A method for predicting steam turbine back pressure trend based on catboost algorithm

Abstract

The invention discloses a method for predicting a back pressure trend of a steam turbine based on a catboost algorithm, relates to the field of machine learning, and comprises the following steps: S1 prepares historical operation data, S2 pre-processes the historical operation data, S3 performs dimension-upgrading processing on feature engineering of the pre-processed data, S4 resamples and screens variables with a large correlation with the back pressure, S5 determines parameters and models the model through catboost, predicts the back pressure within a certain period of time with the screened historical operation data, S6 trains a preparatory model while predicting the trained model, and replaces the current prediction model with an offline trained model when the error of the current prediction model is too large; the method is indirect and easy to operate, can effectively predict the back pressure of the steam turbine, the prediction model and offline training are carried out in parallel, the error of the prediction model can be compensated in time, and the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, belongs to power plant operation and maintenance technology, and specifically relates to a method for predicting turbine back pressure trend based on CatBoost algorithm. Background Art

[0002] With the rapid development of the electric power industry, more and more large-scale units have been put into operation. The increase in unit capacity has made their structure and system increasingly complex. How to ensure that these units can operate safely and reliably is of great significance to the development of the national economy.

[0003] In a sense, trend prediction is a function that is more urgently needed in on-site production. Compared with fault diagnosis, on-site personnel are more concerned about the state of the unit's continued operation and hope to predict future trends. Therefore, it is more important for enterprises to understand the future development trend of the equipment. Trend prediction technology can understand the future development trend of the equipment based on the analysis of its current operating status, take some preventive measures to avoid defects before the equipment occurs, and improve the safety, stability and reliability of equipment operation. In addition, with the help of trend prediction technology, it is also conducive to timely adjustment of the operation mode and improve the economy of the unit. The state detection and trend prediction of the unit are inseparable from the maintenance of the unit. Improving the availability of the unit to reduce the cost of power generation is one of the main measures to achieve this goal. Therefore, the research on trend prediction technology has important scientific significance and research value.

[0004] With the development of science and technology, there are more and more trend prediction methods, and the prediction accuracy is getting higher and higher. However, each trend prediction method has its advantages and disadvantages. CatBoost is a machine learning library open sourced by the Russian search giant in 2017. It is also a type of Boosting family algorithm. It is still an improved implementation under the GBDT algorithm framework. It is a GBDT framework based on the symmetric decision tree algorithm with few parameters, support for categorical variables and high accuracy. CatBoost handles gradient deviation and prediction offset problems to improve the accuracy and generalization ability of the algorithm.

[0005] CatBoost algorithm has obvious advantages over other network models. First, it can avoid the slow training speed of artificial neural network models and the need for manual settings and a large number of hyperparameter adjustments during the training process. Secondly, CatBoost algorithm can also obtain very good results using default parameters, reducing the time spent on parameter adjustment. At the same time, it can be based on GPU gradient boosting algorithms (such as Figure 3CatBoost, like all standard gradient boosting algorithms, builds a new tree to fit the gradient of the current model. However, all classic boosting algorithms suffer from overfitting problems caused by biased pointwise gradient estimates. Many algorithms that use GBDT technology (e.g., XGBoost, LightGBM) build a tree in two stages: selecting a tree structure and calculating the values ​​of leaf nodes after the tree structure is fixed. In order to select the best tree structure, the algorithm enumerates different splits, builds trees with these splits, calculates the values ​​in the obtained leaf nodes, then scores the obtained trees, and finally selects the best split. The values ​​of leaf nodes in both stages are calculated as approximations of gradients or Newton step sizes. CatBoost uses an unbiased estimate of the gradient step size in the first stage, and the second stage is performed using the traditional GBDT scheme.

[0006] In thermal power generation equipment, steam turbines are important equipment and need to be maintained in a timely manner. The impact of steam turbine back pressure on steam turbines is particularly important. However, the monitoring of back pressure in power plants can only be carried out in real time to obtain back pressure data. Once the back pressure exceeds expectations, the steam turbine may have been damaged. If the change of steam turbine back pressure can be predicted in advance, the turbine can be shut down for maintenance before the back pressure changes, which can avoid serious damage to the steam turbine. However, there is almost no relevant technology for predicting steam turbine back pressure in the prior art. Summary of the invention

[0007] In order to solve the deficiencies in the prior art, a method for predicting the back pressure trend of a steam turbine is provided. The technical solution adopted by the present invention is:

[0008] A method for predicting a steam turbine back pressure trend based on a catboost algorithm comprises the following steps:

[0009] S1. Prepare historical operation data of steam turbine related to back pressure;

[0010] S2. Preprocessing the above historical operation data;

[0011] S3, reconstructing each pre-processed historical operation data into multiple features through feature engineering dimensionality increase processing;

[0012] S4. Analyze the resampled historical operation data and select the variables with a greater correlation with the turbine back pressure;

[0013] S5. Determine the adjustable parameters of the model, which include loss function, maximum number of trees, learning rate, regularization parameter, and sample rate. Use catboost to build a model and use the filtered historical operation data to predict the back pressure in a certain period of time in the future through the model.

