A method and device for controlling combustion pulsation in heavy-duty gas turbines

Abstract

This invention discloses a method and apparatus for combustion pulsation control of a heavy-duty gas turbine. The method includes acquiring test data from the combustion chamber of the heavy-duty gas turbine and processing the test data to obtain first sample data of various operating parameters. Based on the first sample data, a second sample data of key features is determined using a random forest model. Based on the second sample data, a first decision tree is determined using a decision tree model. A pulsation window region is determined based on the first decision tree, and combustion pulsation control of the heavy-duty gas turbine is performed based on the pulsation window region. This invention reduces reliance on human experience by controlling combustion pulsation in a heavy-duty gas turbine through a pulsation window region, improving the reliability and practicality of combustion pulsation identification, providing clear support for monitoring and regulating the combustion state of the gas turbine, and thus improving the efficiency and accuracy of combustion adjustment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of combustion engine control, in particular to a heavy-duty gas turbine combustion pulsation control method and device. BACKGROUND

[0002] Currently, the combustion chamber of the heavy-duty gas turbine adopts dry lean premixed combustion technology to achieve low emission targets. Among them, the dry lean premixed combustion technology reduces the temperature of the combustion zone by pre-mixing fuel and excess air, thereby effectively suppressing the generation of thermal nitrogen oxides. However, in order to achieve this goal, the fuel equivalence ratio needs to be controlled in the range close to the lean extinction limit, making the combustion process sensitive and unstable. Based on this, the combustion heat release rate is easily coupled with the acoustic mode and flow field structure in the combustion chamber.

[0003] Specifically, when the periodic fluctuations of the heat release rate in the combustion process resonate with the inherent acoustic mode of the combustion chamber, a thermoacoustic oscillation phenomenon occurs, which can be manifested as pressure pulsation (combustion pulsation) and flickering of flame shape. If the amplitude of the combustion pulsation exceeds the safety threshold, it may cause thermal fatigue damage to high-temperature components, flame extinction, or even destruction of the combustion chamber structure, thereby affecting the safety and economy of the combustion engine. For the above-mentioned combustion pulsation problem, the existing technology usually adopts a control strategy based on the combination of theoretical analysis and engineering experience. Therefore, it is necessary to control the combustion pulsation of the heavy-duty gas turbine to avoid the amplitude of the combustion pulsation exceeding the safety threshold.

[0004] In related technologies, relying on artificial experience, conventional operating parameters such as fuel flow, air flow, and outlet temperature are selected as key features and analysis basis to represent the combustion state, and combined with pre-set pulsation amplitude thresholds at different frequencies, it is judged whether the current combustion state is within the safe allowable range. When the pulsation amplitude exceeds the safety threshold, the test personnel determine the adjustment direction and amplitude of the operating parameters according to the calculation results of the theoretical model and accumulated engineering experience, and through repeated trials, the combustion state is returned to the safe interval. Among them, the above technical solution is characterized by experience-oriented feature selection and combustion adjustment method, which has been applied in the test development and actual operation of the gas turbine.

[0005] However, the theoretical calculation method used in the above technical solution is low in efficiency and sensitive to parameter changes, and needs to be recalculated every time the adjustment is made, and there is often a large error between the actual test and the calculation result, so the reliability is low; the regulation process relies on the parameter adjustment range set by experience, and lacks clear and quantitative pulsation trigger boundary judgment basis. Therefore, the test personnel lack clear and intuitive decision support in the adjustment process, which affects the efficiency and accuracy of the combustion adjustment. SUMMARY

[0006] The application provides a heavy gas turbine combustion pulsation control method and device to solve the technical problems of low reliability, low efficiency and precision of combustion adjustment in the prior art.

[0007] To this end, the application provides a heavy gas turbine combustion pulsation control method, which can determine a first decision tree by using a random forest model and a decision tree model, and perform combustion pulsation control on the heavy gas turbine based on a pulsation window area determined by the first decision tree, so that the random forest model and the decision tree model can be used to perform combined analysis on features, the pulsation window area can be obtained by quantitative division of a pulsation trigger boundary, the combustion pulsation control can be performed on the heavy gas turbine through the pulsation window area, the dependence on artificial experience is reduced, the reliability and practicability of combustion pulsation identification are improved, clear support is provided for monitoring and regulation of the combustion state of the gas turbine, and the efficiency and precision of combustion adjustment are improved.

[0008] Another object of the application is to provide a heavy gas turbine combustion pulsation control device.

[0009] To achieve the above object, in one aspect, the application provides a heavy gas turbine combustion pulsation control method, comprising:

[0010] Obtaining heavy gas turbine combustion chamber test data, and performing data processing on the combustion chamber test data to obtain first sample data of each working condition parameter;

[0011] Based on the first sample data, determining second sample data of key features by using a random forest model;

[0012] Based on the second sample data, determining a first decision tree by using a decision tree model;

[0013] Determining a pulsation window area based on the first decision tree, and performing combustion pulsation control on the heavy gas turbine based on the pulsation window area.

[0014] The heavy gas turbine combustion pulsation control method of the application can further have the following additional technical features:

[0015] In the embodiment of the application, the data processing on the combustion chamber test data to obtain first sample data of each working condition parameter comprises:

[0016] Based on the data acquisition time sequence in the combustion chamber test data, determining a feature matrix of each working condition parameter;

[0017] Performing standardization processing on the feature matrix of each working condition parameter to obtain a standardized feature matrix;

[0018] The pulsation type labels are marked to the pulsation numerical values of each time sequence multi-band in the combustion chamber test data based on each pulsation threshold criterion, to obtain pulsation data;

[0019] The standardized feature matrix and the pulsation data are determined as first sample data of each working condition parameter.

