An online monitoring method for nitrogen oxide emission concentration of coal-fired units based on transfer learning

Abstract

The present invention belongs to the field of electric power, and specifically is an online monitoring method for nitrogen oxide emission concentration of coal-fired units based on transfer learning. The method includes the following steps: S100: collecting historical operation data of the unit as sample data, and preprocessing the sample data; S200: constructing a labeled source domain data set and an unlabeled target domain data set based on the sample data; S300: reading the training data sets of the source domain and the target domain, pre-training the CNN‑RBFNN model, and obtaining a nitrogen oxide emission concentration prediction model; S400: using a test set to verify the effectiveness of the nitrogen oxide emission concentration prediction model; S500: putting it into use online. The present invention implements online soft measurement of nitrogen oxide concentration of coal-fired units based on transfer learning, overcomes the problem of model mismatch of traditional soft measurement models caused by frequent peak regulation of units, and further improves the measurement accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of electric power, and specifically is an online monitoring method for nitrogen oxide emission concentration of a coal-fired unit based on transfer learning. Background Art

[0002] my country's energy structure has changed. The installed capacity of new energy sources such as photovoltaic and wind power has continued to increase. Thermal power units frequently participate in the deep peak regulation of the power grid, which puts higher requirements on the energy saving and intelligent monitoring technology of the units. As an important environmental protection indicator, the nitrogen oxide emission concentration of the unit is of great significance for real-time and accurate intelligent monitoring.

[0003] At present, online methods for measuring nitrogen oxide concentration emissions from thermal power units are mainly divided into two categories: one is various flue gas measuring instruments that rely on hardware, and the other is soft measurement systems that are independent of hardware.

[0004] With the continuous improvement of sensor technology, the measurement accuracy of flue gas measuring instruments is also constantly improving. These measuring devices use electrochemical principles, infrared, microwave and other technologies to measure NO in flue gas. X Direct measurement, such as the currently widely used CEMS system, SCR system, and SNCR system. Hardware measurement devices have high measurement accuracy in the early stage of use, but such devices are generally installed near the flue at the tail of the boiler, with a harsh working environment. When they run for a long time without timely maintenance, the credibility of the measurement results is significantly reduced.

[0005] The other type is a soft measurement system that is independent of hardware. This type of soft measurement system is based on either mechanism analysis or data-driven artificial intelligence. The measurement method based on mechanism analysis uses basic thermodynamic physics theorems, combined with the laws of element conservation and mass conservation, to simulate NO with operating parameters and boiler structural parameters as boundary conditions. X Emissions, this method simplifies the coal combustion process and uses a large number of empirical formulas, and the measurement results are difficult to meet the requirements; data-driven artificial intelligence measurement methods have gained wider attention. This method does not require a detailed mechanism process when modeling, and mostly uses the operating parameters that affect the generation of nitrogen oxides as the input of the model. Based on a large amount of high-quality historical operation data of the unit, the model is established using artificial intelligence methods to predict NO in real time. XGenerated concentration. The popularity of distributed control systems and plant-level information systems in my country's thermal power units provides strong data support for the promotion of soft measurement technology. There are mainly autoregressive linear prediction models, artificial neural network prediction models, support vector machine prediction models, etc. These models have higher measurement accuracy than mechanism models, but they are shallow machine learning methods with limited data feature extraction capabilities. The prediction model based on deep learning methods further reduces the prediction error of nitrogen oxide emission concentration. With the continuous improvement of basic theoretical research and the advancement of modeling methods, the measurement accuracy of soft measurement methods is constantly improving, and the actual investment cost is extremely small. At present, the soft measurement method has the following problems: the operation of thermal power units under conventional load (50%-100% BMCR) is considered in the design. With the deep peak regulation of thermal power units, low load, especially ultra-low load operation, is becoming more and more frequent, typical conditions and atypical conditions appear alternately, and multimodal characteristics have become a common feature in the current unit production process. When the unit operates under atypical conditions, due to the different data distribution from typical conditions, the distribution of real-time data and modeling sample data is quite different. The traditional machine learning effect may be very poor, resulting in mismatch of the soft measurement model and reduced measurement accuracy. Summary of the invention

[0006] In order to solve the above problems, the present invention provides a method for online monitoring of nitrogen oxide emission concentration of coal-fired units based on transfer learning. When the unit operates under atypical conditions, the combustion knowledge under typical conditions is extracted and transferred to atypical conditions to obtain a soft measurement model of nitrogen oxide emission of the unit suitable for atypical conditions.

