Method for predicting rotating speed of power generation steam turbine based on differential thermal inertia
By employing differentiated thermal inertia processing, full-process 0-1 constraints, and an adaptive ARIMAX model, the problems of thermal inertia homogenization, numerical drift, and poor model adaptability in turbine speed prediction are solved, achieving high-precision and high-stability speed prediction suitable for turbine DCS systems.
Patent Information
- Application Number
- CN202610176090.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-19
AI Technical Summary
Existing methods for predicting turbine speed suffer from problems such as homogenization of thermal inertia processing, numerical drift, insufficient accuracy of feature selection, and poor model adaptability, resulting in large prediction errors and low stability.
By employing differentiated thermal inertia processing, full-process 0-1 forced constraints, 3D IMF component screening, and adaptive ARIMAX models, and through differentiated lag step adaptation, full-process dimensionless constraints, and adaptive model order adjustment, accurate predictions are achieved at high sampling rates.
It improves feature extraction accuracy, avoids numerical drift, enhances model adaptability, and achieves high-precision and high-stability speed prediction, making it suitable for real-time prediction in steam turbine DCS systems.
Smart Images
Figure CN122064989A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of turbine control and prediction technology, and in particular to a method for predicting the speed of a power generation turbine based on differentiated thermal inertia. Background Technology
[0002] The main turbine speed is a core parameter characterizing the unit's operating status, and accurate speed prediction is crucial for load regulation, early fault warning, and improved operational stability. Existing turbine speed prediction methods are mostly based on traditional time series models or single feature extraction techniques, which have the following prominent drawbacks in practical applications:
[0003] Homogenization of thermal inertia processing: The thermal inertia characteristics of thermodynamic parameters such as lubricating oil temperature, vapor pressure, and condenser vacuum are significantly different (temperature parameters have much greater thermal inertia than vacuum parameters). However, existing methods use a uniform number of lag steps for processing, which cannot fit the actual response law of the thermodynamic system, resulting in low feature extraction accuracy and large prediction error.
[0004] Numerical drift is a prominent issue: both the rotational speed and the input parameters are dimensionless numbers of 0-1, but traditional methods lack full-process numerical constraints. During model training and extrapolation, the predicted values are prone to deviating from the reasonable range due to the amplification of singular values and parameter overflow, thus losing their physical meaning.
[0005] Insufficient feature selection accuracy: After wavelet decomposition, IMF components are often selected using a single dimension (such as correlation), ignoring the error characteristics and frequency adaptability of the components, resulting in the inclusion of invalid noise components and affecting prediction accuracy.
[0006] Poor model adaptability: The ARIMAX model has a fixed order, which cannot cope with the changes in the stationarity of residuals under high sampling rates. When the residuals are non-stationary, the fixed order is still used, resulting in a significant decrease in prediction stability. Summary of the Invention
[0007] To address the aforementioned technical problems, this invention proposes a method for predicting the speed of a power generation turbine based on differentiated thermal inertia. This prediction method is suitable for accurate prediction of the main turbine speed under a high sampling rate of 1ms and can be integrated into the turbine distributed control system (DCS) to provide data support for stable unit operation and fault early warning.
[0008] A method for predicting the rotational speed of a power generation turbine based on differentiated thermal inertia, the specific process of which is as follows:
[0009] S1, Environment Initialization and Parameter Configuration: Setting the Sampling Interval Defining lag time based on thermodynamic parameters and physical properties Through formula Calculate the number of differential lag steps When adapting to a 1ms sampling interval;
[0010] S2, Data Reading and Preprocessing:
[0011] Load or generate turbine operating data using formulas Apply full-process 0-1 forced constraints to all data, and split the training set and validation set, where, For the dimensionless parameters after constraint, Here are the original parameter values, min(x) is the minimum value of x, and max(x) is the maximum value of x. It is the minimum value;
[0012] The effective IMF components are screened using a three-dimensional IMF component screening formula, which is as follows:
[0013] In the formula, the correlation coefficient is calculated as follows:
[0014] In the formula, Let i be the normalized signal of the i-th parameter. For the k-th IMF component, for The mean, for The mean of the signal, where N is the signal length;
[0015] The formula for calculating the root mean square error is: ;
[0016] The formula for calculating the centroid of frequency is: In the formula, For Fast Fourier Transform, For modulo operation, Sampling frequency, This is the frequency value;
[0017] S3, Multi-input Feature Extraction: Wavelet decomposition is performed on each thermodynamic parameter, effective IMF components are selected through a three-dimensional screening formula, and features are reconstructed after noise reduction;
[0018] S4, Differentiated thermal inertia processing: Based on the \tau_i obtained in step S1, add corresponding hysteresis features to each parameter to form the final feature matrix and execute the 0-1 constraints of step S2 again;
[0019] S5, Trend Reconstruction: Perform singular value decomposition, permutation entropy calculation and K-means clustering on the feature matrix to extract low-entropy trend terms as exogenous variables;
[0020] S6, Adaptive ARIMAX Prediction: via formula The model order is dynamically adjusted, and the output range is constrained by the formula in step S2 after prediction.
