A method for online determination of the operational health status of a central gate under throttling heating boundary conditions

Abstract

This invention discloses an online method for judging the operational health status of the central valve under throttling heating boundary conditions. By extracting information from characteristic parameters under throttling conditions, this invention comprehensively integrates and utilizes multi-source information, sequentially extracting multi-dimensional fault state information across categories for individual characteristic parameters, ensuring the accuracy and completeness of fault state information extraction. Furthermore, this invention achieves the identification and prediction of fault state-sensitive physical quantities through dimensionality reduction decomposition of the array matrix. Combining artificial neural network sample training with targeted focusing on expert databases, this invention can achieve high-precision evaluation of the operational health status of the central valve and real-time online judgment of its operational health status under throttling heating boundary conditions, while significantly reducing the amount of sample data used. This is of great significance for promoting the safe and reliable operation of heating systems under partial load conditions.

Description

Technical Field

[0001] This invention belongs to the field of steam turbine power generation technology, specifically relating to an online method for determining the operational health status of a central valve under throttling heating boundary conditions. Background Technology

[0002] As wind and solar power account for an increasingly larger share of the power system and gradually become the main component, the significant randomness, intermittency, and volatility of renewable energy pose a huge challenge to the stability and security of the power system. Enhancing the flexibility of the power system has become key to building a new type of power system. Deep peak shaving will become the norm for thermal power units, laying the foundation for large-scale grid integration of new energy sources.

[0003] High-temperature steam drawn from cold or hot reheat pipelines by thermal power units, after being desuperheated and depressurized, is supplied to heat users. This is an important and widely used method of industrial heating in cogeneration. However, with the normalization of deep peak shaving by the units, on the one hand, the sliding pressure of the cold or hot reheat steam will gradually decrease during operation; on the other hand, due to the influence of boiler reheater heat balance and unit axial thrust, technical problems often arise at the low and medium load boundaries where the pressure and flow of the cogeneration unit's heating main pipe do not meet user needs.

[0004] To ensure user demand under deep peak-shaving conditions, a throttling operation involving the intermediate valve can be adopted. The technical method involves adjusting the opening of the turbine's intermediate valve to throttle the steam, causing the steam entering the turbine's intermediate-pressure cylinder to stagnate in a limited local space before the valve, creating a "pressure buildup" effect. Under the influence of pressure wave energy transfer, this increases the steam pressure in the cold or hot reheat pipelines. Furthermore, after throttling the intermediate valve, the work capacity in the low-pressure flow zone of the unit decreases, leading to a reduction in load. This not only increases heating capacity but also further promotes peak-shaving depth and enhances the unit's thermoelectric decoupling capability.

[0005] Considering that the initial design boundary of conventional large-capacity thermal power units is to carry the grid reference load, the intermediate valve of the unit will be in a normally open and non-adjustable state under normal operation. Therefore, the existing identification method is mainly for the normal operation state of the unit and does not involve the operation condition in which the intermediate valve participates in regulation. Summary of the Invention

[0006] To address the current issue of not addressing the assessment of unit operating health status under the current normalized long-cycle deep peak shaving environment, this invention provides an online assessment method for the operating health status of the central valve under the throttling heating boundary.

[0007] A method for online determination of the operational health status of a central gate under throttling heating boundary conditions includes the following steps:

[0008] S1: Under throttling conditions, feature parameters are collected, and multi-dimensional fault state information is extracted from individual feature parameters in sequence to obtain time series data;

[0009] S2: Perform dimensionality reduction decomposition on the acquired time series data using an array matrix;

[0010] S3: Combine artificial neural networks to perform batch training and periodic training on the extreme boundary data in the time series data after dimensionality reduction decomposition, and complete the learning of the artificial neural network on the training data to obtain the fault state information after dimensionality reduction decomposition.

[0011] S4: Combining expert practical experience, a fault classification and health identification model is constructed. The fault status information extracted based on the feature parameter input after dimensionality reduction decomposition is compared with the feature values ​​of the health identification model with a wide threshold, and the evaluation level is output to realize the assessment of the operational health status of the central gate.

[0012] The characteristic parameters in S1 include the differential pressure value before and after the central valve, the vibration value of the central valve stem, the stress value of the central valve disc, and the oil pressure value of the high-pressure EH oil supply system.

[0013] The specific steps for extracting multi-dimensional fault state information across categories from individual feature parameters in S1 are as follows:

[0014] S1-1: Establish a mathematical model of individual feature parameters, input feature parameters, and extract feature frequencies of distributed and local faults;

[0015] S1-2: Perform time-domain and frequency-domain statistical analysis;

[0016] S1-3: Based on the multi-wavelet analysis method, kurtosis and entropy are used to select the optimal multi-wavelet and extract the pulse index and margin index statistical features of the waveform.

