Convolutional neural network (CNN) and Transformer-based boiler NOx emission prediction method

Abstract

The invention discloses a boiler NOx emission prediction method based on a convolutional neural network (CNN) and a Transform. According to the method, the advantages of the CNN in the aspect of local feature extraction and the powerful capability of Transform in the aspect of long-time-sequence data modeling are fully fused, so that NOx emission prediction is realized. The method can effectively cope with the influence of flexible deep peak value regulation and control on NOx emission, facilitates improvement of emission monitoring and management efficiency, optimization of execution of environmental protection policies, reduction of NOx emission and guarantee of environmental compliance and economic benefits of boiler operation.

Description

Technical Field

[0001] The present invention relates to the field of boiler NOx emission prediction, and in particular to a boiler NOx emission prediction method based on convolutional neural network (CNN) and Transformer. Background Art

[0002] In boiler NOx emission control, accurate NOx emission prediction is a key factor in ensuring environmental compliance and optimizing energy efficiency. Accurate NOx emission prediction helps optimize the combustion process, reduce harmful gas emissions, and ensure efficient and safe boiler operation. However, boiler NOx emissions are affected by a variety of complex factors, such as boiler load, combustion temperature, and air supply. These factors are often highly nonlinear and time-varying, making NOx emission prediction more complex and challenging.

[0003] Traditional boiler NOx emission prediction methods typically rely on physical models and empirical formulas, employing thermodynamic principles to estimate emissions. However, these methods often rely on simplified assumptions, overlooking the complex dynamics of the combustion process and struggling to accurately capture the fluctuating characteristics of NOx emissions under varying operating conditions. Furthermore, with the diversification of boiler operating modes and evolving control strategies, traditional physical models face significant prediction errors in practical applications, particularly in the face of dynamic changes and flexible control strategies, resulting in limited predictive effectiveness. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem of NOx emission prediction and provide a boiler NOx emission prediction method based on convolutional neural network (CNN) and Transformer. The method proposed in the present invention can quickly and accurately predict future NOx emissions according to historical NOx emission patterns.

[0005] The specific technical solutions provided by the present invention are as follows:

[0006] A boiler NOx emission prediction method based on convolutional neural network (CNN) and Transformer includes the following steps:

[0007] S1. Collect relevant characteristic parameter data from various publicly available parameter data of coal-fired boilers and establish a historical case database;

[0008] S2, filter the dataset from the historical case database and perform preprocessing;

[0009] S3. Select most of the preprocessed data as the training data set and select a small part of the data as the validation data set;

[0010] S4. Use the training dataset to train the CNN-Transformer hybrid model. During the training phase, the quality of the model is judged by the R2 value and RMSE value.

[0011] S5. Use the trained CNN-Transformer hybrid model to predict the NOx concentration of emissions.

[0012] Furthermore, in step S1, the collected data is preliminarily cleaned from the publicly available parameter data of coal-fired boilers to remove noise, outliers and invalid data, and a historical case database is constructed.

[0013] Furthermore, in step S2, a portion of data is screened out from the historical case database as an experimental data set, and the screened experimental data set is preprocessed, which is specifically divided into the following steps:

[0014] S31. Perform outlier detection based on fuzzy C-means clustering on the data in the experimental dataset. Identify and reduce the impact of outliers in the data on the model;

[0015] S32, Z-score data standardization. Z-score data standardization converts data into a form with a mean of 0 and a standard deviation of 1, eliminating the dimensional differences between different features and enhancing the model's ability to learn different features. The specific expression of Z-score data standardization is:

[0016]

[0017] Where Z is the normalized value, x is the original data value, μ is the mean of the data, and σ is the standard deviation of the data;

[0018] S33. Variable importance evaluation based on information entropy theory evaluates the importance of each feature input into the model and identifies input variables that play a key role in NOx emission prediction, thereby improving the interpretability of the model and optimizing feature selection. The specific formula is:

[0019] i(X;Y)=H(Y)-H(Y|X)

[0020] Where H(Y) is the entropy of the target variable Y (i.e., the uncertainty of the target variable), and H(Y|X) is the conditional entropy of the target variable Y given the feature X (i.e., the uncertainty of the target variable given X).