[0021] S7, Results Evaluation and Output: Calculate prediction error indicators and generate visualization results and detailed reports.
[0022] As a preferred embodiment of the above technical solution, in step S1, the parameters The value is [8,6,6,6,6,6,4,4,4], which configures the wavelet decomposition layer number, the initial ARIMAX order, and the feature selection threshold.
[0023] As a preferred embodiment of the above technical solution, in step S1, the... The grading rules are as follows: lubricating oil temperature corresponds to... =8ms; corresponding steam pressure after regulating stage, steam pressure before quick-closing valve, and pressure after valves 1-3 of quick-closing valves. =6ms; corresponding steam seal pressure, condenser vacuum, and condenser water level. =4ms, and Follow Dynamic adjustment to always ensure lag time Constant.
[0024] As a preferred embodiment of the above technical solution, in step S2, It is a minimum value and takes the value of This is used to avoid the denominator being 0, and the 0-1 mandatory constraint runs through the entire process of data preprocessing, feature extraction, trend reconstruction, and prediction output, ensuring that the physical meaning of all dimensionless parameters is valid.
[0025] As a preferred embodiment of the above technical solution, in step S3, the three indicators of the three-dimensional screening formula are the correlation coefficient used to measure the correlation between the component and the original signal, the RMSE used to measure the component reconstruction error, and the frequency centroid used to measure the frequency characteristics of the component.
[0026] As a preferred embodiment of the above technical solution, in step S3, when the three threshold conditions of correlation coefficient, RMSE and frequency centroid are met simultaneously, the corresponding IMF component is retained for feature robustness.
[0027] As a preferred embodiment of the above technical solution, in step S6, Q is the LBQ test statistic. The critical value of the chi-square distribution with 20 degrees of freedom and 95% confidence is used to dynamically adjust the model order by judging the stationarity of the residuals, thereby improving the prediction stability.
[0028] The beneficial effects of this invention are as follows:
[0029] 1. More realistic thermal inertia processing: The differentiated lag step adaptation formula accurately maps the inertial characteristics of different thermodynamic parameters, solving the feature distortion problem caused by homogenized processing and significantly improving feature extraction accuracy.
[0030] 2. Completely solve numerical drift: The full-process 0-1 forced constraint formula ensures that the physical meaning of all parameters is valid, avoids overflow of predicted values, and provides a reliable guarantee for the implementation of the project.
[0031] 3. Significantly improved feature quality: The 3D IMF screening formula purifies features from multiple dimensions, eliminates noise interference, and provides high-quality input for subsequent predictions.
[0032] 4. Strong model adaptability: The adaptive ARIMAX formula dynamically adjusts the order to cope with changes in data characteristics under high sampling rates, and the prediction stability is better than that of the fixed-order model.
[0033] 5. High sampling rate and good adaptability: All formulas are compatible with various sampling intervals and can be directly integrated into the steam turbine DCS system to achieve real-time prediction. Attached Figure Description
[0034] Figure 1 This is a flowchart of the present invention.