[0017] The acquired time series data is decomposed into array matrix dimensionality using a multi-objective evolutionary clustering algorithm.

[0018] The specific steps for dimensionality reduction decomposition of array matrices are as follows:

[0019] S2-1 measures the similarity of a finite number of unlabeled time series data by calculating the distance between data points;

[0020] S2-2 relies on a clustering objective function to perform clustering discrimination on time series data, obtain the distinguishability between different classes of data, and obtain the optimal clustering partition by optimizing the objective function;

[0021] S2-3 involves clustering the high-dimensional time series data to find the specific feature subspaces of each data type and assigning weights to the relevant feature dimensions.

[0022] Based on the Pareto dominance relationship, S2-4 performs multi-objective optimization feasibility analysis by sorting multiple objective functions according to their relative importance and solving them sequentially according to their weight coefficients, thus transforming the multi-objective optimization problem into a series of single-objective optimization problems.

[0023] The distance between data points is defined using the Euclidean distance criterion.

[0024] The method for assigning weights to relevant feature dimensions is to decrease the weight coefficients of non-interested features and increase the weight coefficients of interested features, where the weight coefficient formula is:

[0025]

[0026]

[0027] The multi-objective optimization problem is transformed into a series of single-objective optimization problems, defined as follows:

[0028] minF(x)=f1(x),2(x),……f m (x)) T

[0029] T

[0030] st(x) = (x1, x2, ..., n )

[0031] Where x represents the decision variable, st(x) is the n-dimensional decision space, and minF(x) is the set of objective functions. This set of objective functions constitutes the m-dimensional objective space and maps the points in the decision space to the objective space.

[0032] The specific steps for batch training and periodic training of the extreme boundary data in the time series data after dimensionality reduction decomposition are as follows:

[0033] S3-1, Establish forward propagation, and calculate the final output value and the loss value between the output value and the actual value based on the input sample data, the given initial weight values ​​and bias terms.

[0034] S3-2, establish backpropagation, propagate the output back to the input layer layer by layer through the hidden layer in a certain form, and distribute the error to all units in each layer to obtain the error signal of each unit. This error signal is used as the basis for correcting the weights of each unit.

[0035] S3-3 uses gradient descent to update the neural network's own variables, then resubmits these variables into the neural network, adds all the training data, and obtains the updated error function value. This process is repeated until the error function is reduced to a set range, thus completing the neural network's learning of the training data.

[0036] An online health status assessment system for a central gate under throttling heating boundary conditions based on the above method includes a data acquisition module, an array matrix decomposition module, a boundary sample training module, and an operation status assessment module.

[0037] The data acquisition module is used to collect characteristic parameters under throttling conditions, and sequentially extract multi-dimensional fault state information across categories from the individual characteristic parameters to obtain time series data.

[0038] The array matrix decomposition module is used to perform array matrix dimensionality reduction decomposition on the acquired time series data;

[0039] The boundary sample training module is used to perform batch and periodic training on the extreme boundary data in the time series data after dimensionality reduction decomposition, and to complete the learning of the artificial neural network on the training data to obtain the fault state information after dimensionality reduction decomposition.

[0040] The operational status assessment module is used to combine expert practical experience to construct a fault classification and health identification model. It compares the extracted fault status information after dimensionality reduction decomposition based on feature parameter input with the feature values ​​of the health identification model with a wide threshold, outputs the evaluation level, and realizes the assessment of the operational health status of the central gate.

[0041] Compared with existing technologies, the beneficial effects of this invention are as follows: Through its implementation, this invention enables safety early warning and assessment of the turbine's main operation, key components, and major auxiliary systems during throttling heating. This invention comprehensively integrates and utilizes multi-source information by extracting information from characteristic parameters under throttling conditions, sequentially extracting multi-dimensional fault state information across categories for individual characteristic parameters, ensuring the accuracy and completeness of fault state information extraction. This invention achieves the identification and prediction of fault-sensitive physical quantities through dimensionality reduction decomposition of array matrices. This invention combines artificial neural network sample training with targeted focusing on expert databases, significantly reducing the amount of sample data used, enabling high-precision evaluation of the healthy operation status of the turbine's main operating system, and real-time online assessment of the healthy operation status of the main operating system under throttling heating boundaries, which is of great significance for promoting the safe and reliable operation of the unit under partial load conditions. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below:

[0043] Figure 1 This is a flowchart of the online method for determining the operational health status of the central connecting gate under the throttling heating boundary of the present invention. Detailed Implementation

[0044] The present invention will be further described below with reference to the accompanying drawings:

[0045] Taking a certain unit as an example, such as Figure 1 As shown, an online method for determining the operational health status of the central connecting valve under throttling heating boundary conditions is proposed:

[0046] S1: Collect the differential pressure value before and after the central valve under throttling conditions, the vibration value of the central valve stem, the stress value of the central valve disc, and the oil pressure value of the high-pressure EH oil supply system as characteristic parameters input, and extract multi-dimensional fault status information across categories for individual characteristic parameters in turn to obtain time series data;

[0047] S1-1 Establishes a mathematical model of individual feature parameters, inputs feature parameters, and extracts feature frequencies of distributed and local faults;

[0048] S1-2 performs time-domain and frequency-domain statistical analysis;

[0049] S1-3 is based on the multi-wavelet analysis method, which uses kurtosis and entropy to select the optimal multi-wavelet and extracts statistical features of the waveform such as impulse index and margin index.

