Combustion device CO2 emission concentration prediction method based on CNN-LSTM model

Abstract

The invention provides a combustion device CO2 emission concentration prediction method based on a CNN-LSTM model, and the method comprises the following steps: screening an input variable data set according to a correlation coefficient between each input variable and the CO2 emission concentration; performing data conversion processing; constructing a CNN-LSTM model as an initial prediction model, inputting the training set into the initial prediction model for training until the average absolute error loss is less than 5%, and storing the model as a training model; changing hyper-parameters of the initial prediction model, performing multiple times of training, and selecting a training model with the minimum root mean square error and the maximum decision coefficient as a final prediction model; and inputting the relevant characteristic variable parameters of the current time step into the final prediction model to obtain a CO2 emission concentration prediction value of the next time step. The CO2 emission concentration of the combustion device can be accurately predicted in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of combustion product concentration prediction of combustion devices, and in particular to a method for predicting CO2 emission concentration of combustion devices based on a CNN-LSTM model. Background Art

[0002] In the process of promoting the construction of CCUS projects, it is extremely necessary to conduct a comprehensive investigation and evaluation of the carbon emission level of combustion devices. The online flue gas monitoring systems of most combustion devices are mainly used to monitor the emissions of pollutants such as nitrogen oxides, sulfur oxides and carbon monoxide, and lack direct monitoring of CO2 concentration. Establishing the connection between CO2 concentration and the operating parameters of the combustion device and predicting and controlling the changes in CO2 concentration can not only estimate the carbon emission level of the combustion device, but also provide support for the design of the control system of the carbon capture device.

[0003] At present, the existing technology for predicting CO2 emissions mainly focuses on predicting carbon emissions of a certain industry or group in my country from a macro level, and there are few studies on CO2 emissions from the perspective of industrial application and operating conditions. Since the combustion device is a complex system with multiple factors, the CO2 emission concentration of the combustion device is affected by multiple variables, which makes it difficult to accurately predict the CO2 emission concentration of the combustion device.

[0004] At present, the CO2 emission concentration is roughly estimated by relying only on the flue gas CO emission data and fuel data at the combustion device site, which makes it difficult to directly and accurately estimate the combustion status and carbon emission level of the combustion device. Therefore, it is urgent to propose a CO2 emission concentration prediction method based on the operating conditions of the combustion device, in order to provide theoretical support for the low-energy optimization operation of CCUS devices. Summary of the invention

[0005] Based on the above background, the purpose of the present invention is to provide a method for predicting the CO2 emission concentration of a combustion device based on a CNN-LSTM model, which can realize real-time and accurate prediction of the CO2 emission concentration of the combustion device based on the operating conditions of the combustion device.

[0006] To achieve the above object, the present invention provides a method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model, the method comprising the following steps:

[0007] Extract relevant characteristic variables of continuous time series from the historical data records of the control system and emission monitoring system of the combustion device, export the data at fixed time intervals to obtain input variable data sets;

[0008] Continuously monitor the CO2 emission concentration data at the same time interval and match the input variables of the input variable data set in time series, clean the monitoring data, and obtain the output variable data set;

[0009] According to the correlation coefficient between each input variable and CO2 emission concentration, the input variable data set is screened to remove redundant variable data and irrelevant variable data in the input variable data set;

[0010] Perform data conversion on the output variable data set and the screened input variable data set, using 75% of the total data set as the training set and the other 25% as the test set;

[0011] Construct a CNN-LSTM model as the initial prediction model, input the training set into the initial prediction model for training, perform deep feature extraction on the data of the one-dimensional input variable matrix through the convolutional neural network in the initial prediction model, and then input it into the hidden layer of the long short-term memory neural network in the initial prediction model for output; during the training, synchronously calculate the mean absolute error loss between the training set and the test set, and after each round of training, back propagate to update the state vector of the hidden layer of the long short-term memory neural network at the next moment of the initial prediction model, so that the mean absolute error loss between the predicted value of CO2 emission concentration and the actual measured value of CO2 emission concentration is reduced until the mean absolute error loss is less than 5%, and save the model as a training model;

[0012] The hyperparameters of the initial prediction model are changed, and multiple trainings are performed. The training model with the smallest root mean square error and the largest coefficient of determination is selected as the final prediction model.