[0035] Figure 2 This is a comparison chart of the predicted rotational speed and the actual rotational speed according to the present invention. Detailed Implementation
[0036] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0037] like Figure 1 The method for predicting the rotational speed of a power generation turbine based on differentiated thermal inertia is shown below. The specific process is as follows:
[0038] S1, Environment Initialization and Parameter Configuration: Setting the Sampling Interval Defining lag time based on thermodynamic parameters and physical properties Through formula Calculate the number of differential lag steps When adapting to a 1ms sampling interval;
[0039] S2, Data Reading and Preprocessing:
[0040] Load or generate turbine operating data using formulas Apply full-process 0-1 forced constraints to all data, and split the training set and validation set, where, For the dimensionless parameters after constraint, Here are the original parameter values, min(x) is the minimum value of x, and max(x) is the maximum value of x. It is the minimum value;
[0041] The effective IMF components are screened using a three-dimensional IMF component screening formula, which is as follows:
[0042] In the formula, the correlation coefficient is calculated as follows:
[0043] In the formula, Let i be the normalized signal of the i-th parameter. For the k-th IMF component, for The mean, for The mean of the signal, where N is the signal length;
[0044] The formula for calculating the root mean square error is: ;
[0045] The formula for calculating the centroid of frequency is: In the formula, For Fast Fourier Transform, For modulo operation, Sampling frequency, This is the frequency value;
[0046] S3, Multi-input Feature Extraction: Wavelet decomposition is performed on each thermodynamic parameter, effective IMF components are selected through a three-dimensional screening formula, and features are reconstructed after noise reduction;
[0047] S4, Differentiated thermal inertia processing: Based on the \tau_i obtained in step S1, add corresponding hysteresis features to each parameter to form the final feature matrix and execute the 0-1 constraints of step S2 again;
[0048] S5, Trend Reconstruction: Perform singular value decomposition, permutation entropy calculation and K-means clustering on the feature matrix to extract low-entropy trend terms as exogenous variables;
[0049] S6, Adaptive ARIMAX Prediction: via formula The model order is dynamically adjusted, and the output range is constrained by the formula in step S2 after prediction.
[0050] S7, Results Evaluation and Output: Calculate prediction error indicators and generate visualization results and detailed reports.
[0051] As a preferred embodiment of the above technical solution, in step S1, the parameters The value is [8,6,6,6,6,6,4,4,4], which configures the wavelet decomposition layer number, the initial ARIMAX order, and the feature selection threshold.
[0052] In this embodiment, in step S1, the The grading rules are as follows: lubricating oil temperature corresponds to... =8ms; corresponding steam pressure after regulating stage, steam pressure before quick-closing valve, and pressure after valves 1-3 of quick-closing valves. =6ms; corresponding steam seal pressure, condenser vacuum, and condenser water level. =4ms, and Follow Dynamic adjustment to always ensure lag time Constant.
[0053] In this embodiment, in step S2, It is a minimum value and takes the value of This is used to avoid the denominator being 0, and the 0-1 mandatory constraint runs through the entire process of data preprocessing, feature extraction, trend reconstruction, and prediction output, ensuring that the physical meaning of all dimensionless parameters is valid.
[0054] In this embodiment, in step S3, the three indicators of the three-dimensional screening formula are the correlation coefficient used to measure the correlation between the component and the original signal, the RMSE used to measure the component reconstruction error, and the frequency centroid used to measure the frequency characteristics of the component.
[0055] In this embodiment, in step S3, the corresponding IMF component is retained when the three threshold conditions of correlation coefficient, RMSE and frequency centroid are met simultaneously, so as to be used for feature robustness.
[0056] In this embodiment, in step S6, Q is the LBQ test statistic. The critical value of the chi-square distribution with 20 degrees of freedom and 95% confidence is used to dynamically adjust the model order by judging the stationarity of the residuals, thereby improving the prediction stability.
[0057] This application solves the problem of homogenization of thermal inertia by using a differentiated thermal inertia hysteresis step adaptation formula, breaking through the traditional unified hysteresis processing logic. Based on the thermal parameters such as heat capacity, volume, and flow characteristics, a fixed hysteresis time T_{lag,i} is set in different grades. The sampling interval t_s is dynamically adapted through the formula to ensure that the hysteresis time is constant under different sampling rates, and accurately maps the differences in the inertial response of the thermal system.
[0058] The numerical drift problem was solved by using a dimensionless 0-1 forced constraint formula throughout the entire process. This constraint formula is applied to the entire process of data preprocessing, feature extraction, trend reconstruction, and prediction output, forcing all dimensionless parameters to remain within the physically valid 0-1 range. This completely solves the problem of predicted values drifting outside the reasonable range in traditional methods, providing a core guarantee for engineering implementation.
