follows: S301, extracting initial feature data based on the data in the background database of the computer monitoring system; S302, randomly adding noise to the extracted initial feature data, and performing standardization processing, dividing the data into training set data and test set data according to a fixed ratio; S303, setting the maximum number of iterations of the GWO model, the population size, the value range of the number of hidden layer nodes and the noise coverage of the SDAE model, and the search speed range, and randomly initializing the population; S304, inputting the training set data into the SDAE data cleaning model to train each layer of SDAE; S305, calculate the fitness of the calculation particles and determine the reconstruction result With the initial feature sequence The error between MAPE Indicates the performance of the model. The smaller the fitness function value, the better the data cleaning effect. S306, updating the individual optimal position and the group optimal position of the particles of the GWO model, and continuously updating the velocity and displacement of the particles, and judging whether the GWO model has reached the termination condition. If the termination condition is met, the optimal number of hidden layer nodes and noise coverage are obtained, otherwise, the process returns to step S304 to continue iterative updating; the termination condition is set as the fitness value of the particles of the GWO model tends to be stable or the number of iterations reaches the maximum; S307, input the test set data into the reconstructed GWO-SDAE network data cleaning model, reconstruct and repair the abnormal point data, and finally perform denormalization on the results and evaluate them in combination with the evaluation indicators.

[0021] Preferably, in step S4, using a descriptive statistical model to perform a preliminary analysis on the cleaned feature vector values ​​includes: The mean, median, mode, variance and standard deviation corresponding to each eigenvector value are calculated respectively, and the calculation results are compared with the corresponding setting range. If the value range specified by the equipment parameters is exceeded, an alarm is issued for inspection. If all eigenvector values ​​are within the specified range, it is judged that the large hydropower station unit is in good operating condition and the equipment can operate sustainably.

[0022] like Figure 7 As shown, in step S5, the MVO-BiLSTM hybrid model is used to learn the eigenvector value after the preliminary analysis in step S4 and grasp the internal change law, so as to realize the automatic and intelligent prediction and analysis of the eigenvalue. The specific process is as follows: The MVO-BiLSTM hybrid model is used to accurately predict the feature vector data, and the collected feature data values ​​are compared with the predicted values. If the error exceeds the specified range, an alarm signal is issued in time and manual inspection is performed. The specific process is as follows: S501, taking each feature vector value after preliminary analysis by the descriptive statistical model in step S4 as the input of the MVO-BiLSTM hybrid prediction model and performing normalization processing; S502, dividing the normalized data into a training set and a test set according to an 8:2 ratio; S503, setting the maximum number of iterations of the multiverse optimization algorithm MVO, the population size, the value range of the number of neurons m in the hidden layer of the BiLSTM model and the learning rate ε, and the search speed range, and initializing the universe population kernel parameters; S504, input the training set data into the BiLSTM network data prediction model and calculate , , calculate the expansion rate of each universe, implement the roulette mechanism to select cosmic objects, and continuously update the expansion rate of the universe; S505, calculate the particle fitness value and determine whether the iteration termination condition is met: If the termination condition is reached, the optimal number of hidden layer neurons is output. m With learning rate ε If the judgment condition is not met, then return to step S504 to iterate and optimize again; the iteration termination condition is that the particle fitness value tends to be flat or the number of iterations reaches the maximum; S506, using the obtained optimal hyperparameters to rebuild the MVO-BiLSTM hybrid prediction model, taking the test set data as input, performing iterative prediction, outputting the prediction results, and evaluating them in combination with the evaluation indicators. The formula is as follows: ; ; In the formula, y MAPEis the mean absolute percentage error, y RMSE is the root mean square error. The smaller the mean absolute percentage error and the root mean square error, the better the model prediction effect. X act ( i )and X pred ( i ) respectively represent i The actual value and predicted value at the moment; S506, compare the collected characteristic data value with the predicted value, and if the error exceeds the specified range, issue an alarm signal in time and conduct manual inspection.