[0020] In the embodiment of the present application, the second sample data of the key feature is determined based on the first sample data through a random forest, comprising:

[0021] A training subset is generated from the first sample data through Bootstrap;

[0022] Random forest training is performed based on the training subset, to obtain a corresponding second decision tree;

[0023] The contribution degrees of each working condition parameter in the second decision tree are determined;

[0024] The feature importance degrees of each working condition parameter are determined based on the contribution degrees of each working condition parameter in the second decision tree;

[0025] The key feature is determined based on the feature importance degrees of each working condition parameter, and the first sample data corresponding to the key feature is determined as the second sample data of the key feature.

[0026] In the embodiment of the present application, the random forest training based on the training subset is performed to obtain a corresponding second decision tree, comprising:

[0027] A feature subset is randomly selected from the working condition parameters of the training subset;

[0028] An optimal split threshold of each working condition parameter in the feature subset is determined;

[0029] Node splitting is performed by using the optimal split threshold of each working condition parameter, to obtain a third decision tree;

[0030] The parameters of the third decision tree are adjusted by using Bayesian optimization, to obtain a corresponding second decision tree.

[0031] In the embodiment of the present application, the optimal split threshold of each working condition parameter in the feature subset is determined, comprising:

[0032] The optimal split threshold of each working condition parameter in the feature subset is found based on a Gini index.

[0033] In the embodiment of the present application, the first decision tree is determined based on the second sample data through a decision tree model, comprising:

[0034] The weights of different pulsation types in the second sample data are determined;

[0035] calculating a weighted information entropy of the node based on the number and weight of each pulsation type sample in the node;

[0036] determining an optimal split rule of the node based on the weighted information entropy of the node;

[0037] repeating the calculating the weighted information entropy of the node and the determining the optimal split rule of the node to recursively generate a decision tree until a preset stop condition is met, to obtain a first decision tree, wherein the stop condition comprises that a tree depth reaches a first preset value.

[0038] In the embodiments of the present application, the determining the optimal split rule of the node based on the weighted information entropy of the node comprises:

[0039] determining each candidate working condition parameter and a corresponding preset threshold value;

[0040] dividing the node samples into left and right two child nodes based on the candidate working condition parameter and the corresponding preset threshold value, and calculating a weighted information gain after the division based on the weighted information entropy of the node;

[0041] combining the candidate working condition parameter and the preset threshold value corresponding to the maximum weighted information gain to determine as the optimal split rule of the node.

[0042] In the embodiments of the present application, the determining the pulsation window region based on the first decision tree comprises:

[0043] projecting the split rules in the first decision tree into a coordinate system to obtain a plurality of polygon regions;

[0044] determining each of the polygon regions as a visual pulsation trigger window of different pulsation types.

[0045] In the embodiments of the present application, the controlling the combustion pulsation of the heavy-duty gas turbine based on the pulsation window region comprises:

[0046] projecting a working condition parameter of a key feature in the heavy-duty gas turbine into the coordinate system to obtain a running position of the heavy-duty gas turbine in the coordinate system;

[0047] if a distance between the running position and the pulsation window is less than a second preset value, adjusting a key parameter corresponding to the key feature based on the corresponding split rule.

[0048] To achieve the above purpose, another aspect of the present application provides a heavy-duty gas turbine combustion pulsation control device, which comprises:

[0049] The data processing module is configured to acquire heavy-duty gas turbine combustion chamber test data, and perform data processing on the combustion chamber test data to obtain first sample data of each working condition parameter.

[0050] The first determination module is configured to determine second sample data of a key feature based on the first sample data by using a random forest model.

[0051] The second determination module is configured to determine a first decision tree based on the second sample data by using a decision tree model.

[0052] The control module is configured to determine a pulsation window region based on the first decision tree, and perform combustion pulsation control on the heavy-duty gas turbine based on the pulsation window region.

[0053] In the embodiment of the present application, the data processing module is specifically configured to:

[0054] Determine a feature matrix of each working condition parameter based on a data acquisition time sequence in the combustion chamber test data.

[0055] Perform standardization processing on the feature matrix of each working condition parameter to obtain a standardized feature matrix.

[0056] Label a pulsation type tag based on a pulsation threshold standard for a pulsation value of each time sequence multi-band in the combustion chamber test data to obtain pulsation data.

[0057] Determine the standardized feature matrix and the pulsation data as the first sample data of each working condition parameter.

[0058] In the embodiment of the present application, the first determination module is specifically configured to:

[0059] Generate a training subset from the first sample data by using Bootstrap;

[0060] Perform random forest training based on the training subset to obtain a corresponding second decision tree;

[0061] Determine a contribution degree of each working condition parameter in the second decision tree;

[0062] Determine a feature importance of each working condition parameter based on the contribution degree of each working condition parameter in the second decision tree.

[0063] Determine a key feature based on the feature importance of each working condition parameter, and determine the first sample data corresponding to the key feature as second sample data of the key feature.

[0064] In the embodiment of the present application, the first determination module is further configured to:

[0065] randomly selecting a feature subset from the working condition parameters of the training subset;

[0066] determining an optimal split threshold of each working condition parameter in the feature subset;

[0067] splitting the nodes by using the optimal split threshold of each working condition parameter to obtain a third decision tree;

[0068] adjusting parameters of the third decision tree by using Bayesian optimization to obtain a corresponding second decision tree.

[0069] In the embodiment of the present application, the first determining module is further configured to:

[0070] determining the optimal split threshold of each working condition parameter in the feature subset based on a Gini index.

[0071] In the embodiment of the present application, the second determining module is specifically configured to:

[0072] determining weights of different pulsation types in the second sample data;

[0073] calculating a weighted information entropy of the node based on the number and weights of the pulsation type samples in the node;

[0074] determining an optimal split rule of the node based on the weighted information entropy of the node;

[0075] recursively generating a decision tree by repeating the calculation of the weighted information entropy of the node and the determination of the optimal split rule of the node until a preset stop condition is met to obtain a first decision tree, wherein the stop condition includes that a tree depth reaches a first preset value.