[0013] The relevant characteristic variable parameters of the current time step are input into the final prediction model to obtain the predicted value of CO2 emission concentration in the next time step.

[0014] Preferably, the combustion device is a coal-fired boiler, and the relevant characteristic variables include boiler operation characteristic variables, flue gas emission characteristic variables and coal quality characteristic variables. The boiler operation characteristic variables include feed water flow, unit power, furnace pressure, primary air pressure, reheat steam temperature, reheat steam pressure, main steam pressure, main steam temperature, main steam flow, total air volume and total coal consumption. The flue gas emission characteristic variables include flue gas CO concentration, flue gas NO X concentration, flue gas O2 concentration, flue gas SO2 concentration, flue gas flow rate and flue gas humidity, and the coal quality characteristic variables include carbon content.

[0015] Preferably, the cleaning of the monitoring data includes removing outliers and filling in missing values, and the missing values ​​are filled using a linear interpolation method, and the calculation formula is:

[0016]

[0017] In the formula, m is the starting sample number of the missing value, n is the ending sample number of the missing value, and i is the sample number to be processed.

[0018] Preferably, the input variable data set is screened according to the correlation coefficient between each input variable and the CO2 emission concentration, specifically including:

[0019] The Pearson correlation coefficient between each input variable and CO2 emission concentration was calculated using the following formula:

[0020]

[0021] In the formula, x and y are two variable data points, and are the average values ​​of the two variable data respectively;

[0022] The input variables with correlation coefficients greater than 0.4 in the input variable data set are retained, and the input variables with correlation coefficients greater than 0.9 are removed as redundant variables.

[0023] Preferably, data conversion processing is performed on the output variable data set and the screened input variable data set, specifically including:

[0024] The Min-max normalization method is used to convert all variable data into supervised learning data in the interval [-1,1]. The calculation formula is:

[0025]

[0026] In the formula, x i is the i-th data point in the data set, x max and x min are the maximum and minimum values ​​in a certain type of data, respectively, and max and min are the upper and lower limits of the interval required for normalization.

[0027] Preferably, the calculation formula of the mean absolute error loss is:

[0028]

[0029] In the formula, y i and are the actual measured values ​​and predicted values ​​of CO2 emission concentration respectively, and n is the total number of data.

[0030] Preferably, the back propagation specifically includes:

[0031] Use the same computational unit to repeatedly act on the input feature vector x at the current time t t and the state vector h at the previous moment t-1, generating a new state vector h at that moment t ;

[0032] The new state vector h t Backward gradient propagation is performed to calculate the error gradient layer by layer from the output layer to the input layer, and the partial derivatives of the neural network parameters are solved to achieve hyperparameter optimization and minimization of the loss function.

[0033] Preferably, the calculation formula of the root mean square error is:

[0034]

[0035] The calculation formula of the determination coefficient is,

[0036]

[0037] In the above formula, y i , and are the actual measured value, average value and predicted value of CO2 emission concentration respectively, and n is the total number of data.

[0038] Preferably, the hyperparameters of the final prediction model include:

[0039] The input feature dimension is 4, the convolution layer is 2 layers, the pooling layer is maximum pooling, the number of hidden nodes of the long short-term memory neural network is 20 layers, the Dropout is 0.2, the loss function is MAE, the activation function is Relu, the optimization algorithm is Adam, the number of training rounds is 30, and the number of single training parameters Batch size is 24.