[0059] The feature selection accuracy is improved by using a three-dimensional IMF component selection formula. It integrates three-dimensional indicators: correlation (representing the correlation between the component and the original signal), RMSE (representing the component reconstruction accuracy), and frequency centroid (representing the component frequency adaptability). Only IMF components that simultaneously meet the threshold conditions are retained. Compared with single-dimensional selection, it effectively removes noise components and significantly improves feature robustness.
[0060] The adaptive ARIMAX model order adjustment formula improves model adaptability. Based on the LBQ test statistic, the stationarity of the residuals is judged. When the residuals are non-stationary (Q > 31.41), the model order is automatically adjusted from ARIMA(3,1,1) to ARIMA(2,1,1), which breaks through the limitations of traditional fixed-order models, adapts to changes in data characteristics under high sampling rates, and significantly improves prediction stability.
[0061] This application accurately extracts thermodynamic parameter features through a differentiated thermal inertia formula, ensures numerical validity through a full-process 0-1 constraint formula, purifies features with a three-dimensional screening formula, and finally achieves accurate prediction through an adaptive ARIMAX formula. Each step supports the other to build a high-precision and high-stability prediction system.
[0062] The following is a further explanation with reference to specific embodiments:
[0063] This embodiment focuses on the scenario of predicting the speed of a generator turbine with a 1ms sampling interval. The core parameter configuration is as follows: sampling frequency. =1000Hz, training data length Ntrain=800 points (800ms), prediction data length Npred=140 points (140ms), wavelet basis function is db4, decomposition layer number L=4, permutation entropy embedding dimension m=3, delay time =1, number of K-means clusters =2.
[0064] S1: Environment Initialization and Differentiated Parameter Configuration
[0065] Based on the formula for innovation point 1, configure the differentiated lag steps: The ms range is [8ms, 6ms, 4ms]. =1ms, calculate various parameters of this patent. =[8,6,6,6,6,6,4,4,4]; Configure the three-dimensional filtering thresholds (correlation coefficient ≥ 0.6, RMSE ≤ 0.005, frequency centroid ≤ 5Hz), and the initial ARIMAX order (3,1,1).
[0066] S2: Data Preprocessing and End-to-End Constraints
[0067] Apply 0-1 constraints to all data using the formula in Innovation Point 2:
[0068]
[0069] Ensure that all parameters in the training and validation sets are in the range of 0-1.
[0070] S3: Feature Extraction and 3D IMF Screening
[0071] Four-level db4 wavelet decomposition was performed on each of the nine thermodynamic parameters to obtain approximation coefficients and detail coefficients. Effective IMF components were then selected using the three-dimensional screening formula of innovation point 3.
[0072] Calculate the correlation coefficient, RMSE, and frequency centroid of each IMF component. Only retain the components that simultaneously satisfy "correlation coefficient ≥ 0.6, RMSE ≤ 0.005, frequency centroid ≤ 5Hz". Perform soft threshold noise reduction on the retained components and then sum and reconstruct the features. After reconstruction, the range is constrained again by the formula of innovation point 2.
[0073] S4: Differentiated Thermal Inertia Treatment
[0074] Based on the results obtained in step 1 Add corresponding lag features to each parameter: add 8 lag steps to the first parameter, 6 lag steps to the second to sixth parameters, and 4 lag steps to the seventh to ninth parameters. Concatenate these features to form the final feature matrix and execute the formula constraints of innovation point 2.
[0075] Step S5: Trend Reconstruction
[0076] Singular value decomposition is performed on the feature matrix to calculate the permutation entropy of each singular value component. Low-entropy trend terms are extracted by K-means clustering and used as exogenous variables of the ARIMAX model after constraining the range.
[0077] S6: Adaptive ARIMAX Prediction
[0078] Based on innovation point 4, the model order is dynamically adjusted using the formula: the LBQ test statistic Q is calculated; if Q < 31.41, the ARIMA(3,1,1) model is adopted; otherwise, it is adjusted to ARIMA(2,1,1). The model is trained with exogenous variables to predict rotational speeds over 140 steps, and the predicted value range is constrained by the formula in innovation point 2.
[0079] S7: Results Evaluation and Output
[0080] Calculate RMSE, MAE, and MAPE error indices to evaluate the prediction performance.
