Pumped storage unit state trend prediction system

Abstract

The invention discloses a pumped storage group state trend prediction system, which comprises a state data acquisition module, a data primary decomposition and reconstruction module, a data secondary filtering module, a noise reduction effect evaluation module, a state trend prediction module and a prediction effect evaluation module, a new composite mapping group position initialization strategy and a new fitness function improved swarm intelligence optimization algorithm are constructed, multi-data decomposition method hyper-parameter optimization is carried out, a new scale adaptive mixed fuzzy scatter entropy SAMFSE index is designed, a decomposed IMF component is reconstructed, and a multi-scale adaptive mixed fuzzy scatter entropy SAMFSE index is designed. Performing secondary filtering processing on the reconstruction component of the state characteristic data of each subsystem by using a plurality of data filtering methods to realize final noise reduction on the state data; according to the method, an enhanced attention mechanism EAM comprehensively considering all time steps is provided, an EBixLSTM prediction model is established, the state trend prediction result of the pumped storage unit is predicted, the method can be used for implementing preventive state maintenance on the pumped storage unit, and the health operation and maintenance level of the pumped storage unit is improved.

Description

Technical Field

[0001] The present invention relates to the field of hydro-generator sets in power systems, and particularly to a state trend prediction system for pumped storage units. Background Art

[0002] In recent years, due to the increasingly severe climate change situation and the urgent need to develop the domestic economy with high quality, it has become an inevitable choice for our country to adjust the energy structure and develop clean energy such as hydropower. As a key equipment in the power production process, pumped storage units operate in a harsh environment for a long time and start and stop frequently. In addition, with the continuous advancement of the intelligent construction process of pumped storage power stations, the integration and comprehensiveness levels of pumped storage units have been gradually improved, which also puts forward higher requirements for the use and maintenance of the units.

[0003] The traditional maintenance method of pumped storage units mainly adopts a mode combining "post-maintenance" and "planned maintenance", and uses fixed algorithms for health assessment, but it does not solve the problems of limited prediction accuracy caused by insufficient optimization of non-stationary signal decomposition parameters and serious noise interference. Some studies introduce machine learning algorithms to construct early warning models, but still face defects such as insufficient feature extraction and weak time series correlation modeling when dealing with multi-source heterogeneous data. In addition, most of the existing data denoising methods adopt a single filtering strategy, and it is difficult to effectively separate periodic noise and non-stationary interference under complex working conditions. All of the above may lead to the inability to discover potential hidden dangers in a timely and accurate manner, and then lead to the inability to formulate reasonable and effective maintenance plans, resulting in economic losses. Therefore, there is an urgent need for a new, perfect and reasonable system architecture for the state trend prediction of pumped storage units to meet the functional requirements of implementing preventive state maintenance for pumped storage units. Summary of the Invention

[0004] Object of the Invention: To solve the problems mentioned in the background art, the present invention discloses a state trend prediction system for pumped storage units, aiming to solve the technical problems that the existing technology has not considered constructing a perfect and reasonable state trend prediction system for pumped storage units and cannot meet the requirements of implementing preventive state maintenance for the units.

[0005] Technical Solution:

[0006] The present invention discloses a state trend prediction system for pumped storage units. The system includes a state data acquisition module, which is used to use state monitoring sensors and in-machine self-checking devices to monitor the operating states of each subsystem of the pumped storage unit in real time and collect the state characteristic data of each subsystem of the pumped storage unit.

[0007] The primary data decomposition and reconstruction module, connected to the status data acquisition module, is used to construct a composite mapping group position initialization strategy and a fitness function to improve the swarm intelligence optimization algorithm, and design a scale adaptive hybrid fuzzy dispersion entropy SAMFSE index to perform primary decomposition and reconstruction on the subsystem status characteristic data;

[0008] The secondary data filtering module, connected to the primary data decomposition and reconstruction module, is used to perform secondary filtering on the reconstructed components of the subsystem status characteristic data by using data filtering methods to eliminate the non-periodic random noise of the reconstructed signal;

[0009] The noise reduction effect evaluation module, connected to the secondary data filtering module, is used to screen and achieve the optimal data noise reduction through evaluation indexes such as signal-to-noise ratio and root mean square error;

[0010] The status trend prediction module, connected to the noise reduction effect evaluation module, is used to propose an enhanced attention mechanism EAM, establish an EBixLSTM prediction model, combine it with a neural network model to form a parallel prediction structure, introduce CAM to fuse the parallel prediction results, and obtain the status trend prediction result of the pumped storage unit.