[0076] In the embodiment of the present application, the second determining module is further configured to:

[0077] determining each candidate working condition parameter and a corresponding preset threshold;

[0078] dividing the node samples into left and right two child nodes based on the candidate working condition parameter and the corresponding preset threshold, and calculating a weighted information gain after the division based on the weighted information entropy of the node;

[0079] combining the candidate working condition parameter and the preset threshold corresponding to the maximum weighted information gain to determine as the optimal split rule of the node.

[0080] In the embodiment of the present application, the control module is specifically configured to:

[0081] projecting the split rules in the first decision tree into a coordinate system to obtain a plurality of polygonal regions;

[0082] Each of the polygonal regions is determined as a visualized pulsation trigger window of different pulsation types.

[0083] In the embodiments of the present application, the control module is further configured to:

[0084] Project the operating condition parameters of the key features in the heavy-duty gas turbine into the coordinate system to obtain an operating position of the heavy-duty gas turbine in the coordinate system.

[0085] If the distance between the operating position and the pulsation window is less than a second preset value, adjust the key parameters corresponding to the key features based on the corresponding splitting rule.

[0086] The heavy-duty gas turbine combustion pulsation control method and device provided in the embodiments of the present application include obtaining heavy-duty gas turbine combustion chamber test data, and performing data processing on the combustion chamber test data to obtain first sample data of each operating condition parameter; determining second sample data of key features based on the first sample data through a random forest model; determining a first decision tree based on the second sample data through a decision tree model; determining a pulsation window region based on the first decision tree, and performing combustion pulsation control on the heavy-duty gas turbine based on the pulsation window region. Thus, the present application uses the random forest model and the decision tree model to perform combined analysis on the features, realizes quantitative division of the pulsation trigger boundary to obtain the pulsation window region, so that the combustion pulsation control can be performed on the heavy-duty gas turbine through the pulsation window region, the dependence on artificial experience is reduced, the reliability and practicability of combustion pulsation identification are improved, clear support is provided for monitoring and regulation of the combustion state of the gas turbine, and the efficiency and accuracy of combustion adjustment are improved.

[0087] Additional aspects and advantages of the present application will be made apparent from the following description, which, taken in conjunction with the accompanying drawings, which are shown by way of illustration. BRIEF DESCRIPTION OF DRAWINGS

[0088] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which:

[0089] Figure 1 is a flowchart of a heavy-duty gas turbine combustion pulsation control method according to an embodiment of the present application;

[0090] Figure 2 is a ranking result diagram of feature importance in a medium-low frequency pulsation according to an embodiment of the present application;

[0091] Figure 3 is a ranking result diagram of feature importance in a high frequency pulsation according to an embodiment of the present application;

[0092] Figure 4is a schematic diagram of a first decision tree according to an embodiment of the present application;

[0093] Figure 5 is a schematic diagram of a pulsation window according to an embodiment of the present application;

[0094] Figure 6 is a structural schematic diagram of a heavy-duty gas turbine combustion pulsation control device according to an embodiment of the present application. DETAILED DESCRIPTION

[0095] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0096] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0097] The heavy-duty gas turbine combustion pulsation control method and device according to the embodiments of the present application will be described below with reference to the accompanying drawings.

[0098] Figure 1 is a flowchart of a heavy-duty gas turbine combustion pulsation control method according to an embodiment of the present application.

[0099] As shown in Figure 1 , the method can include the following steps:

[0100] Step 101, obtaining heavy-duty gas turbine combustion chamber test data, and performing data processing on the combustion chamber test data to obtain first sample data of each operating condition parameter.

[0101] In the embodiments of the present application, the heavy-duty gas turbine combustion chamber test data can include each operating condition parameter and multi-band pulsation value collected in time sequence. In the embodiments of the present application, the above-mentioned operating condition parameters can include but are not limited to: combustion chamber air flow, center nozzle fuel flow and center nozzle value fuel flow.

[0102] In the embodiments of the present application, the method of performing data processing on the combustion chamber test data to obtain first sample data of each operating condition parameter can include the following steps:

[0103] Step 1011, determining the feature matrix of each operating condition parameter based on the data collection time sequence in the combustion chamber test data.

[0104] In the embodiment of the present application, the characteristic matrix of each working condition parameter is determined from the data collection time sequence based on the combustion chamber test data , wherein represents the total number of samples in the data set (i.e. the number of time points of the test samples), represents the total number of working condition parameters collected; the matrix element represents the specific value of the i-th working condition parameter of the j-th sample (the i-th time point).

[0105] In step 1012, the characteristic matrix of each working condition parameter is standardized to obtain a standardized characteristic matrix.

[0106] In the embodiment of the present application, after the characteristic matrix of each working condition parameter is determined by the above steps, the characteristic matrix of each working condition parameter can be standardized to obtain a standardized characteristic matrix, so as to eliminate the dimensional differences and inconsistency of the numerical intervals between the above working condition parameters.

[0107] In the embodiment of the present application, the data preprocessing of each working condition parameter can be uniformly standardized by the zero mean and unit variance method by using the first formula, wherein the first formula is:

[0108]

[0109] , wherein represents the i-th working condition parameter value of the j-th sample after standardization; is the mean value of the i-th working condition parameter in all samples, defined as:

[0110]

[0111] , wherein is the standard deviation of the i-th working condition parameter in all samples, defined as:

[0112]

[0113] In the embodiment of the present application, the above standardization processing makes all working condition parameters mapped to a unified numerical range with zero mean and unit variance, so as to eliminate the influence of the dimensions of each parameter and effectively enhance the stability and accuracy of the random forest and decision tree model for different gas turbine operating conditions in the subsequent training process.

[0114] ​​​​​​​Step 1013: Based on each pulsation threshold standard, label the pulsation values ​​of each time-series multi-frequency band in the combustion chamber test data with pulsation type labels to obtain pulsation data.

