Main steam flow soft measurement method and system based on machine learning and storage medium
By constructing the TCN-GRU-RF model for main steam flow prediction, the problems of low accuracy and large energy loss in existing technologies are solved, and more efficient and accurate flow monitoring is achieved.
Patent Information
- Application Number
- CN202510869545.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-26
- Publication Date
- 2025-11-28
AI Technical Summary
Existing technologies suffer from low accuracy and large energy loss when measuring the main steam flow of large units, especially with large deviations in model calculation results under deep peak shaving conditions, making it difficult to meet the requirements for high flow measurement.
A machine learning-based TCN-GRU-RF model is used for soft measurement of main steam flow. The TCN-GRU-RF model is constructed for flow prediction through data acquisition, preprocessing, dimensionality reduction by kernel principal component analysis, TCN layer feature extraction, GRU layer time series prediction, and RF layer error correction.
It improves the accuracy of main steam flow prediction, reduces computational complexity, enhances the model's generalization ability, avoids gradient vanishing and gradient exploding problems, and achieves more efficient flow monitoring.
Smart Images
Figure CN121031262A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of thermal parameters monitoring of energy power systems, and particularly relates to a main steam flow soft measurement method and system based on machine learning and a storage medium. BACKGROUND
[0002] The main steam flow, as one of the core parameters of the thermal system of a power plant, has an important influence on the operation and performance of the unit. Accurate prediction of the main steam flow can achieve the goals of operation monitoring, energy efficiency evaluation, safety control and operation optimization of the unit, thereby improving the operation efficiency and reliability of the unit. At present, the measurement technology of the main steam flow of the unit can be mainly divided into direct measurement and indirect measurement according to the measurement principle. The direct measurement mainly adopts the differential pressure method, which is realized by installing a throttling device such as a flow orifice plate or a nozzle flowmeter. Although this way of measurement is intuitive, simple and convenient to operate. However, in the actual operation process of the unit, the steam passing through the throttling device inevitably causes a certain pressure loss, which leads to an increase in system energy loss. In addition, in recent years, a large number of ultra-supercritical units have been put into operation, and the existing standard throttling device has low precision, and even cannot meet the measurement requirements of high flow of large-capacity units, and the throttling loss cannot be ignored. Therefore, the indirect measurement method is usually used to replace the throttling device for measuring the main steam flow of modern large units.
[0003] Based on the Friulli formula, the indirect measurement usually uses the steam turbine stage group before and after the pressure and temperature correction and other thermal parameters to substitute and convert to obtain the value of the main steam flow. However, when the flow path of the steam turbine is changed due to scaling or corrosion, the measurement result of the main steam flow will have a large deviation. Especially in the context of deep peak regulation, the unit needs to be in variable load and low load operation condition for a long time, and the relationship between the regulating stage pressure and the main steam flow is not ideal linear. Although researchers have made many modifications to the Friulli formula, the model calculation result still has a large deviation, and the use range is limited. SUMMARY
[0004] The purpose of the present application is to provide a main steam flow soft measurement method and system based on machine learning and a storage medium to solve the problem that the main steam flow of the unit is difficult to accurately monitor at present.
[0005] To achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:
[0006] A main steam flow soft measurement method based on machine learning, comprising the following steps:
[0007] Step 1, collecting the original data of the operation parameters related to the main steam flow from the power plant monitoring system;
[0008] Step 2: Impute missing and outlier values in the original data obtained in Step 1, and then normalize the data.
[0009] Step 3: Dimensionality reduction is performed on the data processed in Step 2 using kernel principal component analysis (KPCA).
[0010] Step 4: Construct the TCN-GRU-RF model, which includes an input layer, a TCN layer, a GRU layer, an RF layer, and an output layer connected in sequence. The TCN layer extracts the feature information of the main steam flow, the GRU layer receives the output of the TCN layer and optimizes the time series prediction results, and the RF layer corrects the error of the prediction results output by the GRU layer.
[0011] The TCN-GRU-RF model is trained based on the data processed in step 3.
[0012] Step 5: Input the data to be tested into the TCN-GRU-RF model trained in Step 4, and output the measured value of the main steam flow of the thermal power unit.
[0013] Furthermore, in step 2, the missing and outlier values in the original data are replaced using linear interpolation results according to the time labels.