[0011] Further, in the primary data decomposition and reconstruction module, the construction of the composite mapping group position initialization strategy and the fitness function to improve the swarm intelligence optimization algorithm is as follows:

[0012] Construct a new composite mapping group position initialization strategy:

[0013]

[0014] Among them, n represents the total number of population individuals, P i and P i+1 represent the positions of the i-th and the (i + 1)-th population individuals, and α represents the composite mapping state value;

[0015] Based on the statistical principle and the time domain analysis theory, construct a new fitness function:

[0016] FITNESS = min(INDEX)

[0017]

[0018] Among them, FITNESS represents the new fitness function, INDEX represents the constructed new composite index, ESc(IMF) represents the envelope spectrum peak factor value of the decomposed IMF component, PCC(F(t), IMF) represents the Pearson correlation coefficient between the original signal F(t) and the IMF component, and PCC(IMF a , IMF a+1 ) represents the Pearson correlation coefficient between two adjacent IMF components.

[0019] Furthermore, a new scale adaptive mixed fuzzy dispersion entropy SAMFSE index is designed based on fuzzy dispersion entropy and multi-scale theory:

[0020]

[0021] where β represents the mixed fuzzy dispersion scale factor, SIMF b represents the b-th multi-scale sequence constructed based on the IMF component, B represents the length of the IMF component, FDE(SIMF) represents the fuzzy dispersion entropy value of the multi-scale sequence, c represents the number of categories, and d represents the embedding dimension;

[0022] Based on the calculated SAMFSE value of the IMF component, component reconstruction is completed through the following formula:

[0023]

[0024] where RIMF p represents the reconstructed component, and represent the SAMFSE values of the p-th and p+q-th reconstructed components, and e represents the number of IMF components.

[0025] Furthermore, the state trend prediction module designs the EBixLSTM prediction model as follows:

[0026] Divide the dataset corresponding to the optimal noise reduction result into a training set and a test set;

[0027] Considering all time steps of the BixLSTM model comprehensively, a new scoring function and an attention weight allocation method are proposed, a reinforcement attention mechanism EAM is designed, and an EBixLSTM prediction model is established:

[0028]

[0029] where, represents the new scoring function, y t represents the output of the hidden state at time step t of BixLSTM, represents the output of the hidden state at time step jt of BixLSTM, W aj and χ aj represent learnable weight matrices and parameters, [;] represents vector concatenation, aj t (jt) represents the weight of the output of the hidden state at time step jt of BixLSTM in all hidden state outputs, aj t represents the combination of the weights of the outputs of the hidden states at all t time steps of BixLSTM, g t(jt) represents the output vector of the BixLSTM time step jt after the enhanced attention mechanism EAM assigns weights, g t represents the output vectors of all t time steps of the BixLSTM after the enhanced attention mechanism EAM assigns weights.

[0030] Furthermore, the EBixLSTM model is combined with a neural network model to form a parallel prediction structure, and the CAM is introduced to fuse the parallel prediction results, and the state trend prediction result of the pumped-storage unit is obtained through the final model:

[0031]

[0032] Among them, S PM and S EBixLSTM represent the feature sequences extracted by the neural network model PM and the EBixLSTM model, and represent the query sequence, key sequence and value sequence corresponding to the extracted feature sequence, represents and the similarity between, D represents the input dimension, represents the CAM weight, is the final fusion result of the feature sequences extracted by the neural network model PM and the EBixLSTM model.

[0033] Furthermore, the system further includes a prediction effect evaluation module, which is connected to the state trend prediction module and is used to select the method with the best prediction effect through evaluation indicators such as mean absolute error, mean absolute percentage error, root mean square error, and correlation coefficient, so as to achieve the optimal prediction of the state trend of the pumped-storage unit.

[0034] Furthermore, each subsystem of the pumped-storage unit includes a pump-turbine system, a generator motor and its excitation system, a vibration and swing system, a transformer system, a static frequency converter system, a governor system, a phase modulation and water pressure system, an inlet valve system, a bearing cooling system, a compressed air system and a turbine oil system.