[0007] The present invention adopts the following technical scheme: an online monitoring method for nitrogen oxide emission concentration of a coal-fired unit based on transfer learning, comprising the following steps:

[0008] S100: Collect historical operation data of the unit as sample data, and pre-process the sample data;

[0009] S200: Construct a labeled source domain dataset and an unlabeled target domain dataset based on the sample data. The source domain dataset consists of data under typical operating conditions of the unit, and the target domain dataset consists of a small amount of data under atypical operating conditions. 70% of the samples from the source domain dataset and the target domain dataset are extracted as training sets, and the remaining 30% are used as test sets.

[0010] S300: Read the training data sets of the source domain and the target domain, pre-train the CNN-RBFNN model, and obtain a nitrogen oxide emission concentration prediction model;

[0011] S400: Use the test set to verify the effectiveness of the nitrogen oxide emission concentration prediction model;

[0012] S500: put into use online, extract the real-time values ​​of the model input parameters from the power plant SIS system in real time, use fuzzy C-means clustering to determine whether the current operating condition belongs to an atypical operating condition. If it belongs to a conventional operating condition, it is input into the conventional prediction model for soft measurement of nitrogen oxide emission concentration. If it belongs to an atypical operating condition, it is input into the model built by the present invention, and the nitrogen oxide emission concentration at the current moment is output in real time to complete the prediction.

[0013] Furthermore, the preprocessing of the sample data in step S100 includes data cleaning and normalization. Data cleaning is to remove bad pixels in the data and delete missing data and erroneous data.

[0014] Furthermore, in step S200, the source domain dataset is: ; The target domain dataset is ; where x represents the matrix or vector composed of input parameters, and y represents the corresponding nitrogen oxide emission concentration value, , Respectively represent the number of sample data contained in the source domain dataset and the target domain dataset.

[0015] Furthermore, the specific process of step S300 is as follows:

[0016] S301: Initialize model parameters;

[0017] S302: Forward propagation: The source domain dataset and the target domain dataset are respectively input into the CNN convolutional neural network, and after the convolution layer and the pooling layer, the output of the pooling layer is input into the RBF neural network to construct the network parameter optimization objective function, and the function value of the optimization objective function is calculated using the RBF neural network output.

[0018] S303: Back propagation: The present invention uses the optimization objective function as the loss function, and uses the Adam optimization algorithm to propagate it backward layer by layer to update the model weights to achieve knowledge transfer.

[0019] The specific process of step S302 is:

[0020] Calculate the function value of the optimization objective function, the optimization objective function is:

[0021]

[0022] Where N represents the number of samples of source domain data, Indicates i The label value of each sample, Indicates i The predicted value of samples, represents the fully connected layer output of the source domain data, Represents the fully connected layer output of the target domain data.

[0023] The specific process of step S303 is:

[0024] The first step of the back propagation training process is to calculate the total error of the network. The present invention uses the optimized objective function as the loss function, and returns the function value layer by layer to update the weights.

[0025] Furthermore, the specific process of step S400 is to input the target domain training set data into the pre-trained nitrogen oxide emission concentration model for validity verification, and obtain the predicted nitrogen oxide emission concentration for the input data under different modes, and compare it with the actual measured emission value. If the requirements are met, it is put into actual use; if the requirements are not met, the model is retrained and the model parameters are adjusted.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1 The present invention implements online soft measurement of nitrogen oxide concentration in coal-fired units based on transfer learning, overcomes the problem of model mismatch of traditional soft measurement models caused by frequent peak regulation of units, and further improves the measurement accuracy.