[0014] Furthermore, in step 2, the normalization formula is:
[0015]
[0016] Where, x i,min x i,max Let represent the minimum and maximum values of the i-th variable, respectively. x represents the normalized value. i (t) represents the original data.
[0017] Furthermore, in step 3, the dimensionality reduction process of kernel principal component analysis (KPCA) includes: using kernel functions to obtain random higher-order correlations between input variables, and mapping the original dataset to a high-dimensional feature space, thereby making the originally linearly inseparable data linearly separable in the new feature space.
[0018] Furthermore, in step 4, the TCN layer consists of causal convolution, extended convolution, and residual blocks.
[0019] Furthermore, in step 4, the GRU layer achieves temporal feature fusion through update gates and reset gates.
[0020] Furthermore, in step 4, the RF layer consists of multiple different decision trees. Each decision tree is trained independently on a randomly selected part. Since the model features selected by each decision tree are not the same, multiple results are obtained. Based on the results of each decision tree, voting is performed and the average value is taken to obtain the final result.
[0021] Furthermore, in step 4, the TCN-GRU-RF model training includes a hyperparameter tuning process: grid search technology is used to further optimize the hyperparameters of the TCN-GRU-RF model to avoid random errors in the prediction results, and evaluation indicators are calculated for each parameter.
[0022] Furthermore, the hyperparameter tuning process includes:
[0023] (1) TCN module parameter tuning: The optimal value was determined to be 5 in the candidate set of convolution kernel size [3,5,7,9]; the optimal value was determined to be 2 in the candidate set of dilation coefficient [1,2,4]; and the optimal value was determined to be 64 in the candidate set of number of channels [32,64,128].
[0024] (2) GRU module parameter tuning: The optimal value was determined to be 2 from the candidate set of attention heads [1,2,4,8];
[0025] (3) RF module tuning: The number of trees in the random forest is set to 100, the maximum depth is set to 10, and the minimum number of samples in the leaf node is set to 2.
[0026] (4) Training strategy optimization: 256 was selected as the optimal batch size from the candidate batch size set [16,32,64,128,256,512]; the Adam optimizer was used to determine the optimal learning rate as 0.01.
[0027] A main steam flow measurement system, comprising:
[0028] The data acquisition module obtains power plant operating parameters in real time;
[0029] The preprocessing module performs linear interpolation, range normalization, and KPCA dimensionality reduction.
[0030] The model computation module deploys the TCN-GRU-RF model and performs prediction calculations.
[0031] The visualization terminal displays real-time measurements and historical trend curves.
[0032] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described main steam flow measurement method based on TCN-GRU-RF.
[0033] Beneficial Effects: This invention provides a machine learning-based soft measurement method, system, and storage medium for main steam flow. Combining the efficient parallel computing capabilities of TCN and the short prediction time of GRU, a TCN-GRU model is constructed to predict main steam flow, and random forest (RF) is used for error correction. Compared with existing prediction models, this model avoids the gradient vanishing and gradient exploding problems that occur in time series prediction. Furthermore, the introduced GRU ensemble optimizes and integrates the TCN output, further improving the accuracy of the prediction results. Simultaneously, the simple GRU layer structure, combined with the TCN layer, reduces computational complexity while maintaining the accuracy of main steam flow prediction, and also improves the model's performance and generalization ability, facilitating model training. Attached Figure Description
[0034] Figure 1 A flowchart of a main steam flow measurement method based on the TCN-GRU-RF model;
[0035] Figure 2 This is a unit load curve diagram from the embodiment;
[0036] Figure 3 This is a graph showing the results of the kernel principal component analysis in the embodiment;
[0037] Figure 4 This is a diagram of the GRU network structure in the embodiment;
[0038] Figure 5 This is a schematic diagram of the random forest algorithm in the embodiment;
[0039] Figure 6 This is a schematic diagram of the TCN-GRU-RF model in the embodiment;
[0040] Figure 7 This is a graph showing the main steam flow prediction results based on the TCN-GRU-RF model in the embodiment;
[0041] Figure 8 This is a graph showing the main steam flow prediction results based on the GRU model in the embodiment. Detailed Implementation
[0042] The present invention will be further explained below with reference to the accompanying drawings and embodiments.