[0035] Beneficial effects:

[0036] 1. The present invention comprehensively depicts the core steps in the implementation of state maintenance of the unit by using a state data acquisition module, a data primary decomposition and reconstruction module, a data secondary filtering module, a noise reduction effect evaluation module, a state trend prediction module and a prediction effect evaluation module. The system is formed in a modular manner, effectively reducing the complexity of the design of the state trend prediction system, improving the execution level of the relevant functions of the pumped-storage unit and each subsystem, and ensuring the safe, reliable and efficient operation of the power plant.

[0037] 2. The present invention systematically establishes a new population position initialization strategy and a new fitness function for the swarm intelligence optimization algorithm, and proposes a new feature component reconstruction index SAMFSE. The new feature component reconstruction index accurately measures the volatility of the state data of the pumped-storage unit. By component reconstruction, the complexity of the data is reduced, and the primary decomposition and reconstruction of the data improve the prediction accuracy of the subsequent prediction model.

[0038] 3. The state trend prediction module proposed by the present invention designs a new scoring function and an attention weight allocation method based on the traditional attention mechanism, constructs an EBixLSTM prediction model, and forms a parallel prediction structure to deeply extract the spatio-temporal features of the state data of the pumped-storage unit. Finally, the CAM attention mechanism is used to fuse the parallel prediction results, improving the prediction accuracy and enhancing the robustness at the same time.

[0039] 4. The swarm intelligence optimization algorithm, signal decomposition algorithm, and prediction method involved in the data primary decomposition and reconstruction module and the state trend prediction module proposed by the present invention can be increased, decreased, and updated in real time, and the method effects in the module can be optimized in real time, realizing the long-term update and iteration of the state trend prediction system of the pumped-storage unit, outputting the best possible prediction effect, and maintaining the quality of the maintenance plan. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a schematic diagram of the system flow of the present invention;

[0041] Figure 2 is a schematic diagram of the state trend prediction system of the pumped-storage unit in Embodiment 1 of the present invention;

[0042] Figure 3 is a flow chart of the vibration and swing signal trend prediction of the pumped-storage unit in Embodiment 2 of the present invention;

[0043] Figure 4 is the vibration and swing signal trend prediction result of the pumped-storage unit in Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0045] As Figure 1 shown, a state trend prediction system for a pumped-storage unit includes:

[0046] The status data acquisition module is used to use status monitoring sensors and built-in self-test equipment to monitor the operating status of each subsystem of the pumped-storage unit in real time and collect the status characteristic data of each subsystem of the pumped-storage unit;

[0047] The primary data decomposition and reconstruction module is used to construct a new composite mapping group position initialization strategy and a new fitness function to improve the swarm intelligence optimization algorithm, optimize the hyperparameters of various data decomposition methods, and design a new scale adaptive hybrid fuzzy dispersion entropy SAMFSE index to reconstruct the decomposed IMF components, thereby realizing the primary decomposition and reconstruction of the status characteristic data of each subsystem;

[0048] The secondary data filtering module is used to perform secondary filtering on the reconstructed components of the status characteristic data of each subsystem by using various data filtering methods to eliminate the non-periodic random noise in the reconstructed signal and achieve the final noise reduction of the status data;

[0049] The noise reduction effect evaluation module is used to select the method with the best noise reduction effect through evaluation indexes such as signal-to-noise ratio and root mean square error to achieve the optimal noise reduction of the data;

[0050] The status trend prediction module is used to propose a comprehensive attention mechanism EAM considering all time steps, establish an EBixLSTM prediction model, combine it with multiple prediction models to form a parallel prediction structure, introduce CAM to fuse the parallel prediction results, and obtain the status trend prediction result of the pumped-storage unit;

[0051] The prediction effect evaluation module is used to select the method with the best prediction effect through evaluation indexes such as mean absolute error, mean absolute percentage error, root mean square error, and correlation coefficient to achieve the optimal prediction of the status trend of the pumped-storage unit;

[0052] Each subsystem of the pumped-storage unit includes a pump-turbine system, a generator motor and its excitation system, a vibration and swing system, a transformer system, a static frequency converter system, a governor system, a phase modulation and water compression system, an inlet valve system, a bearing cooling system, a compressed air system, and a turbine oil system.