[0028] 2 The device of the present invention has a simple structure, does not require additional measuring devices, has low investment cost and high operating reliability, and can bring direct economic benefits to the power plant. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A model for predicting nitrogen oxide emission concentration;

[0030] Figure 2 Flowchart of the soft sensing method for nitrogen oxide emission concentration. DETAILED DESCRIPTION

[0031] The research on online measurement of pulverized coal flow parameters is divided into two stages: offline and online. In the offline stage, the model input parameters are determined according to the nitrogen oxide generation mechanism, and the effective sample data set is constructed using the historical operation data of the unit, and the prediction model is established based on the transfer learning method; in the online stage, the input parameter operation values ​​are collected in real time and input into the trained prediction model, and the real-time value of nitrogen oxide emission concentration can be output in real time, which helps the on-site staff to understand the combustion status in real time and guide the operation. Through the analysis of the generation mechanism of nitrogen oxides in coal-fired units, the input parameters used in the model are determined as follows: coal feed of each pulverizer, primary air volume of each pulverizer, air-powder mixture temperature at the outlet of each pulverizer, secondary air volume of the furnace, primary and secondary air temperatures, pulverized coal particle size, burnout air rate, and unit load.

[0032] The present invention introduces transfer learning to perform soft measurement on the nitrogen oxide emission concentration of coal-fired units. The nitrogen oxide emission concentration soft measurement model proposed by the present invention is composed of a convolutional neural network (CNN) and a regressor, and the present invention uses a radial basis function neural network (RBFNN) as a regressor.

[0033] A method for online monitoring of nitrogen oxide emission concentration of a coal-fired unit based on transfer learning comprises the following steps:

[0034] S100: Collect historical operation data of the unit as sample data, and pre-process the sample data.

[0035] The DCS and SIS systems of power plants store a wealth of historical and real-time operating data of the units. Model input parameters and NOX-related operating data are selected for preprocessing, including data cleaning and normalization. Data cleaning mainly removes bad points in the data, deletes missing data and erroneous data. After preprocessing, a valid sample data set is obtained.

[0036] S200: Construct a labeled source domain dataset and an unlabeled target domain dataset based on the sample data. The source domain dataset is: ; The target domain dataset is ; where x represents the matrix or vector composed of input parameters, and y represents the corresponding nitrogen oxide emission concentration value, , Respectively represent the number of sample data contained in the source domain dataset and the target domain dataset. The source domain dataset is composed of data under typical operating conditions of the unit, and the target domain dataset is composed of a small amount of data under atypical operating conditions. 70% of the samples of the source domain dataset and the target domain dataset are extracted as training sets, and the remaining 30% are used as test sets.

[0037] The source domain data set is composed of the operating data of the unit under typical operating conditions, and the target domain data set is composed of a small amount of atypical operating condition data. The purpose of the present invention is to extract the combustion knowledge under typical operating conditions and transfer it to the atypical operating condition model to enhance the adaptability of the model to different operating conditions.

[0038] S300: Read the training data sets of the source domain and the target domain, pre-train the CNN-RBFNN model, and obtain a nitrogen oxide emission concentration prediction model.

[0039] The specific process of step S300 is:

[0040] S301: Initialize model parameters.

[0041] S302: Forward propagation: The source domain dataset and the target domain dataset are respectively input into the CNN convolutional neural network, and after the convolution layer and the pooling layer, the output of the pooling layer is input into the RBF neural network to construct the network parameter optimization objective function, and the function value of the optimization objective function is calculated using the RBF neural network output.

[0042] CNN is a deep learning method with strong feature extraction capability. The convolutional neural network model mainly consists of four parts: data input layer, convolution layer, pooling layer, and fully connected layer.

[0043] The data input layer receives the input matrix, and the convolution layer extracts the input data features, which is the key to the convolutional neural network. The convolution layer contains multiple convolution kernels and two key operations: local association and window sliding. The specific operation method of the convolution layer is as follows:

[0044] (1)

[0045] Where: For the l Tier i convolution kernel weight matrix; For the l- 1 layer output; For the l The output of the layer i features; “*” is the convolution operator; is the bias term.

[0046] The pooling layer is usually located between two adjacent convolutional layers, or connects the convolutional layer and the fully connected layer. The pooling layer can effectively reduce the size of the parameter matrix, reduce the number of parameters in the fully connected layer, speed up the network calculation speed and prevent overfitting. The operator of the pooling layer is called the pooling kernel, which scans the feature matrix output by the convolutional layer in a traversal manner. The specific calculation method is formula (2):

[0047] (2)

[0048] Where: For the l Tier i The elements of the feature matrix in the pooling kernel area; After pooling l +1 level i elements in the feature matrix; D j For the j The area covered by the pooling kernel.