[0043] Example
[0044] like Figure 1 As shown, the main steam flow soft measurement method based on machine learning in this embodiment includes the following steps:
[0045] (1) Collect raw data of operating parameters related to main steam flow from the power plant monitoring system, fill in the missing and outlier values in the raw data, normalize the data, and then perform dimensionality reduction processing through kernel principal component analysis (KPCA).
[0046] Modern large-scale generator sets are generally reheat units. After pretreatment, feedwater is pumped into the boiler to exchange heat with the fuel, generating high-temperature, high-pressure steam. This steam then enters the superheater through steam pipes and is reheated to form superheated steam that meets specific pressure and temperature requirements. This superheated steam drives the turbine to perform work, which in turn drives the generator to convert mechanical energy into electrical energy. After energy release, some of the steam is extracted for reheat cycles or heating, while the remaining steam (exhaust steam) is condensed into water in the condenser and returned to the feedwater pump, forming a cycle.
[0047] Based on the above analysis of the unit's steam cycle process, the generation and consumption of main steam in thermal power units involve complex processes and are influenced by multiple factors, including boiler fuel and combustion efficiency, boiler parameters, water circulation system, power generation load, auxiliary equipment requirements, and steam extraction. In conjunction with the recommendations of operation experts, 18 variables were initially selected as auxiliary variables for this experiment, as shown in Table 1.
[0048] Table 1 List of Auxiliary Variables
[0049]
[0050] Taking a 1000MW ultra-supercritical once-through reheat generator unit as the research object, historical data of the unit from 0:00 on March 2, 2025 to 22:40 on March 8, 2025 were extracted from the SIS system, totaling 10,000 samples, with a sampling time of 60 seconds. Figure 2 As can be seen from the load curves shown, this set of data basically includes various operating conditions of the target unit under stable operation and variable load (including load increase and load decrease operation), and the unit load range is [330MW, 1000MW].
[0051] Faced with a large amount of raw data, some data was missing or outliers due to the complex environment of thermal power unit production sites and the influence of signal noise. In this case, firstly, linear interpolation results were used to replace the missing and outliers in the data according to the time label to ensure data integrity and accuracy. Secondly, to avoid gradient explosion in the model output and improve the model's accuracy and training efficiency, the original input data was normalized, as calculated using the following formula:
[0052]
[0053] Where, x i,minx i,max Let represent the minimum and maximum values of the i-th variable, respectively. x represents the normalized value. i (t) represents the original data.
[0054] Since the 18 candidate variables shown in Table 1 exhibit both linear and nonlinear relationships, KPCA is employed to further reduce the dimensionality of the feature data and select the optimal variables to prevent the curse of dimensionality. The main principles are as follows:
[0055] Principal Component Analysis (PCA) is a widely used data analysis method for dimensionality reduction. It projects the original data into a new, lower-dimensional space through linear transformation, aiming to convert high-dimensional datasets into a lower-dimensional representation while preserving key information. However, due to the complexity of thermal power unit systems, the large amount of operational data they generate contains complex nonlinear and strongly coupled relationships, which may prevent PCA from fully capturing the nonlinear characteristics of thermal power unit data. Furthermore, abnormal operations or measurement errors may occur during unit operation, and these outliers may further adversely affect the PCA results. In contrast, Kernel Principal Component Analysis (KPCA), as a nonlinear multivariate statistical technique, has better applicability when processing unit data. KPCA uses kernel functions to obtain stochastic higher-order correlations between input variables and maps the original dataset to a high-dimensional feature space, thus making originally linearly inseparable data linearly separable in the new feature space. The algorithm principle is as follows:
[0056] Suppose that for N standardized sample points a1, a2, ..., a N If each sample point has M features, then the original dataset A can be represented as:
[0057] A = {a1, a2, ..., a} N}, a i ∈R M , i = 1, 2, ..., N;
[0058] In the formula, M is the data dimension, which is also the number of features, and R... M represents the input space, and i represents the index of the sample point.
[0059] To facilitate the mapping of the original input space to the high-dimensional feature space F via nonlinearity, a nonlinear mapping is defined.
[0060] Where Q represents the dimension of the high-dimensional feature space data sample, Q>M, R QThis represents a space of dimension Q.