[0023] Embodiment 2: In step S5, when constructing the MVO-BiLSTM hybrid model, the MVO principle is as follows: The Multiverse Optimization Algorithm (MVO) was proposed in 2016. It plays an important role in hyperparameter optimization. It has the characteristics of fast convergence speed and strong parameter optimization ability. MOV finds the universe with the best expansion rate through the interaction of white holes, black holes, and wormholes in the multiverse. The process of solving the universe with the best expansion rate is the optimization iteration of the model. The specific steps are as follows: (1) Assume that the initial value of MVO is equation (8), and use equation (9) to update each value, where: d , s are the number of variables and the number of universe groups, for i in the universe j variable, for i The normalized expansion rate, is a random number between (0,1); ; ; (2) MVO continuously updates the expansion rate using equation (10), where: The best universe currently j parameters, , They represent the probability of the cosmic distance rate and the existence of cosmic wormholes, , The upper and lower boundaries of the wormhole; ; (3) , is the key parameter of MVO, and the calculation process is as follows: ; , The two extreme values ​​of the probability of the existence of a cosmic wormhole are: T and t is the total number of iterations and the current number of iterations, P To adjust the parameter, the smaller the value, the slower the search speed.

[0024] The BiLSTM principle is as follows: The BiLSTM model is an improvement on the LSTM model. A BiLSTM network model consists of a forward and a reverse LSTM network model. The specific model diagram is as follows: Figure 6 As shown, the implementation process is as follows: ; ; ; In the formula, i t 、f t 、o t Respectively represent the calculation results of the input gate, forget gate and output gate; W ih 、W fh 、W ox and b i 、b f 、 b o Represent the weight matrix and bias term of the corresponding gate respectively; Represents the sigmoid activation function.

[0025] Output at the current moment x t Combined with the previous moment input x t-1 , and combined according to certain weights to form the input for the next moment x t+1 , the input at the previous moment also includes t-1 Hidden layer cell state at the moment h t-1 and memory unit c t-1 , the output at the current moment contains the memory cell state c t With the hidden layer state h t 。

[0026] In the BiLSTM model t The output result of the moment memory module is determined by the output gate and the unit state: ; ; ; In the formula, Indicates time t The unit state input; tanh is the hyperbolic tangent activation function; W c , b c Represent the state weight matrix and bias term of the input layer respectively; It means that each element is multiplied by position.

[0027] The BiLSTM network model can perform bidirectional memory learning, which overcomes the shortcomings of poor feature item state information extraction that may occur in LSTM unidirectional training and learning. It can more fully explore the association information between feature sequences and plays an important role in data cleaning. The BiLSTM model has 2 hyperparameters: the number of hidden layer neurons m With learning rate ε , m Affects the model's fitting effect. ε Determines the training effect of the model.

[0028] Embodiment three: In this embodiment, the vibration swing statistical analysis module of the turbine generator set of the Three Gorges Hydropower Station is taken as an example to introduce. Figure 8 -Attached Fig.11 The results of the 26F hydro-generator unit vibration swing statistics and analysis module are displayed. By selecting the statistical analysis module of the large hydropower station unit operation data, and the corresponding time and unit number, the large hydropower station unit operation data analysis report in step S6 can be obtained; Fig.12 -Attached Fig.14 The results of the vibration swing statistics and analysis module of all units in the Three Gorges Power Station in February 2024 are displayed. The analysis results show that the overall equilibrium of the internal vibration, swing and pressure pulsation of the units in steady-state operation this month is normal. Some units have short-term secondary alarms due to special operating conditions of the units. After adjusting the operating conditions, they have returned to normal. Attachment Figure 8 -Attached Fig.14 The interruption of the middle curve graph and the missing part of the bar graph indicate that the unit is in shutdown state and no data is collected at present.

Claims

1. A smart operation method based on statistics and analysis of operation data of large hydropower station units, characterized in that: The following steps are involved: S1, determining the statistical analysis modules of the operating data of the large hydropower station units and the feature vector categories corresponding to each statistical analysis module; S2, based on the feature vector categories corresponding to each statistical analysis module, automatically extracts data from the computer monitoring system background database and other various online monitoring systems; S3, using the GWO-SDAE model to clean the extracted feature vectors and reconstruct and repair the abnormal point data; S4, setting the variation range of the eigenvector value corresponding to each statistical analysis module, and using the descriptive statistical model to perform a preliminary analysis on the cleaned eigenvector value. If the analysis result is consistent with the device parameter value range, proceed to the next step. If not, issue an alarm for inspection; S5, after descriptive statistical analysis of the eigenvector values, the MVO-BiLSTM hybrid model is used to learn the eigenvector values ​​and extract internal rules, so as to realize automatic and intelligent prediction and analysis of the eigenvalues; the collected eigendata values ​​are compared with the predicted values, and if the error exceeds the specified range, an alarm signal is issued in time and an inspection is carried out, otherwise the next step is carried out; S6, determine the format of the operation data analysis report of the large hydropower station unit, fill in the analysis conclusion of the characteristic value in step S5, further review, revise and finalize, and output the corresponding analysis report.