[0040] Compared with the prior art, the present invention has the following advantages:

[0041] The present invention discloses a method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model. Compared with the method for predicting by using a model trained by machine learning in the prior art, the present invention obtains real-time and continuous combustion device data to analyze the correlation between input variables and output variables, screens feature variables with high correlation and inputs them into the CNN-LSTM model constructed by the present invention, and fits and mines the relationship between CO2 emission concentration and input variables through a neural network, thereby achieving real-time and accurate prediction of CO2 emission concentration of the combustion device, and providing theoretical support for carbon emission level assessment of the combustion device and low-energy consumption optimized operation of CCUS devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0043] Figure 1 It is the CO2 emission concentration data of the combustion device within one week after data cleaning in the present invention;

[0044] Figure 2 It is a heat map of correlation coefficients between each input variable and CO2 emission concentration in the present invention;

[0045] Figure 3 It is the trend of the MAE loss function change of the training set and the test set in the present invention;

[0046] Figure 4 It is a line graph of the true value and predicted value of the CO2 concentration test set in the present invention;

[0047] Figure 5 It is a scatter plot of the true value and predicted value of the CO2 concentration test set in the present invention and a comparison chart of the zero error curve. DETAILED DESCRIPTION

[0048] The present invention relates to deep learning technology, which is an application of computer technology in the field of industrial forecasting and industrial control. In the process of implementing the present invention, the application of multiple software function modules will be involved. The applicant believes that after carefully reading the application documents and accurately understanding the implementation principle and purpose of the present invention, in combination with the existing known technology, those skilled in the art can fully use their software programming skills to implement the present invention. The aforementioned software function modules include but are not limited to: DCS data and CEMS data extraction modules, CO2 emission concentration monitoring and CO2 emission concentration prediction modules, etc. All those mentioned in the application documents of the present invention belong to this category, and the applicant will not list them one by one.

[0049] The technical solution of the present invention is further described in detail below through specific embodiments and in conjunction with the accompanying drawings. It should be understood that the implementation of the present invention is not limited to the following embodiments, and any form of modification and / or change made to the present invention will fall within the protection scope of the present invention.

[0050] In the present invention, unless otherwise specified, all parts and percentages are weight units, and the equipment and raw materials used can be purchased from the market or are commonly used in the art. The methods in the following embodiments, unless otherwise specified, are conventional methods in the art. The components or equipment in the following embodiments, unless otherwise specified, are universal standard parts or components known to those skilled in the art, and their structures and principles can be learned by those skilled in the art from technical manuals or through conventional experimental methods.

[0051] The embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following detailed description, for the convenience of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present invention.

[0052] An embodiment of the present invention provides a method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model, the method comprising the following steps:

[0053] Extract relevant characteristic variables of continuous time series from the historical data records of the control system and emission monitoring system of the combustion device, export the data at fixed time intervals to obtain input variable data sets;

[0054] Continuously monitor the CO2 emission concentration data at the same time interval and match the input variables of the input variable data set in time series, clean the monitoring data, and obtain the output variable data set;

[0055] According to the correlation coefficient between each input variable and CO2 emission concentration, the input variable data set is screened to remove redundant variable data and irrelevant variable data in the input variable data set;

[0056] Perform data conversion on the output variable data set and the screened input variable data set, using 75% of the total data set as the training set and the other 25% as the test set;

[0057] Construct a CNN-LSTM model as the initial prediction model, input the training set into the initial prediction model for training, perform deep feature extraction on the data of the one-dimensional input variable matrix through the convolutional neural network in the initial prediction model, and then input it into the hidden layer of the long short-term memory neural network in the initial prediction model for output; during the training, synchronously calculate the mean absolute error loss between the training set and the test set, and after each round of training, back propagate to update the state vector of the hidden layer of the long short-term memory neural network at the next moment of the initial prediction model, so that the mean absolute error loss between the predicted value of CO2 emission concentration and the actual measured value of CO2 emission concentration is reduced until the mean absolute error loss is less than 5%, and save the model as a training model;

[0058] The hyperparameters of the initial prediction model are changed, and multiple trainings are performed. The training model with the smallest root mean square error and the largest coefficient of determination is selected as the final prediction model.