[0081] In this embodiment, with a 1ms sampling interval, the predicted value is always constrained to the range of 0-1, with no numerical drift problem; RMSE ≤ 0.05. Compared with the traditional unified thermal inertia processing method, the prediction accuracy is improved by 18% and the stability by 22%, fully meeting the requirements for real-time prediction of turbine speed. Figure 2 As shown.
[0082] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting the rotational speed of a power generation turbine based on differentiated thermal inertia, characterized in that: The specific process is as follows: S1, Environment Initialization and Parameter Configuration: Setting the Sampling Interval Defining lag time based on thermodynamic parameters and physical properties Through formula Calculate the number of differential lag steps When adapting to a 1ms sampling interval; S2, Data Reading and Preprocessing: Load or generate turbine operating data using formulas Apply full-process 0-1 forced constraints to all data, and split the training set and validation set, where, For the dimensionless parameters after constraint, Here are the original parameter values, min(x) is the minimum value of x, and max(x) is the maximum value of x. It is the minimum value; The effective IMF components are screened using a three-dimensional IMF component screening formula, which is as follows: In the formula, the correlation coefficient is calculated as follows: In the formula, Let i be the normalized signal of the i-th parameter. For the k-th IMF component, for The mean, for The mean of the signal, where N is the signal length; The formula for calculating the root mean square error is: ; The formula for calculating the centroid of frequency is: In the formula, For Fast Fourier Transform, For modulo operation, Sampling frequency, This is the frequency value; S3, Multi-input Feature Extraction: Wavelet decomposition is performed on each thermodynamic parameter, effective IMF components are selected through a three-dimensional screening formula, and features are reconstructed after noise reduction; S4, Differentiated thermal inertia processing: Based on the \tau_i obtained in step S1, add corresponding hysteresis features to each parameter to form the final feature matrix and execute the 0-1 constraints of step S2 again; S5, Trend Reconstruction: Perform singular value decomposition, permutation entropy calculation and K-means clustering on the feature matrix to extract low-entropy trend terms as exogenous variables; S6, Adaptive ARIMAX Prediction: via formula The model order is dynamically adjusted, and the output range is constrained by the formula in step S2 after prediction. S7, Results Evaluation and Output: Calculate prediction error indicators and generate visualization results and detailed reports.
2. The method for predicting the rotational speed of a power generation turbine based on differentiated thermal inertia according to claim 1, characterized in that: In step S1, the parameters The value is [8,6,6,6,6,6,4,4,4], which configures the wavelet decomposition layer number, the initial ARIMAX order, and the feature selection threshold.
3. The method for predicting the rotational speed of a power generation turbine based on differentiated thermal inertia according to claim 2, characterized in that: In step S1, the The grading rules are as follows: lubricating oil temperature corresponds to... =8ms; corresponding steam pressure after regulating stage, steam pressure before quick-closing valve, and pressure after valves 1-3 of quick-closing valves. =6ms; corresponding steam seal pressure, condenser vacuum, and condenser water level. =4ms, and Follow Dynamic adjustment to always ensure lag time Constant.
4. The method for predicting the rotational speed of a power generation turbine based on differentiated thermal inertia according to claim 1, characterized in that: In step S2, It is a minimum value and takes the value of This is used to avoid the denominator being 0, and the 0-1 mandatory constraint runs through the entire process of data preprocessing, feature extraction, trend reconstruction, and prediction output, ensuring that the physical meaning of all dimensionless parameters is valid.
5. The method for predicting the rotational speed of a power generation turbine based on differentiated thermal inertia according to claim 1, characterized in that: In step S3, the three indicators of the three-dimensional screening formula are the correlation coefficient used to measure the correlation between the component and the original signal, the RMSE used to measure the component reconstruction error, and the frequency centroid used to measure the frequency characteristics of the component.
6. The method for predicting the rotational speed of a power generation turbine based on differentiated thermal inertia according to claim 5, characterized in that: In step S3, the corresponding IMF component is retained when the three threshold conditions of correlation coefficient, RMSE and frequency centroid are met simultaneously, so as to be used for feature robustness.
7. The method for predicting the rotational speed of a power generation turbine based on differentiated thermal inertia according to claim 1, characterized in that: In step S6, Q is the LBQ test statistic. The critical value of the chi-square distribution with 20 degrees of freedom and 95% confidence is used to dynamically adjust the model order by judging the stationarity of the residuals, thereby improving the prediction stability.