[0021] S34. Use the sliding window method to perform steady-state detection on the sampled data. Define a fixed-size window in the input experimental data and slide it gradually. Calculate the mean and variance within the window to determine whether the system has reached a steady state. The specific formula is:

[0022]

[0023] Among them, w is the size of the sliding window, x i is the i-th data in the sequence.

[0024]

[0025] Among them, w is the size of the sliding window, x i is the i-th data in the sequence, μ t is the mean of the data points in the sliding window at time t.

[0026] Furthermore, in step S3, the preprocessed experimental data set is divided into a training data set and a validation data set.

[0027] Furthermore, in step S4, the preprocessed data is used to train the CNN-Transformer hybrid model, which is specifically divided into the following steps:

[0028] S51. Convert various parameter data of coal-fired boilers into time series data and extract key features: unit load, total fuel quantity, and maximum coal feed quantity, as input features of the model;

[0029] S52. CNN is used to extract local features from the input time series data. CNN performs convolution operations on the time series through a sliding window, thereby capturing local patterns and time dependencies in various parameter data of coal-fired boilers. Through multi-layer convolution operations, higher-level time series features can be gradually extracted and a fixed-dimensional feature vector is generated as the input of the Transformer. The specific formula for convolution is as follows:

[0030]

[0031] Where (x(t)) is the input signal, w(k) is the value of the convolution kernel (filter) at time (k), and K is the length of the convolution kernel;

[0032] S53. The features extracted by CNN are input into the Transformer model. The Transformer model models the input features through the self-attention mechanism to capture the long-term dependencies in the data. Through the multi-layer self-attention mechanism and feed-forward network, the time series of various parameter data of coal-fired boilers are globally modeled to improve the prediction accuracy of NOx emission concentration. The calculation formula of the self-attention mechanism is as follows:

[0033]

[0034] Among them, Q is the query, K is the key, and V is the value.

[0035] The calculation formula of the feedforward network is as follows:

[0036] FeedForward(X)=max(0,XW1+b1)W2+b2

[0037] Among them, W1 and W2 are weight matrices, b1 and b2 are bias terms;

[0038] S54. Optimize the CNN and Transformer models together in end-to-end training. Optimize the entire model through a joint loss function. Adjust the convolution kernel weights of the CNN and the parameters of the self-attention layer in the Transformer. The local features extracted by the CNN cooperate with the global dependencies captured by the Transformer to improve the prediction ability of the model. By comparing with the actual NOx emission concentration data, the evaluation indicators such as the root mean square error (RMSE) and the coefficient of determination (R) are used. 2 To evaluate the model performance, the specific formula of RMSE is as follows:

[0039]

[0040] Where n is the number of samples, y i is the actual value of the i-th sample, is the predicted value of the i-th sample, R 2 The calculation formula is as follows:

[0041]

[0042] Among them, y i is the actual value of the i-th sample, is the predicted value of the i-th sample, is the mean of the true values.

[0043] Furthermore, in step S5, after the training is completed, the test set data is input into the trained CNN-Transformer model, and the model outputs the NOx emission concentration prediction result.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The present invention proposes a boiler NOx emissions prediction method based on a convolutional neural network (CNN) and a Transformer, which can significantly improve the accuracy of NOx emission predictions and effectively adapt to dynamic changes in the boiler combustion process. Compared with traditional prediction methods based on physical models, the present invention uses a convolutional neural network to automatically extract local features from combustion data, capture periodic fluctuations in factors such as temperature and load, and accurately identify global dependencies over long periods of time through the Transformer's self-attention mechanism. This method can effectively cope with the complex nonlinear fluctuations in boiler NOx emissions and reduce the impact of external environmental and operational changes on prediction accuracy.