2. According to claim 1, a smart operation method based on statistics and analysis of operation data of large hydropower station units is characterized in that: In step S2, automatically extracting data from the computer monitoring system backend database and other various online monitoring systems includes: S201, according to the time difference of extracting characteristic information of different equipment in large hydropower stations, set extraction conditions corresponding to characteristic information of different equipment, the conditions include: N hours after the turbine generator set is shut down, N hours before the turbine generator set is started and connected to the grid, the turbine generator set is connected to the grid and generates electricity, the active power of the unit is greater than the preset value, the delay is Q hours, and N hours before the turbine generator set is shut down and completely stopped; The grid-connected power generation status of a hydro-generator unit refers to: The unit speed is greater than 95% of the rated speed, the main synchronous switch is in the closed position, the stator three-phase current is greater than 5% of the rated current, and the three-phase terminal voltage is greater than 90% of the rated voltage. The grid-connected shutdown state of a hydro-generator unit refers to: The unit speed is less than 1%, the guide vane is in the fully closed position, the main synchronous switch is in the open position, the automatic lock spindle is in operation, and the water system is shut down. S202, setting multiple sensor data collection methods, stipulating that data information is recorded once every 1 second, and each sensor collects 86,400 data information every day; all data information collected by all sensors of the hydropower station is transmitted to the computer monitoring database; S203, extracting effective feature information according to the time differentiation requirement for extracting feature information of different devices in step S201.

3. The intelligent operation method based on statistics and analysis of operation data of large hydropower station units according to claim 1 is characterized in that: In step S3, the method for constructing the GWO-SDAE model includes constructing a GWO model part: The parameter optimization process of the GWO model is divided into surrounding prey, chasing prey and attacking prey. First, the wolf pack is divided into four levels: α, β, δ and ω. The roles of the wolf packs at levels ω, δ, β, and α are: the main force of the encirclement target, determining the direction of the encirclement target, finding the encirclement target and the leading role in the encirclement process. The process of encircling the prey is: ； In the formula, , Represents the current position of the gray wolf and the position after iteration, is the location of the prey, D is the distance between the wolf and the prey, A and C are random vectors, is a linearly decreasing value between [2,0], , is a random number between [0, 1]; By iteratively updating the fitness value of the gray wolf, the optimal solution, excellent solution and suboptimal solution are obtained respectively: α, β, δ, ω The hunting direction and position update of the wolf are jointly determined by α, β, and δ. The specific formula is as follows: ； ； ； Where D1, D2 and D3 are α, β and δ respectively, the distance between the wolf and the prey; A1, A2, A 3、 C1, C2 and C3 are the random vectors of α, β and δ wolves respectively; X1, X2 and X3 are the current positions of α, β and δ wolves respectively. , and These are the prey locations of α, β, and δ wolves respectively.

4. The intelligent operation method based on statistics and analysis of operation data of large hydropower station units according to claim 3 is characterized in that: In step S3, the method for constructing the GWO-SDAE model includes constructing a SDAE data cleaning model: (1) In the initial feature sequence Randomly add noise data to get the latest feature sequence ; (2) For new feature sequences Encode and obtain the DAE features of the first layer : ； In the formula, , are the weight matrix and bias vector, is the sigmoid activation function; (3) Decode the DAE features of the first layer Perform reconstruction and repair to obtain the reconstruction result: ； In the formula, , are the weight matrix and bias vector of the decoder respectively; (4) Determine the reconstruction results With the initial feature sequence The error between MAPE Indicates the quality of model performance: ； Use micro gradient descent to further reduce and Error between: ； ； In the formula, x ij , z ij Respectively represent the values ​​before and after data cleaning using the SDAE model, a and b is the type and dimension of the feature vector set, and Respectively represent i Layer neurons pass n The weights and bias vectors after the residual iteration are calculated backwards. r and m Indicates the current training times and training number of samples, l r is the learning rate of the model; (5) According to the principle of layer-by-layer training of the SDAE model, continue to train and learn each layer of DAE separately, determine the size of the key parameters of each layer, and finally form a complete SDAE model.