[0049] After the convolution layer and the pooling layer, the output of the pooling layer is input to the radial basis neural network. The RBF network is a feedforward neural network with three layers of neurons, including an input layer, a hidden layer, and an output layer. The present invention uses a Gaussian function as the kernel function of the hidden layer, and the output of the jth neuron in the hidden layer is h j for:

[0050] (3)

[0051] in, The hidden layer j The coordinate vector of the center point of the Gaussian basis function of each neuron; The hidden layer j The width of the Gaussian basis function of each neuron. x is the output of the convolutional network.

[0052] Assume that the weight matrix between the hidden layer and the output layer is w , then the output of the RBF neural network y That is the predicted value of nitrogen oxide emission concentration, which can be expressed by formula (4):

[0053] (4)

[0054] In order to improve the prediction accuracy of the model under multiple working conditions, this paper proposes a network parameter fine-tuning method for predicting the nitrogen oxide emission concentration of coal-fired units based on transfer learning, thereby realizing the knowledge transfer of combustion states under multi-modal conditions. The established coal-fired unit nitrogen oxide emission concentration prediction model is as follows Figure 1 When optimizing network parameters, there are two main goals: the first goal is to minimize the prediction error of nitrogen oxide emission concentration, that is, to minimize the prediction loss; the second goal is to minimize the feature distribution difference distance between the source domain and the target domain under a given source domain, and keep the common knowledge part unchanged as much as possible. Based on the above two optimization goals, the following optimization objective function is established:

[0055] 1) Minimize prediction error

[0056] By minimizing the absolute value of the error between the predicted value and the measured value, the objective function is shown in formula (5):

[0057] (5)

[0058] Where N represents the number of samples of source domain data. Indicates i The label value of each sample, Indicates i The predicted value of a sample.

[0059] 2) Minimize the difference in feature distribution between the source domain and the target domain

[0060] Minimizing the output of the fully connected layer of the source domain and the target domain means minimizing the difference distance of the feature distribution between the source domain and the target domain. The objective function is shown in formula (6):

[0061] (6)

[0062] in, represents the fully connected layer output of the source domain data, Represents the fully connected layer output of the target domain data.

[0063] Combining the two objectives, we get the following optimization objective function to fine-tune the structural parameters of the network:

[0064] (7)

[0065] Calculate the function value of the optimization objective function: where N represents the number of samples of the source domain data, Indicates i The label value of each sample, Indicates i The predicted value of samples, represents the fully connected layer output of the source domain data, Represents the fully connected layer output of the target domain data.

[0066] S303: Back propagation: The present invention uses the optimized objective function as the loss function, and uses the following method to propagate it backward layer by layer to update the model weights to achieve knowledge transfer.

[0067] The principle of updating network weights in the present invention is:

[0068]

[0069]

[0070]

[0071]

[0072]

[0073]

[0074] in, and are the first-order momentum term and the second-order momentum term, respectively. and are the exponential decay rates of the first-order moment estimate and the second-order moment estimate, respectively. and are their respective correction values. represents the weight matrix for t iterations, Represents the gradient size of the loss function for the weight matrix at the tth iteration, is the learning rate, It is a number close to 0, avoiding the denominator being 0.

[0075] S400: Verify the effectiveness of the nitrogen oxide emission concentration prediction model; input the target domain training set data into the pre-trained nitrogen oxide emission concentration model for effectiveness verification, obtain the predicted nitrogen oxide emission concentration for the input data under different modes, and compare it with the actual measured emission value. If it meets the requirements, it will be put into actual use. If it does not meet the requirements, retrain the model and adjust the model parameters.

[0076] S500: put into use online, extract the real-time values ​​of the model input parameters from the power plant SIS system in real time, use fuzzy C-means clustering to determine whether the current operating condition belongs to an atypical operating condition. If it belongs to a conventional operating condition, it is input into the conventional prediction model for soft measurement of nitrogen oxide emission concentration. If it belongs to an atypical operating condition, it is input into the model built by the present invention, and the nitrogen oxide emission concentration at the current moment is output in real time to complete the prediction.