[0061] Further through a i Mapped to This represents the result of mapping the i-th sample point to a high-dimensional space, realizing the transformation of the original dataset from a low-dimensional feature space to a high-dimensional feature space. Therefore, the covariance matrix C of the mapped data in the high-dimensional feature space F is... F As shown below:
[0062]
[0063] Among them, superscript Represents the transpose of a matrix. The covariance matrix C represents the result of mapping the i-th sample point to a high-dimensional space. F Let V be a square matrix of order Q, and assume its eigenvalues are λ. Then the eigenvectors V satisfy:
[0064] λV=C F V;
[0065] Furthermore, when λ≠0, the eigenvector V can be linearly represented as:
[0066]
[0067] Where, α i This represents the coefficients of the linear combination.
[0068] Define the kernel function matrix K, and let Where i and j represent the sample point indices, and K ij The high-dimensional inner product of sample pairs can be simplified by combining the above three equations to obtain:
[0069]
[0070] Where α represents a vector consisting of N linear combination coefficients, and the superscript... This represents the transpose of a matrix.
[0071] At this point, the problem of finding the eigenvectors is transformed into finding the corresponding α values. KPCA obtains eigenvalues and eigenvectors through the kernel function matrix, and selects the eigenvector with the highest cumulative contribution rate as the new principal component to form a new feature space, thereby achieving dimensionality reduction of the data.
[0072] The results of the kernel principal component analysis of the candidate variables are as follows: Figure 3 As shown. Since the contribution rates of principal components 13-18 are relatively low and insignificant, therefore... Figure 3 Only the specific contribution rates of the first 12 principal components are shown. According to Figure 3It can be seen that the first four principal components represent the largest contribution rates, with the first kernel principal component (KPCA1) contributing as much as 40% of the variance. The cumulative variance contribution curve shows that the first eight principal components contribute 98% of the total, essentially covering the main information of the original data. The contribution rates of the ninth and subsequent principal components are nearly horizontal, and the cumulative variance contribution curve also tends to flatten, indicating that the contribution rates from the ninth kernel principal component onwards are very low, containing very little original information. Therefore, the first eight principal component variables are selected as the final input to the model.
[0073] (2) Construct a TCN-GRU-RF model, which includes an input layer, a TCN layer, a GRU layer, an RF layer, and an output layer connected in sequence. The TCN layer extracts the characteristic information of the main steam flow, the GRU layer receives the output of the TCN layer and optimizes the time series prediction results, and the RF layer corrects the error of the prediction results output by the GRU layer. Train the TCN-GRU-RF model based on the data processed in step (1).
[0074] TCN is a time-series-based convolutional neural network architecture composed of causal convolution, expanded convolution, and residual blocks. In predicting main steam flow, TCN layers can extract feature information about the main steam flow, effectively addressing time-series problems. In causal convolution, layers strictly adhere to causal relationships; the value of each layer at time t depends on the values of the previous layer at time t and before time t, fully utilizing relevant historical information. To address the issue of excessive causal convolution layers, expanded convolution is introduced to increase the receptive field, allowing for the acquisition of more historical information with fewer layers. Residual blocks effectively solve the gradient vanishing and gradient exploding problems found in deep network models, enabling faster model convergence. The formula for one-dimensional expanded causal convolution is:
[0075]
[0076] In the formula, k is the kernel size, f is the filter, x is the input sequence, and s-di indicates that convolution operation is performed only on the historical input.
[0077] GRU (Gated Recurrent Unit), as another special form of recurrent neural network (RNN), is similar to LSTM in that it also has a gating mechanism. However, it simplifies the structure of LSTM to some extent, reducing the number of gates and making the network easier to train and compute. Compared to LSTM, GRU only contains an update gate z. t and reset door r tThe system employs two gating units. The update gate determines the degree of influence of the previous hidden state on the current state, while the reset gate controls how the previous hidden state is combined with the current input. Through the control of these two gates, the GRU can dynamically decide whether to retain past information and how to integrate past information with the current input, thereby better capturing long-term dependencies in time series data.