[0050] S2: Data mining is performed on the acquired time series data using a multi-objective evolutionary clustering algorithm to achieve dimensionality reduction decomposition of the array matrix;

[0051] S2-1 measures the similarity of a finite number of unlabeled time series data by calculating the distance between data points. The distance between data points is defined using the Euclidean distance criterion. The expression for the Euclidean distance is:

[0052]

[0053] In the formula d i,j x is the node spacing. ik Let x be the coordinates of the target node. jk Euclidean coordinates;

[0054] S2-2 relies on a clustering objective function to perform clustering discrimination on time series data, obtain the discriminative power between different classes of data, and obtain the optimal clustering partition by optimizing the objective function. The objective function is as follows:

[0055]

[0056] S2-3 For clustering the high-dimensional time series data, we find the specific feature subspaces of each data type, assign weights to relevant feature dimensions, reduce the weight coefficients of non-interesting features and increase the weight coefficients of interested features, so that the discriminative power of the time series data clustering is clear and reasonable. The formula for the weight coefficients is as follows:

[0057]

[0058]

[0059] Based on the Pareto dominance relationship, S2-4 performs multi-objective optimization feasibility analysis. Multiple objective functions are ranked according to their relative importance, and then solved sequentially based on their weight coefficients. This transforms the multi-objective optimization problem into a series of single-objective optimization problems, defined as follows:

[0060] minF(x)=f1(x),2(x),……f m (x)) T

[0061] T

[0062] st(x) = (x1, x2, ..., n )

[0063] Where x represents the decision variable, st(x) is the n-dimensional decision space, and minF(x) is the set of objective functions. This set of objective functions constitutes the m-dimensional objective space and maps the points in the decision space to the objective space.

[0064] S3: Combine artificial neural networks to perform batch training and periodic training on the extreme boundary data in the time series data after dimensionality reduction decomposition, and complete the learning of the artificial neural network on the training data to obtain the fault state information after dimensionality reduction decomposition.

[0065] S3-1, Establish forward propagation: Based on the input time series data, given initial weight values ​​and bias terms, calculate the final output value and the loss value between the output value and the actual value.

[0066] S3-2, establish backpropagation, propagate the output back to the input layer layer by layer through the hidden layer in a certain form, and distribute the error to all units in each layer to obtain the error signal of each unit. This error signal is used as the basis for correcting the weights of each unit.

[0067] S3-3 uses the gradient descent method to update the neural network's own variables. The neural network's own variables are then substituted back into the neural network, along with all the training data, to obtain the updated error function value. The above steps are repeated continuously until the error function is reduced to a set range, thus completing the neural network's learning of the training data.

[0068] S4: Combining expert practical experience, a fault classification and health identification model is constructed. The fault status information extracted based on the feature parameter input after dimensionality reduction decomposition is compared with the feature values ​​of the health identification model with a wide threshold, and the evaluation level is output to realize the assessment of the operational health status of the central gate.

[0069] An online health status assessment system for a central gate under a throttling heating boundary using the above method includes a data acquisition module, an array matrix decomposition module, a boundary sample training module, and an operational status assessment module.

[0070] The data acquisition module is used to collect characteristic parameters under throttling conditions, and sequentially extract multi-dimensional fault state information across categories from the individual characteristic parameters to obtain time series data.

[0071] The array matrix decomposition module is used to perform array matrix dimensionality reduction decomposition on the acquired time series data;

[0072] The boundary sample training module is used to perform batch and periodic training on the extreme boundary data in the time series data after dimensionality reduction decomposition, and to complete the learning of the artificial neural network on the training data to obtain the fault state information after dimensionality reduction decomposition.

[0073] The operational status assessment module is used to combine expert practical experience to construct a fault classification and health identification model. It compares the extracted fault status information after dimensionality reduction decomposition based on feature parameter input with the feature values ​​of the health identification model with a wide threshold, outputs the evaluation level, and realizes the assessment of the operational health status of the central gate.

[0074] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still make modifications or equivalent substitutions to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention are within the protection scope of the claims of the present invention pending approval.