[0053] The specific implementation method of the status data acquisition module is as follows:

[0054] Use status monitoring sensors and built-in self-test equipment to monitor the operating status of each subsystem of the pumped-storage unit in real time. As shown in Table 1, it includes: using status monitoring sensors and built-in self-test equipment to monitor the water level, pressure, speed, flow rate, temperature, head, and efficiency of the pump-turbine system in real time; using status monitoring sensors and built-in self-test equipment to monitor the motor current, voltage, power, excitation current, voltage, power, temperature, vibration, and noise of the generator motor and its excitation system in real time; using status monitoring sensors and built-in self-test equipment to monitor the frame vibration, stator vibration, top cover vibration, bearing vibration, bearing swing, and spindle swing of the vibration and swing system in real time; using status monitoring sensors and built-in self-test equipment to monitor the three-phase current, active power, reactive power, temperature, and pressure of the transformer system in real time; using status monitoring sensors and built-in self-test equipment to monitor the current, voltage, speed, rotor position signal, power factor, harmonic content, temperature, and pressure of the static frequency converter system in real time; using status monitoring sensors and built-in self-test equipment to monitor the guide vane opening, rotor speed, temperature, and pressure of the governor system in real time; using status monitoring sensors and built-in self-test equipment to monitor the water level, temperature, pressure, and flow rate of the phase modulation and water compression system in real time; using status monitoring sensors and built-in self-test equipment to monitor the temperature, pressure, flow rate, and vibration of the inlet valve system in real time; using status monitoring sensors and built-in self-test equipment to monitor the temperature, pressure, and flow rate of the bearing cooling system in real time; using status monitoring sensors and built-in self-test equipment to monitor the temperature, pressure, flow rate, and digital input / output signals of the compressed air system in real time; using status monitoring sensors and built-in self-test equipment to monitor the viscosity, acid value, saponification value, temperature, pressure, and flow rate of the turbine oil system in real time; using status monitoring sensors and built-in self-test equipment to monitor the water quality and abnormal unit drive of other systems in real time, and finally collect the state characteristic data of each subsystem of the pumped-storage unit;

[0055] Table 1

[0056]

[0057] Table 1 designs the corresponding relationship between each subsystem of the pumped-storage unit and the types of monitoring data, making up for the lack of sufficient types of monitoring data for reference in the status monitoring of each subsystem of the pumped-storage unit, and can provide a complete data reserve for the implementation of the status trend prediction of the pumped-storage unit.

[0058] Example 1

[0059] Such as Figure 2The figure shows a schematic diagram of a state trend prediction system for a pumped - storage unit proposed in Embodiment 1 of the present invention, which includes a total of 6 levels: a state data acquisition layer, a primary data decomposition and reconstruction layer, a secondary data filtering layer, a noise reduction effect evaluation layer, a state trend prediction layer, and a prediction effect evaluation layer. Descriptions are given separately below.

[0060] State data acquisition layer;

[0061] State monitoring sensors and in - machine self - inspection devices are arranged in each subsystem of the pumped - storage unit (pump - turbine system, generator - motor and its excitation system, transformer system, static frequency converter system, governor system, phase - modulation and water - pressuring system, inlet valve system, bearing cooling system, compressed air system, and turbine oil system) to monitor the real - time operating state of each subsystem and collect state characteristic data.

[0062] Primary data decomposition and reconstruction layer;

[0063] Construct a new composite mapping group position initialization strategy and a new fitness function to improve the swarm intelligence optimization algorithm, perform hyperparameter optimization for various data decomposition methods, and design a new scale - adaptive hybrid fuzzy dispersion entropy SAMFSE index to reconstruct the decomposed IMF components, thereby realizing the primary decomposition and reconstruction of the state characteristic data of each subsystem.

[0064] Secondary data filtering layer;

[0065] Use various data filtering methods to perform secondary filtering on the reconstructed components of the state characteristic data of each subsystem, eliminate the non - periodic random noise in the reconstructed signal, and achieve the final noise reduction of the state data.

[0066] Noise reduction effect evaluation layer;

[0067] Select the method with the best noise reduction effect through evaluation indicators such as signal - to - noise ratio and root - mean - square error to achieve the optimal noise reduction of the data.

[0068] State trend prediction layer;

[0069] Propose a comprehensive consideration of all time - step enhanced attention mechanism EAM, establish an EBixLSTM prediction model, combine it with multiple prediction models to form a parallel prediction structure, introduce CAM to fuse the parallel prediction results, and obtain the state trend prediction result of the pumped - storage unit.