[0059] The relevant characteristic variable parameters of the current time step are input into the final prediction model to obtain the predicted value of CO2 emission concentration in the next time step.

[0060] Taking a 300MW coal-fired boiler as an example, each step of the method of the present invention is specifically described as follows.

[0061] From the historical data records of the DCS control system and the CEMS flue gas continuous emission monitoring system of the combustion device, the relevant characteristic variables of the continuous time series are extracted, and the data are exported at fixed time intervals to obtain the input variable data set. The sampling interval of all variables in the data sample is 1 minute, and the data volume is about more than 18,000 groups of samples, which are continuous data in the time series.

[0062] The extracted relevant characteristic variables include boiler operation characteristic variables, flue gas emission characteristic variables and coal quality characteristic variables. Specifically, they include 11 boiler operation characteristic variables: feed water flow (t / h), unit power (MW), furnace pressure (Pa), primary air pressure (KPa), reheat steam temperature (℃), reheat steam pressure (MPa), main steam pressure (MPa), main steam temperature (℃), main steam flow (t / h), total air volume (t / h) and total coal consumption (t / h); 6 flue gas emission characteristic variables: flue gas CO concentration (mg / Nm 3 ), flue gas NO X Concentration (mg / Nm 3 )、Flue gas O2 concentration (%)、Flue gas SO2 concentration (mg / Nm 3 ), flue gas flow (10,000 Nm 3 / h), flue gas humidity (%) and one coal quality characteristic variable: carbon content Cd (%).

[0063] Sampling was performed from the flue, and the CO2 concentration emission data at the same time interval was continuously monitored by a flue gas analyzer. The input variables were matched in time series, and the monitoring data was cleaned to obtain a CO2 emission concentration data set with a time interval of 1 minute for two consecutive weeks for the training set and validation set of the model. The data for one week was as follows: Figure 1 shown.

[0064] According to the normal range of CO2 emission concentration, the monitoring data is cleaned. Specifically, it includes removing abnormal values ​​and filling in missing values. The linear interpolation method is used to fill in the missing values, and the calculation formula is:

[0065]

[0066] In the formula, m is the starting sample number of the missing value, n is the ending sample number of the missing value, and i is the sample number to be processed.

[0067] Based on the calculation of the Pearson correlation coefficient, redundant variables and irrelevant variables are eliminated, and input variables with high correlation are screened. Specifically, the Pearson correlation coefficient between each input variable and CO2 emission concentration is calculated, input variables with correlation coefficients greater than 0.4 are retained, and redundant variables with correlations greater than 0.9 are eliminated.

[0068] The calculation formula of Pearson correlation coefficient is:

[0069]

[0070] In the formula, x and y are the data points of two variables, and are the average values ​​of the two variables.

[0071] The correlation coefficients between each input variable and the CO2 emission concentration as the output variable are shown in Table 1.

[0072] Table 1 Correlation coefficients between input variables and CO2 emission concentration

[0073]

[0074]

[0075] According to Table 1, five input variables with correlation greater than 0.4 were screened out, including unit power (MW), total coal consumption (t / h), flue gas CO concentration (mg / Nm3), flue gas O2 concentration (%) and flue gas flow rate (10,000 Nm3 / h).

[0076] The correlation between the input variables is as follows: Figure 2 As shown in the figure, the correlation between unit power (MW) and total coal consumption (t / h) is as high as 0.94, which is a redundant variable, so one of the characteristic variables is eliminated.

[0077] The screened input variables and CO2 emission concentration data were converted, and then 75% of the entire data set was used as training set data and 25% was used as test set data.

[0078] The data conversion process specifically includes converting all variable data into supervised learning data in the interval [-1,1] based on the influence of the input variables on the output variables. The conversion method is Min-max normalization, and the calculation method is shown in the following formula:

[0079]

[0080] In the formula, xi is the i-th data point in the data set, x max and x min are the maximum and minimum values ​​in a certain type of data, respectively, and max and min are the upper and lower limits of the interval required for normalization.