[0115] In this embodiment of the invention, the pulsation values ​​of each time-series multi-frequency band are labeled with pulsation type tags at different time points according to preset low, medium, and high frequency pulsation threshold standards. Among them, 0 represents no pulsation, 1 represents low-to-medium frequency pulsation (LFD / MFD), and 2 represents high-frequency pulsation (HFD).

[0116] Furthermore, in this embodiment of the invention, the aforementioned pulsation label uses a three-value encoding: 0 represents no pulsation (NoPulse), 1 represents low-to-medium frequency pulsation (LFD / MFD), and 2 represents high-frequency pulsation (HFD). To allow the model to subsequently focus on the two discrimination tasks of "whether low-to-medium frequency pulsation occurs" and "whether high-frequency pulsation occurs," the three-value label needs to be split into two non-overlapping binary labels so that the model can focus on the two pulsation types respectively.

[0117] ,

[0118] in, This indicates whether low- to mid-frequency pulsations occur at that moment, and This indicates whether high-frequency pulsations occur at that moment.

[0119] Step 1014: The standardized feature matrix and pulsation data are determined as the first sample data for each operating condition parameter.

[0120] In this embodiment of the invention, after obtaining the standardized feature matrix through the above steps, it is possible to perform sampling at each sampling time. All normalized chemical condition parameters are arranged sequentially into eigenvectors. ,in, , , And so on. In this embodiment of the invention, The characteristic vector of combustion chamber airflow at sampling time i The feature vector of fuel flow rate at the center nozzle at sampling time i, The feature vector of the fuel flow rate at the central nozzle at sampling time i.

[0121] Step 102: Based on the first sample data, determine the second sample data of key features using a random forest model.

[0122] In this embodiment of the invention, after obtaining the first sample data through the above steps, the second sample data of key features can be determined based on the first sample data using a random forest model.

[0123] Specifically, in the embodiments of the present application, the method for determining the second sample data of the key features based on the first sample data through the random forest model can include the following steps:

[0124] Step 1021, generating a training subset from the first sample data through Bootstrap.

[0125] Step 1022, performing random forest training based on the training subset to obtain a corresponding second decision tree.

[0126] In the embodiments of the present application, the method for performing random forest training based on the training subset to obtain a corresponding second decision tree can include the following steps:

[0127] Step 1, randomly selecting a feature subset from the working condition parameters of the training subset.

[0128] Step 2, determining the optimal split threshold of each working condition parameter in the feature subset.

[0129] In the embodiments of the present application, in the training process, the samples in the node are continuously divided into left and right child nodes according to whether a working condition parameter exceeds a corresponding threshold value . For example, assuming that the combustion chamber air flow rate is used for splitting, the division rule can be represented as:

[0130]

[0131] wherein, is the left node divided by the combustion chamber air flow rate and the corresponding threshold value , is the right node divided by the combustion chamber air flow rate and the corresponding threshold value .

[0132] In the embodiments of the present application, the method for determining the optimal split threshold of each working condition parameter in the feature subset can include:

[0133] Finding the optimal split threshold of each working condition parameter in the feature subset based on the Gini index.

[0134] In the embodiments of the present application, the method for finding the optimal split threshold of each working condition parameter in the feature subset based on the Gini index can include: determining the corresponding Gini index reduction amount based on the Gini index of each division of the feature subset, and determining the division threshold corresponding to the maximum value of the Gini index reduction amount of each division as the optimal split threshold of each working condition parameter.

[0135] In the embodiments of the present application, the Gini index can be specifically: let be the node , the number of samples with the middle label being middle-frequency pulsation or high-frequency pulsation is denoted as , and the total number of samples of the node is denoted as , then the positive class ratio is , and the Gini index corresponding to the node is , based on which, the Gini indexes of the left and right child nodes are calculated respectively by the above method, that is, and , and the Gini index reduction amount corresponding to the above division is:

[0136]

[0137] In the embodiments of the present application, the Gini index reduction amount describes the improvement range of the "pulsation / non-pulsation" distinguishing ability at the threshold .

[0138] Step 3: Split the node using the optimal split threshold of each working condition parameter to obtain a third decision tree.

[0139] Step 4: Adjust the parameters of the third decision tree using Bayesian optimization to obtain a corresponding second decision tree.

[0140] In the embodiments of the present application, the overall hyperparameters of the third decision tree can be optimized by the Bayesian optimization model, so that the second decision tree has a more optimal decision tree structure and performance.

[0141] Step 1023: Determine the contribution degree of each working condition parameter in the second decision tree.

[0142] In the embodiments of the present application, if the working condition parameter is selected as a split variable multiple times at different levels of the second decision tree, the contribution of the working condition parameter to each node is added one by one to obtain the contribution degree of the working condition parameter in the second decision tree.

[0143] Step 1024: Determine the feature importance of each working condition parameter based on the contribution degree of each working condition parameter in the second decision tree.

[0144] In the embodiments of the present application, after determining the contribution degree of each working condition parameter in the second decision tree by the above steps, the contribution degree of each working condition parameter in the second decision tree can be determined as the feature importance of each working condition parameter .

[0145] Step 1025: Determine the key feature based on the feature importance of each working condition parameter, and determine the second sample data of the key feature as the first sample data corresponding to the key feature.

[0146] In the embodiment of the present application, after the feature importance of each working condition parameter is determined through the above steps, the key features can be determined based on the feature importance of each working condition parameter.

[0147] In the embodiment of the present application, the method for determining the key features based on the feature importance of each working condition parameter can include: ranking the feature importance of all working condition parameters in descending order, determining the top pre-set number of working condition parameters in the obtained ranking result as the key working condition parameters of the corresponding pulsation type, and merging the key working condition parameters of different pulsation types to obtain the key features.

[0148] In the embodiment of the present application, Figure 2 A ranking result diagram of feature importance in a low-frequency pulsation in the embodiment of the present application is shown in FIG. 2. Figure 3 A ranking result diagram of feature importance in a high-frequency pulsation in the embodiment of the present application is shown in FIG. 3. Figure 2 And Figure 3 The features in FIGS. 2 and 3 can be inconsistent.