Claims

1. An online method for judging the health state of a throttling heat supply boundary lower intermediate door operation, characterized in that, Comprise the following steps: S1: Collecting characteristic parameters in throttling state, the characteristic parameters include the differential pressure value before and after the middle gate, the valve rod vibration value of the middle gate, the disc stress value of the middle gate and the oil pressure value of the high pressure EH oil supply system; sequentially extracting cross-gate multi-dimensional fault state information for single characteristic parameters, and obtaining time series data; The specific steps of extracting cross-gate multi-dimensional fault state information for single characteristic parameters are as follows: S1-1: Establishing a mathematical model of single characteristic parameters, inputting characteristic parameters, and extracting characteristic frequencies of distributed and local faults; S1-2: Performing time domain and frequency domain statistical analysis; S1-3: Based on multi-wavelet analysis method, using kurtosis and entropy to select optimal multi-wavelet, extracting pulse index and margin index statistical characteristics of waveforms; S2: Array matrix dimensionality reduction decomposition is performed on the obtained time series data, and the specific steps are as follows: S2-1: The similarity of time series data is measured by calculating the distance between data points; S2-2: The time series data is clustered and distinguished by relying on a clustering objective function, the distinguishability between different classes of data is obtained, and the best clustering division is obtained by optimizing the objective function; S2-3: For the processed high-dimensional time series data clustering, the specific characteristic subspace of each class of data is found, and the relevant characteristic dimensions are weighted; S2-4: According to the Pareto dominance relationship, a multi-objective optimization problem is solved, multiple objective functions are sorted according to the relative importance, and the objective function is solved according to the weight coefficient in turn, and the multi-objective optimization problem is converted into a series of single-objective optimization problems; S3: Batch training and periodic training are performed on the limit boundary data in the time series data after dimensionality reduction decomposition by combining artificial neural network, and the learning of the training data by artificial neural network is completed, and the fault state information after dimensionality reduction decomposition is obtained; S4: A fault classification and health identification model is constructed according to the actual experience of experts, the characteristic values of the fault state information after dimensionality reduction decomposition based on the input of characteristic parameters are compared with the health identification model, and an evaluation level is output, so as to realize the evaluation of the running health state of the middle gate.

2. The method according to claim 1, characterized in that, The time series data obtained is subjected to array matrix dimensionality reduction decomposition based on a multi-objective evolutionary clustering algorithm.

3. The method according to claim 1, characterized in that, The distance between data points is defined by the Euclidean distance criterion.

4. The method according to claim 1, characterized in that, The method of weighting the relevant characteristic dimensions is to reduce the weight coefficient of non-attention characteristics and increase the weight coefficient of attention characteristics, wherein the weight coefficient formula is: 。 5. The method according to claim 1, characterized in that, The multi-objective optimization problem is converted into a series of single-objective optimization problems, and the problem definition is as follows: wherein denotes a decision variable, is an n-dimensional decision space, is a set of objective functions that constitutes an m-dimensional objective space and maps points in the decision space to the objective space.

6. The method according to claim 1, characterized in that, The specific steps of batch training and periodic training on the limit boundary data in the time series data after dimensionality reduction decomposition are as follows: S3-1, establish forward propagation, calculate the final output value and the loss value between the output value and the actual value according to the input sample data, the given initial weight value and the value of the bias term; S3-2, establish back propagation, the output is transmitted to the input layer through the hidden layer in a certain form, and the error is allocated to all units in each layer to obtain the error signal of each layer unit, which is used as the basis for correcting the weight of each unit. S3-3 uses gradient descent method to update the neural network's own variables, re-enters the variables into the neural network, adds all the training data, and obtains the updated error function value. The above steps are repeated until the error function is reduced to the set range, and the neural network learning of the training data is completed.

7. An on-line system for judging the health state of a throttling heat supply boundary lower intermediate door according to the method of any one of claims 1-6, characterized in that, The data acquisition module, the array matrix decomposition module, the boundary sample training module, and the running state evaluation module are included. The data acquisition module is used for collecting characteristic parameters under throttling state, extracting cross-door multi-dimensional fault state information for single characteristic parameters in sequence, and obtaining time series data. The array matrix decomposition module is used for array matrix dimensionality reduction decomposition of the obtained time series data. The boundary sample training module is used for batch training and periodic training of limit boundary data in the dimensionality reduction decomposed time series data in combination with an artificial neural network, and learning of the training data by the artificial neural network, to obtain the dimensionality reduction decomposed fault state information. The running state evaluation module is used for constructing a fault classification and health identification model in combination with expert actual experience, comparing the extracted dimensionality reduction decomposed fault state information based on characteristic parameter input with characteristic values of the health identification model in a wide threshold, and outputting an evaluation grade to realize evaluation of the running health state of the middle joint door.

Images