[0014] S6. Deploy the offline trained model and start the scheduled task to train the model. When the trained model is inconsistent with the current running model, replace the running model.

[0015] Furthermore, the historical operation data of the steam turbine related to the back pressure in step S1 include: unit load, ambient temperature, fan power, feed water flow, cold section reheat mother pipe steam pressure, main steam valve front steam pressure, and steam turbine back pressure.

[0016] Furthermore, the preprocessing of step S2 includes missing values, outliers, noise, and normalization and resampling processing.

[0017] Furthermore, the missing value processing adopts Lagrange interpolation, the outlier processing adopts Raifudat's law, and the noise processing adopts quintic cubic smoothing filtering.

[0018] Furthermore, the adjustable parameters in step S6 include loss function, maximum number of trees, learning rate, regularization parameter, and sample rate.

[0019] By adopting the above method, the back pressure of the steam turbine can be predicted, so as to know the trend change of the back pressure of the steam turbine in advance, and provide a reference for the power plant operators. The scheme of the present invention is simple, operable, easy to implement by computer program, and easy to deploy on site. It can effectively predict the development trend of the back pressure of the steam turbine, and has great promotion and utilization value.

[0020] Compared with other neural network algorithms, the modeling using the CatBoost algorithm shows better robustness and has faster training and prediction speeds. The predicted back pressure trend can provide guidance for power plant operators.

[0021] Continuously training the prepared models during model prediction can timely replace inaccurate models and improve prediction accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of the method for predicting turbine back pressure trend based on the machine learning model in the present invention.

[0023] Figure 2 This is a flow chart of the turbine back pressure trend prediction structure in the present invention.

[0024] Figure 3 This is a flow chart of the basic principle of the gradient decision tree in the present invention.

[0025] Figure 4 The following is a back pressure trend prediction diagram of a 600MW unit in a power plant at a certain time in the embodiment. DETAILED DESCRIPTION

[0026] In order to better understand the technical solution of the present invention, the technical solution is further described in detail below:

[0027] This prediction method is based on the catboost algorithm to predict the trend of turbine back pressure. Figure 1 The following steps are involved:

[0028] S1. Prepare the historical operation data of the steam turbine related to back pressure; these historical operation data include unit load, ambient temperature, fan power, feed water flow, cold section reheat mother pipe steam pressure, main steam valve front steam pressure, steam turbine back pressure and other data, all of which are historical record data collected in real time by sensors.

[0029] S2. Preprocess the above historical operation data; preprocessing includes missing values, outliers, noise and normalization processing. Preprocessing is used to eliminate values ​​that have no reference value in the data, and the preprocessed data is resampled (historical data of different variables taken at fixed time intervals must be interpolated to complete the data, i.e. resampled, and the resampling time interval is 1 minute), and it is ensured that the processed data is at least greater than 60,000.

[0030] S3. Reconstruct the pre-processed historical operation data into multiple features through feature engineering dimensionality increase processing.

[0031] For example: the historical operating data are reconstructed into 5 features through feature engineering and dimensionality upgrading. The number of screened historical operating data is S, and the number of input variables of the trend prediction model is 5×S.

[0032] S4. Analyze the resampled historical operation data and select the variables with a greater correlation with turbine back pressure.

[0033] S5. Determine the adjustable parameters of the model, including loss function, maximum number of trees, learning rate, regularization parameter, and sample rate, and use catboost for modeling.

[0034] The back pressure in a certain period of time in the future is predicted by using the data that has been screened out and processed by dimensionality upgrade.

[0035] Taking the prediction of back pressure in the next 30 minutes as an example, assuming that the screened auxiliary variables are {Q1, Q2, …, QS}, and the resampled data totals N, the constructed input training data is shown in the figure:

[0036]

[0037] Among them, Q11, Q12, Q13, Q14, and Q15 represent auxiliary variables that are upgraded to 5 features after feature engineering of auxiliary variable Q1. Similarly, QS1, QS2, QS3, QS4, and QS5 represent auxiliary variables that are upgraded to 5 features after feature engineering of auxiliary variable QS. Q1-1 and Q1-2 represent the first and second data of auxiliary variable Q1, respectively, and so on.