[0149] In addition, in the embodiment of the present application, the pre-set number can be set as needed, such as 5.

[0150] In the embodiment of the present application, after the key features are determined through the above steps, the first sample data corresponding to the key features can be determined as the second sample data of the key features.

[0151] In step 103, a first decision tree is determined based on the second sample data through a decision tree model.

[0152] In the embodiment of the present application, after the second sample data is obtained through the above steps, a first decision tree can be determined based on the second sample data through a decision tree model.

[0153] In the embodiment of the present application, the method for determining the first decision tree based on the second sample data through the decision tree model can include the following steps:

[0154] In step 1031, the weights of different pulsation types in the second sample data are determined.

[0155] In the embodiment of the present application, the weights of different pulsation types in the second sample data can be determined as needed, so that the decision data pays more attention to the scarce pulsation samples in the sample imbalance scenario. For example, different weights are set for no pulsation, low-frequency pulsation and high-frequency pulsation: the weight of no pulsation is 1, and the weights of low-frequency pulsation and high-frequency pulsation are both 2.

[0156] In step 1032, the weighted information entropy of the node is calculated based on the number and weight of each pulsation type sample in the node.

[0157] In the embodiments of the present application, the weighted information entropy of the node can be calculated based on the number and weight of each pulsation type sample in the node. For example, in the embodiments of the present application, it is assumed that the number of samples corresponding to no pulsation, low-frequency pulsation and high-frequency pulsation in the node is

[0158]

[0159]

[0160]

[0161] In step 1033, the optimal split rule of the node is determined based on the weighted information entropy of the node.

[0162] In the embodiments of the present application, after the weighted information entropy of the node is determined through the above steps, the optimal split rule of the node can be determined based on the weighted information entropy of the node.

[0163] In the embodiments of the present application, the method for determining the optimal split rule of the node based on the weighted information entropy of the node can include the following steps:

[0164] In step 10331, each candidate working condition parameter and the corresponding preset threshold value are determined.

[0165] In the embodiments of the present application, each candidate working condition parameter and the corresponding preset threshold value can be determined as needed. For example, in the embodiments of the present application, it is assumed that the candidate working condition parameter is the combustion chamber air flow

[0166] In the embodiments of the present application, the candidate working condition parameter can be a working condition parameter in the key feature.

[0167] In step 10332, the node samples are divided into left and right two child nodes based on the candidate working condition parameter and the corresponding preset threshold value, and the weighted information gain after the division is calculated based on the weighted information entropy of the node.

[0168] In the embodiments of the present application, after the candidate working condition parameter and the corresponding preset threshold value are obtained through the above steps, the node samples can be divided into left and right two child nodes based on the candidate working condition parameter and the corresponding preset threshold value, and the weighted information gain after the division is calculated based on the weighted information entropy of the node.​​​​​​​​​​​

[0169] In the embodiment of the present application, the method for calculating the weighted information gain after division based on the weighted information entropy of the node can comprise: calculating the weighted information gain after division based on the weighted information entropy of the node by a second formula, wherein the second formula is:

[0170]

[0171] wherein, is the sample quantity of the left node after division and is the sample quantity of the right node after division.

[0172] In the embodiment of the present application, the preset threshold value of the candidate working condition parameter is represented as "zero mean-unit variance" in the training stage, and needs to be calculated back to the real dimension when visualized and applied.

[0173] Step 10333, combining the candidate working condition parameter corresponding to the maximum weighted information gain and the preset threshold value to determine the optimal split rule of the node.

[0174] Step 1034, repeating the above steps 1032-1033 to recursively generate a decision tree until a preset stop condition is met, to obtain a first decision tree, wherein the stop condition comprises that the tree depth reaches a first preset value.

[0175] In the embodiment of the present application, the first preset value can be set as needed, such as 3, that is, after training is completed, the algorithm outputs a first decision tree with a depth limited to 3, that is, the longest path from the root node to any leaf node does not exceed 3 times of splitting.

[0176] In the embodiment of the present application, after obtaining the first decision tree through the above steps, the sample mean and standard deviation of the working condition parameter can be obtained as and and the normalized threshold value in the first decision tree to obtain the working condition boundary corresponding to the working condition parameter as:

[0177]

[0178] In the embodiment of the present application, after obtaining the working condition parameter, the division threshold value and the working condition boundary corresponding to each node through the above steps, the working condition parameter, the division threshold value and the working condition boundary corresponding to each node can be added to each node in the first decision tree, so that the threshold condition in the first decision tree can be corresponded to the working condition boundary, thereby the decision rule output by the first decision tree can be more intuitively understood and applied.

[0179] In the embodiment of the present application, the first decision tree uses a few key operating parameters such as a specific equivalence ratio, temperatures at different positions, etc. Each split normalizes the division threshold to the original dimension and adds it to the node box of the corresponding node. In addition, different colors are used to represent the classification results of each node (non-pulsation, low-frequency pulsation, high-frequency pulsation) during visualization. The purer the color, the higher the proportion of the same type of samples in the leaf node. Figure 4 A schematic diagram of a first decision tree proposed in the embodiment of the present application is shown in FIG. 2. Figure 4 The visualization classification results of the first decision tree are shown in FIG. 3. The color of each leaf node represents the pulsation classification results of the time sample points in this part, and the key node is marked with the boundary condition of the normalized operating parameter for reference of combustion pulsation analysis. Figure 4

[0180] In the embodiment of the present application, the performance of the above-mentioned random forest in the classification of combustion pulsation can also be measured on the basis of the node classification of the above-mentioned first decision tree, including the low-frequency accuracy, the high-frequency accuracy, and the overall pulsation accuracy including non-pulsation samples. For example, in the embodiment of the present application, it is assumed that represents the index set of all pulsation samples, is the decision result of the first decision tree on the i th sample, at this time, the overall pulsation accuracy can be determined by the third formula, wherein the third formula is:

[0181]

[0182] Step 104, determining a pulsation window region based on the first decision tree, and controlling the combustion pulsation of the heavy-duty gas turbine based on the pulsation window region.