5. The intelligent operation method based on statistics and analysis of operation data of large hydropower station units according to claim 4 is characterized in that: In step S3, the GWO-SDAE model is used to clean the extracted feature vectors and reconstruct and repair the abnormal point data. The specific process is as follows: S301, extracting initial feature data based on the data in the background database of the computer monitoring system; S302, randomly adding noise to the extracted initial feature data, and performing standardization processing, dividing the data into training set data and test set data according to a fixed ratio; S303, setting the maximum number of iterations of the GWO model, the population size, the value range of the number of hidden layer nodes and the noise coverage of the SDAE model, and the search speed range, and randomly initializing the population; S304, inputting the training set data into the SDAE data cleaning model to train each layer of SDAE; S305, calculate the fitness of the calculation particles and determine the reconstruction result With the initial feature sequence The error between MAPE Indicates the performance of the model. The smaller the fitness function value, the better the data cleaning effect. S306, updating the individual optimal position and the group optimal position of the particles of the GWO model, and continuously updating the velocity and displacement of the particles, and judging whether the GWO model has reached the termination condition. If the termination condition is met, the optimal number of hidden layer nodes and noise coverage are obtained, otherwise, the process returns to step S304 to continue iterative updating; the termination condition is set as the fitness value of the particles of the GWO model tends to be stable or the number of iterations reaches the maximum; S307, input the test set data into the reconstructed GWO-SDAE network data cleaning model, reconstruct and repair the abnormal point data, and finally perform denormalization on the results and evaluate them in combination with the evaluation indicators.

6. The intelligent operation method based on statistics and analysis of operation data of large hydropower station units according to claim 5 is characterized in that: In step S4, a preliminary analysis of the cleaned feature vector values ​​using a descriptive statistical model includes: The mean, median, mode, variance and standard deviation corresponding to each eigenvector value are calculated respectively, and the calculation results are compared with the corresponding setting range. If the value range specified by the equipment parameters is exceeded, an alarm is issued for inspection. If all eigenvector values ​​are within the specified range, it is judged that the large hydropower station unit is in good operating condition and the equipment can operate sustainably.

7. The intelligent operation method based on statistics and analysis of operation data of large hydropower station units according to claim 6 is characterized in that: In step S5, the MVO-BiLSTM hybrid model is used to learn the eigenvector values ​​after the preliminary analysis in step S4 and grasp the internal change rules, so as to realize the automatic and intelligent prediction and analysis of the eigenvalues. The specific process is as follows: The MVO-BiLSTM hybrid model is used to accurately predict the feature vector data, and the collected feature data values ​​are compared with the predicted values. If the error exceeds the specified range, an alarm signal is issued in time and manual inspection is performed. The specific process is as follows: S501, taking each feature vector value after preliminary analysis by the descriptive statistical model in step S4 as the input of the MVO-BiLSTM hybrid prediction model and performing normalization processing; S502, dividing the normalized data into a training set and a test set according to an 8:2 ratio; S503, setting the maximum number of iterations of the multiverse optimization algorithm MVO, the population size, the value range of the number of neurons m in the hidden layer of the BiLSTM model and the learning rate ε, and the search speed range, and initializing the universe population kernel parameters; S504, input the training set data into the BiLSTM network data prediction model and calculate , , calculate the expansion rate of each universe, implement the roulette mechanism to select cosmic objects, and continuously update the expansion rate of the universe; S505, calculate the particle fitness value and determine whether the iteration termination condition is met: If the termination condition is reached, the optimal number of hidden layer neurons is output. m With learning rate ε If the judgment condition is not met, then return to step S504 to iterate and optimize again; the iteration termination condition is that the particle fitness value tends to be flat or the number of iterations reaches the maximum; S506, using the obtained optimal hyperparameters to rebuild the MVO-BiLSTM hybrid prediction model, taking the test set data as input, performing iterative prediction, outputting the prediction results, and evaluating them in combination with the evaluation indicators. The formula is as follows: ； ； In the formula, y MAPE is the mean absolute percentage error, y RMSE is the root mean square error. The smaller the mean absolute percentage error and the root mean square error, the better the model prediction effect. X act ( i )and X pred ( i ) respectively represent i The actual value and predicted value at the moment; S506, compare the collected characteristic data value with the predicted value, and if the error exceeds the specified range, issue an alarm signal in time and conduct manual inspection.

Images