[0038] The output training data after the back pressure historical operation data is reconstructed into 30 features through feature engineering dimensionality processing is shown in the figure:

[0039]

[0040] Among them, backpressure 1, backpressure 2, backpressure 3...backpressure 30 represent variables after backpressure feature engineering. Similarly, QS1, QS2, QS3, QS4, QS5 represent auxiliary variables after auxiliary variable QS feature engineering. Backpressure-6 and backpressure-7 represent the sixth and seventh data of backpressure respectively, and so on.

[0041] Referring to the two sets of tables above, when the input training data and the data of back pressure 1 are used as the training data of the model, the prediction model of the steam turbine back pressure in the first minute in the future will be obtained. Similarly, the prediction models of the steam turbine back pressure in the 2nd to 30th minutes in the future will be obtained respectively. For example, if the prediction model of the steam turbine back pressure in the next 30 minutes is needed, the input training model and the data of back pressure 30 will be used as the model training data.

[0042] S6. Reference Figure 2 Deploy the offline trained model and start the scheduled task to train the model. When the trained model is inconsistent with the current running model, replace the running model.

[0043] The following is a further detailed description of data selection, preprocessing and data screening methods:

[0044] The selected historical operation data S = {k1, k2, ..., k n ,b}, where k1,k2,…,k n ,b is the positive ordering of multidimensional vectors over time;

[0045] When dealing with outliers, we first calculate the standard deviation σ of the data, and then eliminate the data with a deviation greater than 3σ according to statistical theory.

[0046] The formula for dealing with data noise is as follows:

[0047]

[0048]

[0049]

[0050]

[0051]

[0052] Where: y k is the data that needs to be processed currently, y k-1 ,y k-2 The two data before the current data, y k+1 ,y k+2 are the two data before the current data. The formula requires that the number of nodes k is greater than or equal to 5. When the number of nodes is greater than 5, the two ends of the data to be processed use formula 1, formula 2, formula 4, and formula 5 respectively, and the rest of the data to be processed all use formula 3 for data processing to achieve the purpose of symmetry.

[0053] The data normalization formula is as follows:

[0054]

[0055] Where: x n is the normalized data, x0 is the original data, x max is the maximum value in the sample value range, x min The minimum value in the sample range.

[0056] To perform principal component analysis, follow these steps:

[0057] For principal component analysis using partial least squares regression, first set m independent variables x1, x2, ... x m and p dependent variables y1, y2, ...y p , extract the first component u1 from the independent variable set, and also extract the first component v1 from the dependent variable set, and then establish the dependent variables y1, y2, ... y p If the fitting result reaches the expected value, the algorithm is terminated. Otherwise, the second component is extracted until sufficient accuracy is obtained. When extracting these two components, there are two requirements for regression analysis:

[0058] (1) The components extracted from the variable set should carry as much information as possible from their respective sets.

[0059] (2) The correlation between v1 and u1 can reach the maximum.

[0060] Assume there are p dependent variables y1, y2, ...y p and m independent variables x1, x2, ... x mAll are standardized variables, and the n-time standardized data matrix of the independent variable combination and dependent variable group is as follows:

[0061]

[0062] The specific steps of partial least squares regression analysis are as follows:

[0063] (1) Extract the first components u1 and v1 of the two variable groups respectively, where u1 is the independent variable set X = [x1,…,x m ] T A linear combination of:

[0064] u1=α 11 x1+…+α 1m x m =ρ (1)T X

[0065] v1 is the set of independent variables Y = [y1,…,y p ] T A linear combination of:

[0066] v1=β 11 y1+…+β 1p y p =γ (1)T Y

[0067] The score vector for the first pair of components can be calculated from the standard matrices A and B for the two sets of variables:

[0068]

[0069]

[0070] The covariance Cov(u1, v1) of u1 and v1 can be calculated based on the inner product of the score vectors of the first pair of components, so the above two requirements can be converted into an extreme value problem with mathematical conditions:

[0071]

[0072]

[0073] The solution to this problem is to use the Lagrange multiplier method, and the problem will be transformed into finding the unit vector ρ (1) and γ (1) ,make Reach the maximum, so by calculating the m*m matrix M=A T BB T The eigenvectors and eigenvalues ​​of A, and the largest eigenvalue of M is The corresponding unit eigenvector is the desired solution ρ (1) , while γ (1)It can be obtained by (1) Calculated.