[0081] A CNN-LSTM model was constructed as the initial prediction model, and the training set was input into the initial prediction model for training. After deep feature extraction of the data of the one-dimensional input variable matrix was performed through the convolutional neural network in the initial prediction model, the data was input into the hidden layer of the long short-term memory neural network in the initial prediction model for output. During the training, the mean absolute error loss between the training set and the test set was calculated synchronously. After each round of training, the state vector of the hidden layer of the long short-term memory neural network of the initial prediction model at the next moment was updated by back propagation, so that the mean absolute error loss between the predicted value of CO2 emission concentration and the actual measured value of CO2 emission concentration was reduced until the mean absolute error loss was less than 5%, and the model was saved as a training model.

[0082] The calculation formula for the mean absolute error loss is,

[0083]

[0084] In the formula, y i and They represent the actual value and predicted value of CO2 emission concentration respectively; n represents the total number of data in the data set in the characteristic variable.

[0085] The gradual reduction of the loss function means that when the maximum number of training rounds is reached, the change of the loss function in each round decreases, such as Figure 3 As shown in the figure, the trend of MAE loss function changes for training and test sets is shown.

[0086] Backpropagation update specifically refers to using the same computing unit to repeatedly act on the input feature vector x at the current time t t and the state vector h at the previous moment t-1 , generating a new state vector h at that moment t ; The new state vector h that will be used to update t Backward gradient propagation is performed to calculate the error gradient layer by layer from the output layer to the input layer, and the partial derivatives of the neural network parameters are efficiently solved to achieve hyperparameter optimization and minimization of the loss function.

[0087] The hyperparameters of the initial prediction model are changed, and multiple trainings are performed. The multiple training models obtained are screened, and the training model with the smallest root mean square error and the largest coefficient of determination is selected as the final prediction model.

[0088] The formula for calculating the root mean square error is,

[0089]

[0090] The calculation formula of the coefficient of determination is,

[0091]

[0092] In the formula, y i , and They represent the actual measured value, average value and predicted value of CO2 emission concentration respectively; n represents the total number of data in the data set in the characteristic variable.

[0093] The model with the smallest RMS error and the largest coefficient of determination was used as the hyperparameters of the final prediction model as shown in Table 2.

[0094] Table 2 Hyperparameters of the final prediction model

[0095]

[0096]

[0097] The relevant characteristic variable parameters of the current time step are input into the final prediction model to obtain the predicted value of CO2 emission concentration in the next time step.

[0098] 25% of the test set is input into the final prediction model for inference, and the final result is as follows Figure 4 The changing trend between the true value and the predicted value of CO2 emission concentration is shown in the figure. The time interval is 1 minute. The root mean square error of the final test set is 0.103 and the determination coefficient is 0.980. Figure 5 The relationship between the scatter distribution of the predicted value of CO2 emission concentration and the zero error curve is shown. It can be seen that the CO2 emission concentration prediction method of the combustion device based on the CNN-LSTM model has a high accuracy in the prediction results of the CO2 emission concentration of the combustion device.

Claims

1. A method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model, characterized in that: The method comprises the following steps: Extract relevant characteristic variables of continuous time series from the historical data records of the control system and emission monitoring system of the combustion device, export the data at fixed time intervals to obtain input variable data sets; Continuously monitor the CO2 emission concentration data at the same time interval and match the input variables of the input variable data set in time series, clean the monitoring data, and obtain the output variable data set; According to the correlation coefficient between each input variable and CO2 emission concentration, the input variable data set is screened to remove redundant variable data and irrelevant variable data in the input variable data set; Perform data conversion on the output variable data set and the screened input variable data set, using 75% of the total data set as the training set and the other 25% as the test set; Construct a CNN-LSTM model as the initial prediction model, input the training set into the initial prediction model for training, perform deep feature extraction on the data of the one-dimensional input variable matrix through the convolutional neural network in the initial prediction model, and then input it into the hidden layer of the long short-term memory neural network in the initial prediction model for output; during the training, synchronously calculate the mean absolute error loss between the training set and the test set, and after each round of training, back propagate to update the state vector of the hidden layer of the long short-term memory neural network at the next moment of the initial prediction model, so that the mean absolute error loss between the predicted value of CO2 emission concentration and the actual measured value of CO2 emission concentration is reduced until the mean absolute error loss is less than 5%, and save the model as a training model; The hyperparameters of the initial prediction model are changed, and multiple trainings are performed. The training model with the smallest root mean square error and the largest coefficient of determination is selected as the final prediction model. The relevant characteristic variable parameters of the current time step are input into the final prediction model to obtain the predicted value of CO2 emission concentration in the next time step.