[0183] In the embodiment of the present application, after obtaining the first decision tree through the above steps, the pulsation window region can be determined based on the first decision tree, and the combustion pulsation of the heavy-duty gas turbine can be controlled based on the pulsation window region.

[0184] In the embodiment of the present application, the method for determining the pulsation window region based on the first decision tree and controlling the combustion pulsation of the heavy-duty gas turbine based on the pulsation window region can include the following steps:

[0185] Step 1041, projecting the split rule in the first decision tree into a coordinate system to obtain a plurality of polygon regions.

[0186] Step 1042, determining each polygon region as a visual pulsation trigger window of different pulsation types.

[0187] ​​In this embodiment of the invention, the method of projecting the splitting rules in the first decision tree onto a coordinate system to obtain multiple polygonal regions may include: based on the splitting rules in the first decision tree and the boundary conditions of the operating parameters, projecting the first decision tree onto a two-dimensional feature plane to obtain corresponding multiple polygonal regions.

[0188] In this embodiment of the invention, a scatter plot can also be drawn on a two-dimensional feature coordinate system to display the sample distribution of no-pulsation, low-to-medium frequency pulsation, and high-frequency pulsation, so as to clearly show the pulsation window region divided by the decision tree, such as... Figure 5 As shown, after the first decision tree is split and projected onto the two-dimensional feature plane, each split corresponds to a horizontal or vertical straight line, thereby dividing the plane into several regions to show the distribution range of low-frequency or high-frequency pulsations in the actual data.

[0189] For example, when the splitting condition of a node in the first decision tree is or When this happens, a vertical or horizontal dividing line will be drawn, thus forming a pulse trigger window. and Features can be assigned to different operating condition parameters. and These are the corresponding partitioning thresholds.

[0190] In this embodiment of the invention, the method for combustion pulsation control of a heavy-duty gas turbine based on a pulsating window region may include the following steps:

[0191] Step a: Project the operating parameters of the key features of the heavy-duty gas turbine into the coordinate system to obtain the operating position of the heavy-duty gas turbine in the coordinate system.

[0192] Step b: If the distance between the running position and the pulsation window is less than the second preset value, then adjust the key parameters corresponding to the key features based on the corresponding splitting rules.

[0193] In this embodiment of the invention, the operating parameters of the current heavy-duty gas turbine are used as... For example, among which, This refers to the airflow rate in the combustion chamber. The center nozzle fuel flow is projected into the same coordinate system to obtain the operation position of the heavy-duty gas turbine in the coordinate system. Based on this, if the distance between the operation position and the pulsation window is greater than a third preset value, it indicates that the operation position is in the safe area at this time, which means that the current air flow-center nozzle fuel flow combination of the combustion chamber is far away from the thermal-acoustic oscillation boundary, and the test operating condition can be continuously expanded. If the distance between the operation position and the pulsation window is less than a second preset value, it indicates that the operation position is close to the low-frequency or high-frequency pulsation window at this time, which means that the pulsation trigger is imminent, and the operating condition parameters need to be immediately corrected according to the threshold indicated by the leaf node. The second preset value and the third preset value can be set as needed.

[0194] In the embodiment of the present application, the method for adjusting the key parameters corresponding to the key features can include fine-tuning along the vertical or horizontal direction of the first decision tree split line to make the operating point retreat from the edge of the pulsation window to the safe area. For example, the one-time parameter adjustment vector is used to accurately execute:

[0195]

[0196] In the embodiment of the present application, the complex thermal-acoustic coupling can be formulated into intuitive safe area / dangerous area discrimination through the visual pulsation window, which greatly reduces the dependence on experience and improves the efficiency of combustion adjustment and the safety of long-term operation of the gas turbine.

[0197] The heavy-duty gas turbine combustion pulsation control method provided in the embodiment of the present application includes obtaining test data of a combustion chamber of a heavy-duty gas turbine, and performing data processing on the test data of the combustion chamber to obtain first sample data of each operating condition parameter. Based on the first sample data, second sample data of key features is determined through a random forest model. Based on the second sample data, a first decision tree is determined through a decision tree model. A pulsation window area is determined based on the first decision tree, and the heavy-duty gas turbine is controlled for combustion pulsation based on the pulsation window area. Thus, the random forest model and the decision tree model are used to analyze the features in combination to quantitatively divide the pulsation trigger boundary to obtain the pulsation window area, so that the heavy-duty gas turbine can be controlled for combustion pulsation through the pulsation window area, the dependence on artificial experience is reduced, the reliability and practicality of combustion pulsation identification are improved, clear support is provided for monitoring and regulation of the combustion state of the gas turbine, and the efficiency and accuracy of combustion adjustment are improved.

[0198] To implement the above-mentioned embodiments, as Figure 6 shown, the present embodiment further provides a heavy-duty gas turbine combustion pulsation control device, which can include:

[0199] The data processing module 601 is configured to obtain test data of a combustion chamber of a heavy-duty gas turbine, and perform data processing on the test data of the combustion chamber to obtain first sample data of each operating condition parameter.

[0200] The first determining module 602 is configured to determine second sample data of a key feature based on first sample data by using a random forest model.

[0201] The second determining module 603 is configured to determine a first decision tree based on the second sample data by using a decision tree model.

[0202] The control module 604 is configured to determine a pulsation window region based on the first decision tree, and perform combustion pulsation control on the heavy-duty gas turbine based on the pulsation window region.

[0203] In the embodiment of the present application, the data processing module 601 is specifically configured to:

[0204] determine a feature matrix of each working condition parameter based on a time sequence of data acquisition in the combustion chamber test data;

[0205] perform standardization processing on the feature matrix of each working condition parameter to obtain a standardized feature matrix;

[0206] label a pulsation type tag based on each time sequence multi-band pulsation value in the combustion chamber test data to obtain pulsation data;

[0207] determine the standardized feature matrix and the pulsation data as the first sample data of each working condition parameter.

