A boiler performance prediction method based on operating condition division and neural network

Abstract

The invention discloses a boiler performance prediction method based on operating condition division and neural network, which relates to an online prediction method for coal-fired boiler performance, and solves the technical problem that the prediction of boiler performance under atypical operating conditions is not accurate enough and is prone to distortion. The key points of the technical solution are to collect boiler operation data and boiler performance indicators to obtain original data sets, the boiler performance indicators include boiler efficiency and NOx emission concentration, the boiler efficiency is indirectly obtained through a preset efficiency calculation module, and the NOx emission concentration is obtained by a continuous emission detector; abnormal data and non-steady-state data are eliminated through data preprocessing to obtain a static data set; secondly, the boiler operating condition division standard is obtained through a K-means clustering method, and the error square sum is used as an evaluation indicator to evaluate the effectiveness of the clustering method; finally, a data pool is established for each boiler operating condition, and a BP neural network model is established separately for each operating condition to achieve the purpose of accurately predicting boiler performance.

Description

Technical Field

[0001] The present application relates to an online prediction method for coal-fired boiler performance, and in particular to a boiler performance prediction method based on operating condition division and neural network. Background Art

[0002] The accurate establishment of the combustion model is an important step in achieving the optimization of coal-fired boiler combustion. Because there are still many difficulties in using mechanism modeling, intelligent algorithms such as neural network algorithms are currently used to directly establish the mapping model between boiler input parameters and output parameters. The processing of the data set required to establish the model from the original data collected from the power plant generally includes outlier removal and steady-state detection. The purpose of this processing is to screen out abnormal data caused by instrument failure and unstable data when the boiler changes operating conditions. However, boiler combustion is a complex process with multiple operating conditions. The above data processing method can meet the model establishment requirements when the data volume is small, but when the time span of the boiler performance to be predicted becomes larger, the data volume of the typical operating condition of the boiler will greatly exceed other operating conditions, making the model biased towards typical operating conditions in order to meet the accuracy requirements during training, which will cause the model to be distorted when predicting other operating conditions, so it is urgent to further process the data set used by the model. Summary of the invention

[0003] The present application provides a boiler performance prediction method based on operating condition division and neural network. Its technical purpose is to divide the data of each operating condition of the boiler and establish a neural network model according to the operating condition to achieve accurate prediction of boiler performance.

[0004] The above technical objectives of this application are achieved through the following technical solutions:

[0005] A boiler performance prediction method based on operating condition division and neural network, comprising:

[0006] S1: Obtain boiler operation data and boiler performance indicators, pre-process the operation data and performance indicators, and obtain a static data set for modeling to avoid the influence of boiler variable operating condition data.

[0007] Among them, boiler performance indicators include boiler efficiency and NOx emission concentration. Boiler efficiency is indirectly obtained through a preset efficiency calculation module, and NOx emission concentration is obtained by a continuous emission detector.

[0008] S2: Use the K-means clustering method to divide the static data set to obtain division results, and establish different data pools according to the division results.

[0009] S3: Establish BP neural network models for each operating condition in different data pools, and predict boiler performance through each BP neural network model.

[0010] Each data set is divided into training set and test set in a ratio of 7:3. The training set and test set are selected by random sampling to improve the generalization ability of the model.

[0011] In addition, it can be determined whether the output of each BP neural network model meets the accuracy requirement. If it does not meet the accuracy requirement, return to step S3 to adjust the BP neural network model of the corresponding working condition.

[0012] Furthermore, in step S1, the preprocessing method includes an outlier elimination method and a steady-state discrimination method;

[0013] The method of eliminating outliers includes defining parameter ranges based on practical experience or processing data through the 3σ principle;

[0014] The steady-state determination method is a sliding window method, which determines whether the fluctuation range of the data in the window meets the requirements of the steady-state operating conditions in the boiler test manual.