[0077] In view of the problem that when the unit operates under atypical conditions, the distribution of operating data is quite different from that of data under conventional operating conditions, the traditional soft measurement model has model mismatch and the measurement accuracy is reduced, this application proposes to use transfer learning to learn knowledge from conventional operating conditions and transfer this knowledge to atypical conditions, so as to construct a nitrogen oxide concentration prediction model suitable for the unit's full operating conditions.

Claims

1. A method for online monitoring of nitrogen oxide emission concentration of coal-fired units based on transfer learning, characterized in that: The following steps are included: S100: Collect historical operation data of the unit as sample data, and pre-process the sample data; S200: Construct a labeled source domain dataset and an unlabeled target domain dataset based on the sample data. The source domain dataset consists of data under typical operating conditions of the unit, and the target domain dataset consists of a small amount of data under atypical operating conditions. 70% of the samples from the source domain dataset and the target domain dataset are extracted as training sets, and the remaining 30% are used as test sets. S300: Read the training data sets of the source domain and the target domain, pre-train the CNN-RBFNN model, and obtain a nitrogen oxide emission concentration prediction model; The specific process of step S300 is: S301: Initialize model parameters; S302: Forward propagation: Input the source domain dataset and the target domain dataset into the CNN convolutional neural network respectively, and after the convolution layer and the pooling layer, input the output of the pooling layer into the RBF neural network, construct the network parameter optimization objective function, and use the RBF neural network output to calculate the function value of the optimization objective function; S303: Back propagation: The optimization objective function is used as the loss function, and the Adam optimization algorithm is used to propagate the optimization function backward layer by layer to update the model weights to achieve knowledge transfer. S400: Use the test set to verify the effectiveness of the nitrogen oxide emission concentration prediction model; S500: It is put into use online and extracts the real-time values ​​of the model input parameters from the power plant SIS system in real time. It uses fuzzy C-means clustering to determine whether the current operating condition is an atypical condition. If it is a conventional condition, it is input into the conventional prediction model for soft measurement of nitrogen oxide emission concentration. If it is an atypical condition, it is input into the established model and outputs the current nitrogen oxide emission concentration in real time to complete the prediction.

2. The method for online monitoring of nitrogen oxide emission concentration of coal-fired units based on transfer learning according to claim 1 is characterized in that: The preprocessing of the sample data in step S100 includes data cleaning and normalization. Data cleaning is to remove bad pixels in the data and delete missing data and erroneous data.

3. The method for online monitoring of nitrogen oxide emission concentration of coal-fired units based on transfer learning according to claim 1 is characterized in that: In step S200, the source domain dataset is: ; The target domain dataset is ; where x represents the matrix or vector composed of input parameters, and y represents the corresponding nitrogen oxide emission concentration value, , Respectively represent the number of sample data contained in the source domain dataset and the target domain dataset.

4. The method for online monitoring of nitrogen oxide emission concentration of coal-fired units based on transfer learning according to claim 1 is characterized in that: The specific process of step S302 is: Calculate the function value of the optimization objective function, the optimization objective function is: Where N represents the number of samples of source domain data, Indicates i The label value of each sample, Indicates i The predicted value of samples, represents the fully connected layer output of the source domain data, Represents the fully connected layer output of the target domain data.

5. The method for online monitoring of nitrogen oxide emission concentration of coal-fired units based on transfer learning according to claim 1 is characterized in that: The specific process of step S303 is: The first step of the back-propagation training process is to calculate the total error of the network, use the optimized objective function as the loss function, return the function value layer by layer, and update the weights.

6. The method for online monitoring of nitrogen oxide emission concentration of coal-fired units based on transfer learning according to claim 1 is characterized in that: The specific process of step S400 is to input the target domain training set data into the pre-trained nitrogen oxide emission concentration model for validity verification, obtain the predicted nitrogen oxide emission concentration for the input data under different modes, and compare it with the actual measured emission value. If it meets the requirements, it is put into actual use. If it does not meet the requirements, the model is retrained and the model parameters are adjusted.

Images