[0046] The boiler NOx emissions prediction method of this invention reduces reliance on manually set parameters in traditional models while rapidly responding to changes in boiler load and combustion conditions, improving boiler operating efficiency and environmental compliance. This method demonstrates strong adaptability and is applicable to a variety of complex boiler operating scenarios, with high prediction accuracy and good generalization capabilities. Its versatility and efficiency lend it broad potential for practical application in boiler emissions control, providing important technical support for boiler operation optimization, environmental monitoring, and NOx emissions control. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Flow chart of the steps of the present invention

[0048] Figure 2 This is the model prediction effect diagram (50 samples are selected for display) DETAILED DESCRIPTION

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] like Figure 1 As shown, the present invention proposes a boiler NOx emission prediction method based on convolutional neural network (CNN) and Transformer. The specific implementation steps in the embodiment are as follows:

[0051] S1. Collect relevant characteristic parameter data from various publicly available parameter data of coal-fired boilers, such as unit load, total fuel quantity, maximum coal feed quantity, total air volume, main steam pressure, main steam temperature, total primary air volume, total secondary air volume, induced draft fan flue gas flow rate, furnace pressure, flue gas oxygen content at air preheater inlet, primary air temperature at air preheater inlet, primary air temperature at air preheater outlet, main feed water temperature at economizer inlet, feed water flow at economizer inlet, flue gas temperature at economizer outlet, flue gas temperature at air preheater inlet, fan inlet temperature, flue gas temperature at air preheater outlet, main feed water pressure, net flue gas SO discharged from total exhaust port. 2The measured values ​​of content, net flue gas O2 content of total outlet discharge and the establishment of a historical case database are shown in Table 1. The measured values ​​of net flue gas NOx content of total outlet discharge of the selected data range from [0.03, 45.40];

[0052] Table 1: Coal-fired boiler parameter data

[0053]

[0054]

[0055] S2. Filter out a portion of data from the historical case database as the experimental data set, and preprocess the filtered experimental data set. The specific steps are as follows:

[0056] S21. Perform outlier detection based on fuzzy C-means clustering on the data in the experimental dataset. Identify and reduce the impact of outliers in the data on the model;

[0057] S22. Z-score data normalization. Z-score data normalization converts data into a form with a mean of 0 and a standard deviation of 1, eliminating the dimensional differences between different features and enhancing the model's ability to learn different features. The specific expression of Z-score data normalization is:

[0058]

[0059] Where Z is the normalized value, x is the original data value, μ is the mean of the data, and σ is the standard deviation of the data;

[0060] S23. Variable importance evaluation based on information entropy theory. Each feature input into the model is evaluated for importance, identifying input variables that play a key role in NOx emission prediction, thereby improving the interpretability of the model and optimizing feature selection. The specific formula is:

[0061] I(X;Y)=H(Y)-H(Y|X)

[0062] Where H(Y) is the entropy of the target variable Y (i.e., the uncertainty of the target variable), and H(Y|X) is the conditional entropy of the target variable Y given the feature X (i.e., the uncertainty of the target variable given X).

[0063] S24. Use the sliding window method to perform steady-state detection on the sampled data. The sliding window value is set to 5. Define a fixed-size window in the input experimental data and slide it gradually. Calculate the mean and variance within the window to determine whether the system has reached steady state. The specific formula is:

[0064]

[0065] Among them, w is the size of the sliding window, x i is the i-th data in the sequence.

[0066]

[0067] Among them, w is the size of the sliding window, x i is the i-th data in the sequence, μ t is the mean of the data points in the sliding window at time t;

[0068] S3, the preprocessed experimental data set is divided into a training data set and a validation data set in a ratio of 8:2. Specifically, the training data set consists of 8065 data items and the test data set consists of 2016 data items;

[0069] S4. Use the preprocessed data to train the CNN-Transformer hybrid model. The specific steps are as follows:

[0070] S41. Convert various parameter data of the coal-fired boiler into time series data and extract key features: unit load, total fuel quantity, maximum coal feed quantity, total air volume, main steam pressure, main steam temperature, total primary air volume, total secondary air volume, induced draft fan flue gas flow rate, furnace pressure, flue gas oxygen content at the air preheater inlet, primary air temperature at the air preheater inlet, primary air temperature at the air preheater outlet, main feed water temperature at the economizer inlet, feed water flow rate at the economizer outlet, flue gas temperature at the air preheater inlet, flue gas temperature at the air preheater outlet, main feed water pressure, as input features of the model;

[0071] S42. Utilize CNN to extract local features from the input time series data. CNN performs convolution operations on the time series using a sliding window, thereby capturing local patterns and temporal dependencies in various parameter data of coal-fired boilers. Through multiple layers of convolution, higher-level time series features can be gradually extracted, generating fixed-dimensional feature vectors as input to the Transformer. The specific convolution formula is as follows:

[0072]

[0073] Where (x(t)) is the input signal, w(k) is the value of the convolution kernel (filter) at time (k), and K is the length of the convolution kernel;

[0074] S43. The features extracted by the CNN are input into the Transformer model. The Transformer model uses the self-attention mechanism to model the input features and capture long-term dependencies in the data. A multi-layer self-attention mechanism and feed-forward network are used to globally model the time series of various parameter data of coal-fired boilers to improve the prediction accuracy of NOx emission concentration. The calculation formula of the self-attention mechanism is as follows:

[0075]

[0076] Among them, Q is the query, K is the key, and V is the value.

[0077] The calculation formula of the feedforward network is as follows:

[0078] FeedForward(X)=max(0,XW1+b1)W2+b2

[0079] Among them, W1 and W2 are weight matrices, b1 and b2 are bias terms;

[0080] S44. Optimize the CNN and Transformer models together in end-to-end training. Optimize the entire model through a joint loss function, adjusting the convolution kernel weights of the CNN and the parameters of the self-attention layer in the Transformer. In this way, the local features extracted by the CNN and the global dependencies captured by the Transformer work together to improve the model's predictive ability. By comparing with actual NOx emission concentration data, evaluation indicators such as root mean square error (RMSE) and coefficient of determination (R) are used. 2 To evaluate the model performance. The specific formula of RMSE is as follows:

[0081]

[0082] Where n is the number of samples, y i is the actual value of the i-th sample, is the predicted value of the i-th sample.

[0083] R 2 The calculation formula is as follows:

[0084]

[0085] Among them, y i is the actual value of the i-th sample, is the predicted value of the i-th sample, is the mean of the true values;

[0086] S5. The model was trained for 2000 rounds. After training, the best model from the 2000 rounds was selected for testing. The test data was input into the trained CNN-Transformer model, which then output the NOx emission concentration prediction results. See Table 2: Comparison of actual and predicted NOx values.

[0087] Table 2: Comparison of actual and predicted NOx values

[0088]

[0089]

[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A boiler NOx emission prediction method based on convolutional neural network (CNN) and Transformer, characterized in that: The steps include: S1. Collect relevant characteristic parameter data from various publicly available parameter data of coal-fired boilers and establish a historical case database; S2, filter the dataset from the historical case database and perform preprocessing; S3. Select most of the preprocessed data as the training data set and select a small part of the data as the validation data set; S4. Use the training data set to train the CNN-Transformer hybrid model. During the training phase, rely on R 2 The value and RMSE value can be used to judge the quality of the model; S5. Use the trained CNN-Transformer hybrid model to predict the NOx concentration of emissions.

2. The boiler NOx emission prediction method based on convolutional neural network (CNN) and Transformer according to claim 1, characterized in that: In step S1, the collected data is preliminarily cleaned from the publicly available parameter data of coal-fired boilers to remove noise, outliers, and invalid data, and a historical case database is constructed.