[0208] In the embodiment of the present application, the first determining module 602 is specifically configured to:

[0209] generate a training subset from the first sample data by using Bootstrap;

[0210] perform random forest training based on the training subset to obtain a corresponding second decision tree;

[0211] determine a contribution degree of each working condition parameter in the second decision tree;

[0212] determine a feature importance of each working condition parameter based on the contribution degree of each working condition parameter in the second decision tree;

[0213] determine a key feature based on the feature importance of each working condition parameter, and determine the first sample data corresponding to the key feature as second sample data of the key feature.

[0214] In the embodiment of the present application, the first determining module 602 is further configured to:

[0215] randomly select a feature subset from the working condition parameters of the training subset;

[0216] determine an optimal split threshold of each working condition parameter in the feature subset;

[0217] Split the nodes by using the optimal split threshold of each working condition parameter to obtain a third decision tree;

[0218] Adjust the parameters of the third decision tree by using Bayesian optimization to obtain a corresponding second decision tree.

[0219] In the embodiment of the application, the first determination module 602 is further configured to:

[0220] Find the optimal split threshold of each working condition parameter in the feature subset based on the Gini index.

[0221] In the embodiment of the application, the second determination module 603 is specifically configured to:

[0222] Determine the weights of different pulsation types in the second sample data;

[0223] Calculate the weighted information entropy of the node based on the number and weights of the pulsation type samples in the node;

[0224] Determine the optimal split rule of the node based on the weighted information entropy of the node;

[0225] Recursively generate a decision tree by repeatedly calculating the weighted information entropy of the node and determining the optimal split rule of the node until a preset stop condition is met to obtain a first decision tree, wherein the stop condition includes that the tree depth reaches a first preset value.

[0226] In the embodiment of the application, the second determination module 603 is further configured to:

[0227] Determine each candidate working condition parameter and a corresponding preset threshold;

[0228] Divide the node samples into two child nodes based on the candidate working condition parameter and the corresponding preset threshold, and calculate the weighted information gain after the division based on the weighted information entropy of the node;

[0229] Combine the candidate working condition parameter and the preset threshold corresponding to the maximum weighted information gain to determine the optimal split rule of the node.

[0230] In the embodiment of the application, the control module 604 is specifically configured to:

[0231] Project the split rule in the first decision tree into a coordinate system to obtain a plurality of polygonal regions;

[0232] Determine each polygonal region as a visual pulsation trigger window of a different pulsation type.

[0233] In the embodiment of the application, the control module 604 is further configured to:

[0234] The working condition parameters of the key features in the heavy-duty gas turbine are projected into the coordinate system to obtain a running position of the heavy-duty gas turbine in the coordinate system.

[0235] If the distance between the running position and the pulsation window is less than the second preset value, the key parameters corresponding to the key features are adjusted based on the corresponding splitting rule.

[0236] In the heavy-duty gas turbine combustion pulsation control device, the first sample data of each working condition parameter is obtained by acquiring the test data of the combustion chamber of the heavy-duty gas turbine and processing the test data of the combustion chamber; the second sample data of the key features is determined based on the first sample data through a random forest model; the first decision tree is determined based on the second sample data through a decision tree model; the pulsation window area is determined based on the first decision tree, and the heavy-duty gas turbine is controlled for combustion pulsation based on the pulsation window area. Thus, the random forest model and the decision tree model are used to analyze the features in combination to quantitatively divide the pulsation trigger boundary to obtain the pulsation window area, so that the heavy-duty gas turbine can be controlled for combustion pulsation through the pulsation window area, the dependence on artificial experience is reduced, the reliability and practicality of combustion pulsation identification are improved, clear support is provided for monitoring and regulation of the combustion state of the gas turbine, and the efficiency and accuracy of combustion adjustment are improved.

[0237] It should be understood that the various forms of flow shown above can be used to reorder, add or delete steps. For example, the steps described in the present disclosure can be executed in parallel, sequentially or in a different order, as long as the desired results of the technical solutions of the present disclosure can be achieved, which is not limited herein.

[0238] The above specific embodiments do not constitute a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements and improvements within the spirit and principles of the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A method of heavy-duty gas turbine combustion pulsation control, characterized by, The method comprises: obtaining heavy gas turbine combustor test data, and performing data processing on the combustor test data to obtain first sample data of each operating condition parameter; based on the first sample data, determining second sample data of key features through a random forest model; based on the second sample data, determining a first decision tree through a decision tree model; based on the first decision tree, determining a pulsation window region, and based on the pulsation window region, performing combustion pulsation control on the heavy gas turbine.

2. The method of claim 1, wherein, The data processing on the combustor test data to obtain first sample data of each operating condition parameter comprises: based on the data collection time sequence in the combustor test data, determining a feature matrix of each operating condition parameter; performing standardization processing on the feature matrix of each operating condition parameter to obtain a standardized feature matrix; based on each pulsation threshold standard, marking the pulsation type label of the pulsation value of each time sequence multi-band in the combustor test data to obtain pulsation data; determining the standardized feature matrix and the pulsation data as the first sample data of each operating condition parameter.

3. The method of claim 1, wherein, The determination of second sample data of key features based on the first sample data through a random forest comprises: generating a training subset from the first sample data through Bootstrap; based on the training subset, performing random forest training to obtain a corresponding second decision tree; determining the contribution degree of each operating condition parameter in the second decision tree; based on the contribution degree of each operating condition parameter in the second decision tree, determining the feature importance of each operating condition parameter; based on the feature importance of each operating condition parameter, determining key features, and determining the first sample data corresponding to the key features as the second sample data of the key features.