[0015] Furthermore, in step S2, the static data set is divided using a K-means clustering method, including:

[0016] Select K initial cluster centers based on the distance maximization principle;

[0017] Calculate the data D collected at time t i The sum of squared errors with the initial cluster center is used to obtain the partition result, which is expressed as:

[0018]

[0019] Among them, K represents the cluster center, D i =(d(1) t ,…,d(m) t ,…,d(M) t ,P),d(m) t represents the mth boiler operation data collected at time t, P represents the index of boiler operating condition classification, and M represents the total number of operation data; the static data set S is represented by S = {D1,…,D i ,…D n}.

[0020] The essential requirement of formula (1) is to solve the sum of squares of errors between various types of data after clustering and the cluster center. In the power plant's operating condition division, the sum of squares of errors between the performance index and the performance index of the cluster center is solved. In order to avoid the influence of the randomly selected initial cluster center on the clustering result, the principle of maximizing the distance between each other is introduced when initializing the cluster center. Through this principle, a large amount of power plant operation data and multiple iterations can obtain a stable cluster center.

[0021] The beneficial effects of the present application are: obtaining the original data set by collecting boiler operation data and boiler performance indicators, obtaining a static data set by eliminating abnormal data and non-steady-state data through data preprocessing, obtaining the boiler operating condition classification standard through the K-means clustering method, establishing a data pool for each boiler operating condition, and establishing a neural network model according to the operating condition to achieve the purpose of accurately predicting boiler performance.

[0022] This application performs separate modeling according to different operating conditions of the boiler. The model error is smaller than the error of the overall modeling and the accuracy is higher; and the trained network is closer to the actual operating conditions, which is conducive to the subsequent optimization of the boiler parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flowchart of the method described in this application;

[0024] Figure 2 It is a schematic diagram of the change of the sum of squared errors with the number of clusters;

[0025] Figure 3 It is a structural diagram of the BP neural network model;

[0026] Figure 4 Schematic diagram of model training for a specific embodiment. DETAILED DESCRIPTION

[0027] The technical solution of the present application will be described in detail below with reference to the accompanying drawings.

[0028] Figure 1 This is a flow chart of the method described in this application, and the specific process will not be repeated. This article uses the operating data of a 1000MW unit in a power station to illustrate the technical solution.

[0029] (1) Obtaining the original boiler operation data: Extract the data of the unit's continuous operation for 46 days, with a sampling frequency of 1 minute, totaling 66,240 data points. The outliers in the original data were eliminated by parameter range delineation and the 3σ principle, leaving 65,299 data points after elimination. The steady-state operating conditions were determined by the sliding window method, and a total of 30,720 data points were obtained.

[0030] (2) Use the K-means clustering method to partition the static data set, for example, select boiler load and relative coal quality coefficient as the partitioning index. The data set obtained by the power plant DCS system is S = {D1,…,D i ,…D n}, where D i =(d(1) t ,…,d(m) t ,…,d(M) t , P), d(k) trepresents the mth boiler operating parameter collected at time t, and P represents the boiler load and relative coal quality coefficient. K cluster centers are selected according to the distance maximization principle, and the distance between each operating parameter and the cluster center is calculated. The sum of square errors between the overall data set and the cluster center is calculated according to formula (1). Figure 2 The diagram shows the change of the calculated error sum of squares with the load and relative coal quality coefficient cluster number. For the calculation of the load cluster number, about 40% of the loads are excluded at the beginning, because this type of load is far away from other loads and can be obviously classified into one category. Finally, Figure 2 The points in the middle ellipse are taken as the final number of clusters, which takes into account the principle of small sum of square errors and prevents the division from being too detailed.

[0031] Finally, the working conditions are divided into 4 categories according to the load, and 3 categories according to the relative coal quality coefficient. The clustering process is as follows: Figure 2 As shown. Finally, the boiler operating conditions are divided into 12 categories, and the division results are shown in Table 1. The number of data samples occupied by each operating condition is different. The data samples are concentrated in the operating conditions where the load and coal quality match. For example, the number of samples of 40% load and low coal quality is the largest among the same load conditions, and the number of data samples of the operating conditions of 50% load and 70% load with medium and high coal quality is relatively large, and the number of data samples of the operating conditions of high working conditions and low coal quality is 0. The distribution of data samples is consistent with the actual production situation of the power plant.