[0070] Prediction effect evaluation layer;

[0071] Select the method with the best prediction effect through evaluation indicators such as mean absolute error, mean absolute percentage error, root - mean - square error, and correlation coefficient to achieve the optimal prediction of the state trend of the pumped - storage unit.

[0072] Embodiment 2

[0073] In this embodiment, the state trend prediction of the swing of the water guide bearing in the monitoring data of the vibration and swing system of a pumped-storage unit in a domestic power station is carried out. As Figure 3 shown in the flowchart of the method for predicting the trend of the vibration and swing signal of the pumped-storage unit provided by the present invention, the method specifically includes the following steps:

[0074] Step 1: Use state monitoring sensors and in-machine self-checking equipment to monitor the operating state of the vibration and swing system of the pumped-storage unit in real time, and collect the swing data of the water guide bearing in the Y direction of the vibration and swing system of the pumped-storage unit.

[0075] Step 2: Perform parameter-improved variational mode decomposition (PMVMD) and parameter-improved successive variational mode decomposition (PMSVMD) on the time-series data of the unit vibration and swing signals, and obtain multiple sets of intrinsic mode components {IMF1, IMF2,..., IMF n}.

[0076] (2-1) Take the No. 1 unit of a domestic pumped-storage power station as an analysis case to obtain the original time-series curve of the swing signal of the water guide bearing in the Y direction of the unit.

[0077] (2-2) Perform parameter-improved variational mode decomposition (PMVMD) and parameter-improved successive variational mode decomposition (PMSVMD) on the time-series data of the unit state characteristics, and use the improved Newton-Raphson optimization algorithm (INRBO) and the improved goose optimization algorithm (IGOOSE) to select the optimal parameters in signal decomposition, and obtain multiple sets of intrinsic mode components {IMF1, IMF2,..., IMF n}.

[0078] Through iteration, the optimal parameters of the variational mode decomposition and successive variational mode decomposition methods can be selected. The optimal parameters of different methods are shown in Table 2, where K is the decomposition order and λ is the penalty factor.

[0079] Table 2

[0080]

[0081] The modal components obtained after the vibration and swing signals of the pumped-storage unit are decomposed by PMVMD and PMSVMD. The original time series are respectively decomposed into 6, 4, 7, and 4 modal components.

[0082] Step 3: According to the entropy value analysis theory method, screen the time-series data that can effectively characterize the unit state characteristics from the unit vibration and swing signal data, and perform recombination of multiple modal components to generate a reconstructed characteristic signal.

[0083] (3-1) According to the entropy value reconstruction theory, calculate the SAMFSE of each modal component, and the calculation results are shown in Table 3.

[0084] Table 3

[0085]

[0086]

[0087] (3-2) Based on the SAMFSE values of the calculated IMF components, the component reconstruction is completed through the following formula:

[0088]

[0089] where RIMF p represents the reconstructed component, and represent the SAMFSE values of the p-th and (p+q)-th reconstructed components, and e represents the number of IMF components.

[0090] As can be seen from the above formula, according to the distribution of SAMFSE values, multiple intrinsic mode components are respectively reconstructed into 3, 2, 3, and 2 characteristic components, as shown in Table 4.

[0091] Table 4

[0092]

[0093] The reconstructed characteristic signal curves of the multiple intrinsic mode components after reconstruction significantly reduce the complexity of the original signal.

[0094] Step 4: Realize the secondary filtering of the reconstructed characteristic signal through singular value decomposition (SVD) and wavelet transform (WT), and select the optimal denoising method based on two quantitative evaluation indexes of signal-to-noise ratio (SNR) and root mean square error (RMSE).

[0095] (4-1) Respectively take the reconstructed characteristic signals of PMVMD and PMSVMD as the inputs of SVD and WT to eliminate the non-periodic random noise in the reconstructed signals and obtain the final denoising results.

[0096] (4-2) Calculate the SNR and RMSE of the finally denoised signals after SVD and WT, compare and select the method with better denoising effect to achieve the optimal denoising of the vibration and swing signals of the pumped storage unit.