[0078] Its neuronal internal structure is as follows Figure 4 As shown, the corresponding calculation formula is as follows:
[0079] z t =σ(W z ·[h t-1 ,x t ])
[0080] r t =σ(W r ·[h t-1 ,x t ])
[0081] In the formula, h t-1 It is the hidden state of the previous time step, x t It is the input at the current time step, W Z W r Here, σ represents the weight matrices for the update gate and reset gate, respectively, and σ represents the sigmoid activation function. By resetting the gate and applying the current input, the candidate hidden state is obtained as follows:
[0082]
[0083] Where ⊙ denotes element-wise multiplication, and W is the weight matrix used to calculate the candidate hidden states. Finally, based on the update gate and the candidate hidden states, the hidden state at the current time step is obtained as follows:
[0084]
[0085] By linearly combining past hidden states with candidate hidden states, the update gate controls how much past information to retain, and the candidate hidden states are combined to generate the hidden state for the current time step. Compared to LSTM, the GRU network structure has one less gating component, thus reducing the number of parameters and making it more resource-efficient during model training. However, this also means that the GRU model sometimes cannot provide enough information to uniquely determine the system's state at the next time step, so in most cases, the model training accuracy is not significantly improved. Furthermore, GRU has a disadvantage in parallel processing, as it can only learn sequentially from beginning to end.
[0086] Random Forest (RF) is an ensemble learning algorithm based on decision trees. It solves classification and regression problems by constructing multiple decision trees, achieving high accuracy and good generalization ability. The Random Forest algorithm consists of multiple different decision trees, each trained independently on randomly selected features. Since each decision tree selects different model features, multiple results are obtained. The final result is obtained by using methods such as voting and averaging based on the results of each decision tree. The Random Forest algorithm is described below. Figure 5 As shown.
[0087] Due to the frequent long-term uninterrupted operation of thermal power units, the historical data they generate is typically long-term, large in volume, and complex in relationships. Therefore, this invention combines the efficient parallel computing capabilities of TCN with the short prediction time of GRU to construct a TCN-GRU model for predicting main steam flow. Random forest (RF) is used for error correction. Compared with existing prediction models, this model avoids the gradient vanishing and gradient exploding problems that occur in time series prediction. Furthermore, the introduced GRU ensemble optimizes and integrates the TCN output, further improving the accuracy of the prediction results. Simultaneously, the simple structure of the GRU layer, combined with the TCN layer, reduces computational complexity while maintaining the accuracy of main steam flow prediction, and also improves the model's performance and generalization ability, facilitating model training. A schematic diagram of the constructed TCN-GRU-RF model is shown below. Figure 6 As shown.
[0088] During the experiment, 8000 sets of the processed original dataset were selected as the training set and 2000 sets as the test set in an 8:2 ratio to train the constructed TCN-GRU-RF model. Mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) were used as the statistical measures. 2 Three metrics are used to evaluate model performance, and their definitions are as follows:
[0089]
[0090] In the formula, y i , represents the actual value, the predicted value, and the mean of the actual value, respectively, and n represents the total number of samples.
[0091] To further improve the prediction accuracy and fitting effect of the network model, grid search technology was used to further optimize its hyperparameters. TCN module parameter tuning: the optimal value for kernel size was determined to be 5 from the candidate set [3,5,7,9]; the optimal value for dilation coefficient was determined to be 2 from the candidate set [1,2,4]; the optimal value for channel number was determined to be 64 from the candidate set [32,64,128]. GRU module parameter tuning: the optimal value for attention head number was determined to be 2 from the candidate set [1,2,4,8]. RF module tuning: the number of trees in the random forest was set to 100, the maximum depth was set to 10, and the minimum number of samples per leaf node was set to 2. Training strategy tuning: 256 was selected as the optimal batch size from the candidate set [16,32,64,128,256,512]. The Adam optimizer was used, and the optimal learning rate was determined to be 0.01.
[0092] (3) Input the data to be measured into the TCN-GRU-RF model and output the measured value of the main steam flow of the thermal power unit.
[0093] The mean and standard deviation of the prediction errors of TCN-GRU-RF and GRU networks are shown in Table 2.
[0094] Table 2 shows the predicted performance of TCN-GRU-RF and GRU on the test set.
[0095]
[0096] The TCN-GRU-RF model has an RMSE of 28.356 t / h, a MAE of 20.878 t / h, and R... 2 The value is 0.996, and all performance metrics are superior to the GRU network.
[0097] Based on the above hyperparameter optimization results, the main steam flow rate is predicted using test set data. The TCN-GRU-RF model constructed in this invention can accurately predict its changing trend. The model prediction results can determine the inflection point of the actual change in the main steam flow rate, remain stable within the period of stable flow rate, and maintain a good fit with the actual value, such as... Figure 7 As shown. Figure 8 This is the predicted main steam flow rate for the GRU.