4. The method of claim 3, wherein, The random forest training based on the training subset to obtain a corresponding second decision tree comprises: randomly selecting a feature subset from the operating condition parameters of the training subset; determining the optimal split threshold of each operating condition parameter in the feature subset; performing node splitting using the optimal split threshold of each operating condition parameter to obtain a third decision tree; adjusting the parameters of the third decision tree using Bayesian optimization to obtain a corresponding second decision tree.

5. The method of claim 4, wherein, The determination of the optimal split threshold of each operating condition parameter in the feature subset comprises: based on the Gini index, finding the optimal split threshold of each operating condition parameter in the feature subset.

6. The method of claim 1, wherein, The determination of a first decision tree based on the second sample data through a decision tree model comprises: determining the weight of different pulsation types in the second sample data; based on the number and weight of pulsation type samples in the node, calculating the weighted information entropy of the node; based on the weighted information entropy of the node, determining the optimal split rule of the node; repeating the calculation of the weighted information entropy of the node and the determination of the optimal split rule of the node to recursively generate a decision tree until a preset stopping condition is met, to obtain a first decision tree, wherein the stopping condition comprises that the tree depth reaches a first preset value.

7. The method of claim 6, wherein, The determination of the optimal split rule of the node based on the weighted information entropy of the node comprises: determining each candidate operating condition parameter and a corresponding preset threshold; Divide the node sample into two sub-nodes based on the candidate working condition parameter and the corresponding preset threshold value, and calculate the weighted information gain of the divided node based on the weighted information entropy of the node; Combine the candidate working condition parameter and the preset threshold value corresponding to the maximum weighted information gain to determine the optimal split rule of the node.

8. The method of claim 1, wherein, The method further includes: Project the split rule in the first decision tree into a coordinate system to obtain a plurality of polygonal regions; Determine each polygonal region as a visual pulsation trigger window of a different pulsation type.

9. The method of claim 8, wherein, The method further includes: Project the working condition parameter of the key feature in the heavy-duty gas turbine into the coordinate system to obtain a running position of the heavy-duty gas turbine in the coordinate system; If the distance between the running position and the pulsation window is less than a second preset value, adjust the key parameter corresponding to the key feature based on the corresponding split rule.

10. A heavy-duty gas turbine combustion pulsation control apparatus characterized by comprising: The method further includes: A data processing module configured to obtain heavy-duty gas turbine combustion chamber test data, and perform data processing on the combustion chamber test data to obtain first sample data of each working condition parameter; A first determination module configured to determine second sample data of a key feature based on the first sample data by using a random forest model; A second determination module configured to determine a first decision tree based on the second sample data by using a decision tree model; A control module configured to determine a pulsation window region based on the first decision tree, and perform combustion pulsation control on the heavy-duty gas turbine based on the pulsation window region.

11. The apparatus of claim 10, wherein, The data processing module is specifically configured to: Determine a feature matrix of each working condition parameter based on a data acquisition time sequence in the combustion chamber test data; Perform standardization processing on the feature matrix of each working condition parameter to obtain a standardized feature matrix; Label a pulsation type tag based on a pulsation threshold standard for a pulsation value of each time sequence multi-band in the combustion chamber test data to obtain pulsation data; Determine the standardized feature matrix and the pulsation data as the first sample data of each working condition parameter.

12. The apparatus of claim 10, wherein, The first determination module is specifically configured to: Generate a training subset from the first sample data by using Bootstrap; Perform random forest training based on the training subset to obtain a corresponding second decision tree; Determine the contribution degree of each working condition parameter in the second decision tree; Determine the feature importance of each working condition parameter based on the contribution degree of each working condition parameter in the second decision tree; Determine a key feature based on the feature importance of each working condition parameter, and determine the first sample data corresponding to the key feature as the second sample data of the key feature.

13. The apparatus of claim 12, wherein, The first determination module is further configured to: Randomly select a feature subset from the working condition parameters of the training subset; Determine an optimal split threshold value of each working condition parameter in the feature subset; Perform node splitting by using the optimal split threshold value of each working condition parameter to obtain a third decision tree; Adjust the parameters of the third decision tree by using Bayesian optimization to obtain a corresponding second decision tree.

14. The apparatus of claim 13, wherein, The first determination module is further configured to: The Gini index is used to find the optimal split threshold of each working condition parameter in the feature subset.

15. The apparatus of claim 10, wherein, The second determining module is specifically configured to: determine the weight of each pulsation type in the second sample data; calculate the weighted information entropy of the node based on the number and weight of the samples of each pulsation type in the node; determine the optimal split rule of the node based on the weighted information entropy of the node; repeat the calculation of the weighted information entropy of the node and the determination of the optimal split rule of the node to recursively generate a decision tree until a preset stopping condition is met, to obtain a first decision tree, wherein the stopping condition includes that the tree depth reaches a first preset value.

16. The apparatus of claim 15, wherein, The second determining module is specifically configured to: determine each candidate working condition parameter and the corresponding preset threshold value; divide the node samples into left and right two child nodes based on the candidate working condition parameter and the corresponding preset threshold value, and calculate the weighted information gain after the division based on the weighted information entropy of the node; combine the candidate working condition parameter and the preset threshold value corresponding to the maximum weighted information gain to determine the optimal split rule of the node.

17. The apparatus of claim 10, wherein, The control module is specifically configured to: project the split rules in the first decision tree into a coordinate system to obtain a plurality of polygonal regions; determine each polygonal region as a visual pulsation trigger window of a different pulsation type.

18. The apparatus of claim 17, wherein, The control module is further configured to: project the working condition parameters of the key features in the heavy gas turbine into the coordinate system to obtain the running position of the heavy gas turbine in the coordinate system; if the distance between the running position and the pulsation window is less than a second preset value, adjust the key parameters corresponding to the key features based on the corresponding split rule.

19. An electronic device, comprising: comprise: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the method of any one of claims 1 to 9.

20. A computer storage medium, wherein, The computer storage medium stores computer executable instructions; the computer executable instructions are executed by the processor to implement the method of any one of claims 1 to 9.

Images