3. The boiler NOx emission prediction method based on convolutional neural network (CNN) and Transformer according to claim 1, characterized in that: In step S2, a portion of data is screened from the historical case database as an experimental data set, and the screened experimental data set is preprocessed, which is specifically divided into the following steps: S31. Perform outlier detection based on fuzzy C-means clustering on the data in the experimental dataset to identify and reduce the impact of outliers in the data on the model; S32, Z-score data standardization. Z-score data standardization converts data into a form with a mean of 0 and a standard deviation of 1, eliminating the dimensional differences between different features and enhancing the model's ability to learn different features. The specific expression of Z-score data standardization is: Where Z is the normalized value, x is the original data value, μ is the mean of the data, and σ is the standard deviation of the data; S33. Variable importance evaluation based on information entropy theory evaluates the importance of each feature input into the model and identifies input variables that play a key role in NOx emission prediction, thereby improving the interpretability of the model and optimizing feature selection. The specific formula is: I(X;Y)=H(Y)-H(Y|X) Among them, H(Y) is the entropy of the target variable Y, that is, the uncertainty of the target variable, and H(Y|X) is the conditional entropy of the target variable Y after the given feature X, that is, the uncertainty of the target variable after X is given; S34. Use the sliding window method to perform steady-state detection on the sampled data. Define a fixed-size window in the input experimental data and slide it gradually. Calculate the mean and variance within the window to determine whether the system has reached a steady state. The specific formula is: Among them, w is the size of the sliding window, x i is the i-th data in the sequence; Among them, w is the size of the sliding window, x i is the i-th data in the sequence, μ t is the mean of the data points in the sliding window at time t.

4. The method for predicting boiler NOx emissions based on convolutional neural network (CNN) and Transformer according to claim 1, characterized in that: In step S3, the preprocessed experimental dataset is divided into a training dataset and a validation dataset.

5. The boiler NOx emission prediction method based on convolutional neural network (CNN) and Transformer according to claim 1, characterized in that: In step S4, the preprocessed data is used to train the CNN-Transformer hybrid model, which is specifically divided into the following steps: S51. Convert various parameter data of coal-fired boilers into time series data and extract key features: unit load, total fuel quantity, and maximum coal feed quantity, as input features of the model; S52. CNN is used to extract local features from the input time series data. CNN performs convolution operations on the time series through a sliding window, thereby capturing local patterns and time dependencies in various parameter data of coal-fired boilers. Through multi-layer convolution operations, higher-level time series features are gradually extracted and a fixed-dimensional feature vector is generated as the input of the Transformer. The specific formula for convolution is as follows: Where (x(t)) is the input signal, w(k) is the value of the convolution kernel (filter) at time (k), and K is the length of the convolution kernel; S53. The features extracted by CNN are input into the Transformer model. The Transformer model models the input features through the self-attention mechanism to capture long-term dependencies in the data. Through a multi-layer self-attention mechanism and feed-forward network, the time series of various parameter data of coal-fired boilers are globally modeled to improve the prediction accuracy of NOx emission concentration. The calculation formula of the self-attention mechanism is as follows: Where Q is the query, K is the key, and V is the value. The calculation formula of the feedforward network is as follows: FeedForward(X)=max(0,XW1+b1)W2+b2 Among them, W1 and W2 are weight matrices, b1 and b2 are bias terms; S54. Optimize the CNN and Transformer models together in end-to-end training. Optimize the entire model through a joint loss function. Adjust the convolution kernel weights of the CNN and the parameters of the self-attention layer in the Transformer. The local features extracted by the CNN cooperate with the global dependencies captured by the Transformer to improve the prediction ability of the model. By comparing with the actual NOx emission concentration data, the evaluation indicators such as the root mean square error (RMSE) and the coefficient of determination (R) are used. 2 To evaluate the model performance, the specific formula of RMSE is as follows: Where n is the number of samples, y i is the actual value of the i-th sample, is the predicted value of the i-th sample, R 2 The calculation formula is as follows: Among them, y i is the actual value of the i-th sample, is the predicted value of the i-th sample, is the mean of the true values.

6. The boiler NOx emission prediction method based on convolutional neural network (CNN) and Transformer according to claim 1, characterized in that: In step S5, after the training is completed, the test set data is input into the trained CNN-Transformer model, and the model outputs the NOx emission concentration prediction result.

Images