[0032] Table 1 Sample classification

[0033] Dataset load Coal quality Number of samples 1 [376.77,412.54] (2.23,2.52) 745 2 [376.77,412.54] (2.52,3.07] 328 3 [376.77,412.54] [1.88,2.23] 6878 4 [500.27,588.30] (2.23,2.52) 5978 5 [500.27,588.30] (2,52,3.07] 1900 6 [500.27,588.30] [1.88,2.23] 482 7 [588.32,734.12] (2.23,2.52) 5571 8 [588.32,734.12] (2,52,3.07] 3166 9 [588.32,734.12] [1.88,2.23] 94 10 [734.15,1015.90] (2.23,2.52) 1008 11 [734.15,1015.90] (2,52,3.07] 4195 12 [734.15,1015.90] [1.88,2.23] 0

[0034] (3) A neural network model is established for each operating condition. The input of the model is selected from 40 physical parameters, including load, furnace oxygen content, coal powder temperature, primary and secondary air door opening, and the output is selected from NOx emission concentration. Therefore, the number of input neurons of the neural network model is 40, the output is 1, and the number of hidden layer neurons is selected as 12. The structure of the model is as follows: Figure 3 The training results are shown in Figure 4As shown in the figure, the horizontal axes 1-11 are the training set mean square error and the test set mean square error of each data set in Table 1 after individual training. The bar chart of data set 12 represents the training set mean square error and the test set mean square error of the global data set after modeling for different working conditions. The dotted horizontal line represents the test set mean square error during overall modeling, and the asterisk horizontal line represents the training set mean square error during overall modeling. It can be seen that except for some data sets (1, 2, 9) whose errors are higher than the errors during overall modeling due to the small number of samples, the training set error and the test set error are far apart due to the small amount of data in data set 9. The errors of other data sets (3, 4, 5, 7, 8, 10, 11) are close to or even lower than the errors during overall modeling, and the overall error after modeling for different working conditions is also lower than the error during overall modeling. This is because after modeling for different working conditions, the data samples are more closely connected, which is conducive to the training and learning of the network, so the generated network can better reflect the current working conditions.

[0035] The above are exemplary embodiments of the present application, and the protection scope of the present application is defined by the claims and their equivalents.

Claims

1. A boiler performance prediction method based on operating condition division and neural network, characterized in that: include: S1: Obtain boiler operation data and boiler performance indicators, pre-process the operation data and performance indicators, and obtain a static data set for modeling; wherein the boiler performance indicators include boiler efficiency and NOx emission concentration; S2: using a K-means clustering method to divide the static data set to obtain a division result, and establishing different data pools according to the division result; S3: Establish BP neural network models for each operating condition in different data pools, and predict boiler performance through each BP neural network model; Wherein, in the step S2, the static data set is divided using a K-means clustering method, including: Select K initial cluster centers based on the distance maximization principle; Calculate the data D collected at time t i The sum of squared errors with the initial cluster center is used to obtain the partition result, which is expressed as: Among them, K represents the cluster center, D i =(d(1) t ,…,d(m) t ,…,d(M) t ,P),d(m) t represents the mth boiler operation data collected at time t, P represents the index of boiler operating condition classification, and M represents the total number of operation data; the static data set S is represented by S = {D1,…,D i ,…D n }.

2. The prediction method according to claim 1, characterized in that: In step S1, the preprocessing method includes an outlier elimination method and a steady-state discrimination method; The method of eliminating outliers includes defining parameter ranges based on practical experience or processing data through the 3σ principle; The steady-state determination method is a sliding window method, which determines whether the fluctuation range of the data in the window meets the requirements of the steady-state operating conditions in the boiler test manual.

Images