[0097] As shown in Table 5 are the calculation results of SNR and RMSE of the finally denoised results of eight methods. When the SNR value is large and the RMSE value is small, it indicates that the noise proportion is small. By comparison, it can be seen that the denoising effect of the IGOOSE-PMVMD-SVD method is better. Therefore, the optimal denoising result of this method is used as the input of the subsequent state trend prediction method.

[0098] Table 5

[0099]

[0100] Step 5: The convolutional neural network (CNN) and the temporal convolutional network (TCN) prediction models are respectively combined with the EBixLSTM prediction model to form a parallel prediction structure, and the CAM is introduced to fuse the parallel prediction results, so as to obtain the state trend prediction result of the pumped-storage unit.

[0101] (5-1) The optimal denoising results of the IGOOSE-PMVMD-SVD method are respectively used as the inputs of the parallel prediction models CNN-EBixLSTM-CAM and TCN-EBixLSTM-CAM. Through data analysis, constructing a data set, data set analysis, dividing the training set and the test set, data normalization, creating a network and setting parameters, simulation testing, data denormalization and other steps, the state trend prediction of the optimal denoising results is realized.

[0102] Based on the model training results of the training set, the CNN-EBixLSTM-CAM and TCN-EBixLSTM-CAM methods have a high fitting accuracy, the error between the predicted value and the original value is small, and the model training effect is good.

[0103] As Figure 4 shown are the trend prediction results of the vibration and swing signals of the pumped-storage unit based on the test set by the CNN-EBixLSTM-CAM and TCN-EBixLSTM-CAM methods. From the results shown in the figure, it can be seen that the final prediction curve can well fit the state change curve of the original vibration and swing signal, verifying the effectiveness of the two state trend prediction methods.

[0104] (5-2) Calculate the prediction accuracy indexes of the two parallel prediction methods, including the mean absolute error (MAE), the root mean square error (RMSE), the mean absolute percentage error (MAPE) and the correlation coefficient (R) close to 1. The comparison results of the prediction accuracy are shown in Table 6.

[0105] Table 6

[0106]

[0107] As can be seen from the above data, for the vibration and swing signals of the selected pumped-storage units, the CNN-EBixLSTM-CAM method performs best in all metrics. For both the training set and the test set, it has the smallest MAE, RMSE, and MAPE, and at the same time, the correlation coefficient R is also closest to 1. This means that this method can provide more accurate and reliable prediction results, and the difference between the predicted value and the original value is very small, and it can effectively complete the trend prediction of the vibration and swing signals of the pumped-storage units. Therefore, CNN-EBixLSTM-CAM is the optimal trend prediction method for the vibration and swing signals of the pumped-storage units.

[0108] So far, the optimal methods for the trend prediction of the vibration and swing signals of the pumped-storage units include the optimal noise reduction method IGOOSE-PMVMD-SVD and the optimal trend prediction method CNN-EBixLSTM-CAM.

[0109] The above description of the embodiments enables those skilled in the art to implement or use the present invention. Various modifications to the embodiments will be obvious to those skilled in the art. The general principles of the present invention can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention should not be limited to the embodiments shown herein, but should cover the widest range that conforms to the principles and novel features disclosed in the present invention.

Claims

1. A state trend prediction system for a pumped-storage unit, characterized in that, The system includes: A state data acquisition module, which is used to use state monitoring sensors and built-in self-test equipment to monitor the operating states of each subsystem of the pumped-storage unit in real time and collect the state characteristic data of each subsystem of the pumped-storage unit; A primary data decomposition and reconstruction module, connected to the state data acquisition module, which is used to construct a composite mapping group position initialization strategy and a fitness function-improved swarm intelligence optimization algorithm, and design a scale-adaptive hybrid fuzzy dispersion entropy SAMFSE index to perform primary decomposition and reconstruction on the subsystem state characteristic data; A secondary data filtering module, connected to the primary data decomposition and reconstruction module, which is used to perform secondary filtering on the reconstructed components of the state characteristic data of each subsystem by using data filtering methods to eliminate the non-periodic random noise of the reconstructed signal; A noise reduction effect evaluation module, connected to the secondary data filtering module, which is used to screen and realize the optimal noise reduction of the data through evaluation indexes such as signal-to-noise ratio and root mean square error; A state trend prediction module, connected to the noise reduction effect evaluation module, which is used to propose an enhanced attention mechanism EAM, establish an EBixLSTM prediction model, combine it with a neural network model to form a parallel prediction structure, introduce CAM to fuse the parallel prediction results, and obtain the state trend prediction result of the pumped-storage unit.