2. According to claim 1, a method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model is characterized in that: The combustion device is a coal-fired boiler, and the relevant characteristic variables include boiler operation characteristic variables, flue gas emission characteristic variables and coal quality characteristic variables. The boiler operation characteristic variables include feed water flow, unit power, furnace pressure, primary air pressure, reheat steam temperature, reheat steam pressure, main steam pressure, main steam temperature, main steam flow, total air volume and total coal consumption. The flue gas emission characteristic variables include flue gas CO concentration, flue gas NO X concentration, flue gas O2 concentration, flue gas SO2 concentration, flue gas flow rate and flue gas humidity, and the coal quality characteristic variables include carbon content.

3. The method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model according to claim 1, characterized in that: The cleaning of the monitoring data includes removing abnormal values ​​and filling in missing values. The missing values ​​are filled using a linear interpolation method, and the calculation formula is: In the formula, m is the starting sample number of the missing value, n is the ending sample number of the missing value, and i is the sample number to be processed.

4. The method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model according to claim 1, characterized in that: According to the correlation coefficient between each input variable and CO2 emission concentration, the input variable data set is screened, including: The Pearson correlation coefficient between each input variable and CO2 emission concentration is calculated using the following formula: In the formula, x and y are two variable data points, and are the average values ​​of the two variable data respectively; The input variables with correlation coefficients greater than 0.4 in the input variable data set are retained, and the input variables with correlation coefficients greater than 0.9 are removed as redundant variables.

5. The method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model according to claim 1, characterized in that: Perform data conversion on the output variable data set and the filtered input variable data set, including: The Min-max normalization method is used to convert all variable data into supervised learning data in the interval [-1,1]. The calculation formula is: In the formula, x i is the i-th data point in the data set, x max and x min are the maximum and minimum values ​​in a certain type of data, respectively, and max and min are the upper and lower limits of the interval required for normalization.

6. The method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model according to claim 1, characterized in that: The calculation formula of the mean absolute error loss is: In the formula, y i and are the actual measured values ​​and predicted values ​​of CO2 emission concentration respectively, and n is the total number of data.

7. The method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model according to claim 1 is characterized in that: The back propagation specifically includes: Use the same computational unit to repeatedly act on the input feature vector x at the current time t t and the state vector h at the previous moment t-1 , generating a new state vector h at that moment t ; The new state vector h t Backward gradient propagation is performed to calculate the error gradient layer by layer from the output layer to the input layer, and the partial derivatives of the neural network parameters are solved to achieve hyperparameter optimization and minimization of the loss function.

8. The method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model according to claim 1, characterized in that: The calculation formula of the root mean square error is, The calculation formula of the determination coefficient is, In the above formula, y i , and are the actual measured value, average value and predicted value of CO2 emission concentration respectively, and n is the total number of data.

9. The method for predicting CO2 emission concentration of a combustion device based on a CNN-LSTM model according to claim 1, characterized in that: The hyperparameters of the final prediction model include: The input feature dimension is 4, the convolution layer is 2 layers, the pooling layer is maximum pooling, the number of hidden nodes of the long short-term memory neural network is 20 layers, the Dropout is 0.2, the loss function is MAE, the activation function is Relu, the optimization algorithm is Adam, the number of training rounds is 30, and the number of single training parameters Batch size is 24.

Images