[0098] This embodiment addresses the problem of inaccurate monitoring of main steam flow in steam turbines by proposing a main steam flow measurement method based on the TCN-GRU-RF model. Simulation verification is performed using operating data from a 1000MW ultra-supercritical single-reheat unit. A TCN-GRU model is constructed to predict the main steam flow. Random forest (RF) is used for error correction, and the KPCA algorithm is employed to screen and reduce the dimensionality of the original candidate variables. The top eight principal component variables are selected as the final input to the model, with a cumulative contribution rate of 98%, essentially covering the main information of the original data. The hyperparameter-optimized TCN-GRU-RF model exhibits better prediction performance. Compared to the prediction results with unreduced data, using KPCA for dimensionality reduction effectively reduces the impact on prediction accuracy without losing the main information of the original data, and also reduces model training time.
[0099] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described machine learning-based soft measurement method for main steam flow.
[0100] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A machine learning-based soft measurement method for main steam flow rate, characterized in that, Includes the following steps: Step 1: Collect raw data of operating parameters related to main steam flow from the power plant monitoring system; Step 2: Impute missing and outlier values in the original data obtained in Step 1, and then normalize the data. Step 3: Perform dimensionality reduction on the data processed in Step 2 using kernel principal component analysis; Step 4: Construct the TCN-GRU-RF model, which includes an input layer, a TCN layer, a GRU layer, an RF layer, and an output layer connected in sequence. The TCN layer extracts the feature information of the main steam flow, the GRU layer receives the output of the TCN layer and optimizes the time series prediction results, and the RF layer corrects the error of the prediction results output by the GRU layer. The TCN-GRU-RF model is trained based on the data processed in step 3. Step 5: Input the data to be tested into the TCN-GRU-RF model trained in Step 4, and output the measured value of the main steam flow of the thermal power unit.
2. The method according to claim 1, characterized in that, In step 2, the missing and outlier values in the original data are replaced with linear interpolation results according to the time label, and then the data is normalized.
3. The method according to claim 1, characterized in that, In step 3, the dimensionality reduction process of kernel principal component analysis includes: using kernel functions to obtain random higher-order correlations between input variables, and mapping the original dataset to a high-dimensional feature space, thereby making the originally linearly inseparable data linearly separable in the new feature space.
4. The method according to claim 1, characterized in that, In step 4, the TCN layer consists of causal convolution, extended convolution, and residual blocks.
5. The method according to claim 1, characterized in that, In step 4, the GRU layer achieves temporal feature fusion through update gates and reset gates.
6. The method according to claim 1, characterized in that, In step 4, the RF layer consists of multiple different decision trees. Each decision tree is trained independently on a randomly selected part. Since the model features selected by each decision tree are not the same, multiple results are obtained. Based on the results of each decision tree, voting is performed and the average value is taken to obtain the final result.
7. The method according to claim 1, characterized in that, In step 4, the training of the TCN-GRU-RF model includes a hyperparameter tuning process: grid search technology is used to further optimize the hyperparameters of the TCN-GRU-RF model to avoid random errors in the prediction results, and evaluation indicators are calculated for each parameter.
8. The method according to claim 7, characterized in that, The hyperparameter tuning process includes: (1) TCN module parameter tuning: The optimal value was determined to be 5 in the candidate set of convolution kernel size [3,5,7,9]; the optimal value was determined to be 2 in the candidate set of dilation coefficient [1,2,4]; and the optimal value was determined to be 64 in the candidate set of number of channels [32,64,128]. (2) GRU module parameter tuning: The optimal value was determined to be 2 from the candidate set of attention heads [1,2,4,8]; (3) RF module tuning: The number of trees in the random forest is set to 100, the maximum depth is set to 10, and the minimum number of samples in the leaf node is set to 2. (4) Training strategy optimization: 256 was selected as the optimal batch size from the candidate batch size set [16,32,64,128,256,512]; the Adam optimizer was used to determine the optimal learning rate as 0.
01.
9. A main steam flow measurement system for implementing the method of any one of claims 1-8, characterized in that, include: The data acquisition module obtains power plant operating parameters in real time; The preprocessing module performs linear interpolation, range normalization, and KPCA dimensionality reduction. The model computation module deploys the TCN-GRU-RF model and performs prediction calculations. The visualization terminal displays real-time measurements and historical trend curves.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the main steam flow measurement method based on TCN-GRU-RF as described in any one of claims 1 to 8.