2. The pumped-storage unit state trend prediction system according to claim 1, wherein In the primary data decomposition and reconstruction module, the construction of the composite mapping group position initialization strategy and the fitness function-improved swarm intelligence optimization algorithm is as follows: Construct a new composite mapping group position initialization strategy: Among them, n represents the total number of population individuals, P i and P i+1 represent the positions of the i-th and (i + 1)-th population individuals, and α represents the composite mapping state value; Based on the statistical principle and time-domain analysis theory, construct a new fitness function: FITNESS = min(INDEX) Among them, FITNESS represents the new fitness function, INDEX represents the newly constructed composite index, ESc(IMF) represents the envelope spectrum peak factor value of the IMF component after decomposition, PCC(F(t), IMF) represents the Pearson correlation coefficient between the original signal F(t) and the IMF component, and PCC(IMF a , IMF a+1 ) represents the Pearson correlation coefficient between two adjacent IMF components.

3. The pumped-storage unit state trend prediction system according to claim 2, wherein Design a new scale-adaptive hybrid fuzzy dispersion entropy SAMFSE index according to the fuzzy dispersion entropy and multi-scale theory: Among them, β represents the mixed fuzzy dispersion scale factor, SIMF b represents the b-th multi-scale sequence constructed based on the IMF component, B represents the length of the IMF component, FDE(SIMF) represents the fuzzy dispersion entropy value of the multi-scale sequence, c represents the number of categories, and d represents the embedding dimension; Based on the SAMFSE values of the calculated IMF components, complete the component reconstruction through the following formula: Among them, RIMF p represents the reconstructed component, and represent the SAMFSE values of the p-th and (p + q)-th reconstructed components, and e represents the number of IMF components.

4. The pumped-storage unit state trend prediction system according to claim 1, wherein The state trend prediction module designs the EBixLSTM prediction model as follows: Divide the data set corresponding to the optimal noise reduction result into a training set and a test set; Comprehensively consider all time steps of the BixLSTM model, propose a new scoring function and attention weight allocation method, design an enhanced attention mechanism EAM, and establish an EBixLSTM prediction model: Among them, represents the new scoring function, y t represents the output of the hidden state at time step t of BixLSTM, represents the output of the hidden state at time step jt of BixLSTM, W aj and χ aj represent learnable weight matrices and parameters, [;] represents vector concatenation, aj t (jt) represents the weight of the output of the hidden state at time step jt of BixLSTM among all hidden state outputs, aj t represents the combination of the weights of the hidden state outputs of all t time steps of BixLSTM, g t (jt) represents the output vector of BixLSTM at time step jt after the enhanced attention mechanism EAM assigns weights, g t represents the output vector of BixLSTM at all t time steps after the enhanced attention mechanism EAM assigns weights.

5. The pumped-storage unit status trend prediction system according to claim 4, characterized in that, Combine the EBixLSTM model with a neural network model to form a parallel prediction structure, introduce CAM to fuse the parallel prediction results, and obtain the state trend prediction result of the pumped-storage unit through the final model: Among them, S PM and S EBixLSTM represent the feature sequences extracted by the neural network model PM and the EBixLSTM model, and represent the query sequence, key sequence, and value sequence corresponding to the extracted feature sequence, represents and the similarity between them, D represents the input dimension, represents the CAM weight, is the final fusion result of the feature sequences extracted by the neural network model PM and the EBixLSTM model.

6. The pumped-storage unit status trend prediction system according to claim 1, wherein The system further includes a prediction effect evaluation module, connected to the state trend prediction module, which is used to select the method with the best prediction effect through evaluation indexes such as mean absolute error, mean absolute percentage error, root mean square error, and correlation coefficient to achieve the optimal prediction of the state trend of the pumped-storage unit.

7. The pumped-storage unit status trend prediction system according to claim 1, characterized in that, Each subsystem of the pumped-storage unit includes a pump-turbine system, a generator motor and its excitation system, a vibration and swing system, a transformer system, a static frequency converter system, a governor system, a phase modulation and water compression system, an inlet valve system, a bearing cooling system, a compressed air system, and a turbine oil system.

Images