[0074]

[0075] (2) Construct y1, y2, ...y p Regression on u1 and x1, x2, ... x m Regression on u1. Assume the regression model is:

[0076]

[0077] where σ (1) =[σ 11 ,…,σ 1m ] T , τ (1) =[τ 11 ,…,τ 1p ] T are the parameter vectors in the many-to-one regression model, and A1 and B1 are the residual matrices. The regression coefficient vector σ (1) , τ (1) The least squares estimate of is:

[0078]

[0079] (3) Repeat the above steps with the residual matrix A1 and B1 instead of A and B. If the absolute value of the elements in the matrix B1 is close to 0, it is considered that the regression equation established with the first component has met the requirements and the extraction of components can be stopped. Otherwise, repeat the above steps with the matrix A1 and B1 instead of A and B to obtain:

[0080] ρ (2) =[α 21 ,…,α 2m ] T

[0081] γ (2) =[β 21 ,…,β 2p ] T

[0082] and is the component vector of the second pair of components The vectors of the second pair of components of X and Y are:

[0083]

[0084] (4) Let the rank of the n*m ​​data matrix A be r, then there are r components u1, u2, ..., u r , so that:

[0085]

[0086] put u k =α k1 x1+…+α km x m (k=1,2,……,r) Substitute into Y=u1τ (1) +…+u r τ (r) , that is, the partial least squares regression equation for p independent variables:

[0087] y j =c j1 x1+…+c jm x m

[0088] For regression modeling of single or multiple independent variables, it is not necessary to use all components for regression modeling. So far, in actual analysis and experience, it can be known that the first principal component is the most important and can obtain a lot of information from the original data. As the number of principal components increases, the importance of the components decreases. Generally speaking, the partial least squares method does not need to select the existing r components u1, u2, ..., u r To build a regression equation, but like principal component analysis, only the first L components are selected to obtain a model with good predictive ability. The number of components L extracted for modeling can be determined by cross-validation test. Each time the i-th observation data is discarded, the remaining n-1 observation data are modeled by partial least squares regression method, and the regression equation fitted after extracting h components is then substituted into the fitted regression equation to obtain y j The predicted value at the i-th observation point

[0089] Repeat the above verification for i = 1, 2, 3...n, that is, when h components are extracted, the jth dependent variable y j The sum of squares of the prediction errors is:

[0090]

[0091] Y=[y1,…,y p ] T The sum of squares of the prediction errors is:

[0092]

[0093] In addition, all sample points are used to fit the regression equation containing h components. At this time, the predicted value of the i-th sample point is Then we can define y j The sum of squared errors is:

[0094]

[0095] The sum of squared errors is defined as:

[0096]

[0097] When PRESS(h) reaches the minimum value, the corresponding h is the required number of components l. Usually, PRESS(h) is greater than SS(h), while SS(h) is less than SS(h-1). Therefore, when extracting components, we always hope that the ratio PRESS(h) / SS(h-1) is as small as possible, and the limit value is generally set to 0.05.

[26] ,Right now:

[0098] PRESS(h) / SS(h-1)≤(1-0.05) 2 =0.95 2

[0099] For this purpose, cross validity is defined as:

[0100]

[0101] In this way, a cross-validation test is performed before each modeling step. Then the model accuracy meets the requirements and component extraction can be stopped.

[0102] After using the above prediction model to predict the turbine back pressure, the predicted back pressure trend change is compared with the actual back pressure trend change. Figure 4 As shown in the figure, the trend of back pressure prediction is basically consistent with the actual back pressure trend, which can fully meet the requirements of predicted back pressure.

Claims

1. A method for predicting the back pressure trend of a steam turbine based on a catboost algorithm, comprising the following steps: S1. Prepare historical operation data of the steam turbine related to back pressure, wherein the historical operation data includes unit load, ambient temperature, fan power, feed water flow, cold section reheat mother pipe steam pressure, main steam valve front steam pressure, and steam turbine back pressure; S2. Preprocessing the above historical operation data; S3, reconstructing each pre-processed historical operation data into multiple features through feature engineering dimensionality increase processing; S4. Analyze the resampled historical operation data and select the variables with a greater correlation with the turbine back pressure; S5. Determine the model's adjustable parameters, including loss function, maximum number of trees, learning rate, regularization parameter, and sample rate. Use catboost to build a model and use the filtered historical operation data to predict the back pressure in a certain period of time in the future through the model. S6. Deploy the offline trained model and start the scheduled task to train the model. When the trained model is inconsistent with the current running model, replace the running model.

2. The method for predicting a steam turbine back pressure trend based on a catboost algorithm according to claim 1, characterized in that: The preprocessing of step S2 includes missing values, outliers, noise, and normalization and resampling processing.

3. The method for predicting the back pressure trend of a steam turbine based on the catboost algorithm according to claim 2, characterized in that: The missing value processing adopts Lagrange interpolation, the outlier processing adopts Raifudat's law, and the noise processing adopts five-point cubic smoothing filter.

Images