Flue gas outlet NOx concentration prediction method based on LSTM model real-time calibration

Abstract

The invention discloses a flue gas outlet NOx concentration prediction method based on LSTM model real-time calibration. The flue gas outlet NOx concentration prediction method comprises the following steps: step 1, determining depth related factors; step 2, data acquisition; step 3, data preprocessing; step 4, data division; step 5, constructing an LSTM (Long Short Term Memory) model; 6, performing model training and precision evaluation; step 7, predicting the NOx concentration of the flue gas exhaust port; and 8, calibrating the model in real time. According to the method, accurate prediction of the NOx concentration of the flue gas outlet can be realized; the problem of large lag existing in outlet NOx concentration measurement can be solved, accurate ammonia spraying is realized, and the consumption of a denitration reducing agent is reduced; the method has good generalization ability and strong applicability, and is convenient to popularize in practical application; compared with a traditional statistical model, the method has the advantages that the NOx emission concentration can be predicted more accurately, and a more reliable decision basis is provided for environmental protection and pollution abatement.

Description

Technical Field

[0001] The present invention belongs to the technical field of flue gas emission prediction, and in particular relates to a flue gas outlet NOx prediction system based on real-time calibration of LSTM model. x Concentration prediction method. Background Art

[0002] With the acceleration of industrialization, NO in flue gas x The emission concentration of nitrogen oxides (NOx) is also increasing year by year, which has a serious impact on the environment. At the same time, the country's requirements for environmental protection are constantly increasing, and the industry's air pollution emission standards are also constantly being revised and upgraded. Therefore, it is necessary to accurately predict the NOx at the flue gas outlet. x The concentration of flue gas can achieve fine control of flue gas emission, which is of great significance for environmental protection. x Concentration prediction methods are usually based on statistical models, such as linear regression and support vector machines. However, these methods are not very effective in dealing with nonlinear, time-series, and highly volatile NO x When it comes to emissions data, forecasts are often poor.

[0003] In recent years, the application of deep learning in the field of time series prediction has gradually attracted attention. Among them, the LSTM (Long Short-Term Memory Neural Network) model can effectively process time series data with long-term dependencies due to its unique memory unit structure. It can prevent the problems of gradient disappearance and overfitting in traditional neural network models such as BP neural network, and can effectively filter out noise, thereby obtaining more accurate results. Therefore, we consider applying the LSTM model to NO x Predictions of emission concentrations may have certain advantages.

[0004] However, how to effectively use LSTM models for NO x There are still many challenges in predicting emission concentrations. First, NO x Emission concentration is affected by many factors. How to select appropriate influencing factors as input is crucial to improving prediction accuracy. x Emission concentration data is time-series, and it is necessary to consider the characteristics of time series, such as trends and seasonality. Secondly, the density of data will also have a certain impact on the training of the model. Finally, how to post-process the prediction results is also an important link that affects the final prediction effect.

[0005] In summary, how to apply LSTM model to NO x Emission concentration prediction to achieve more accurate prediction results is a technical problem that urgently needs to be solved. Summary of the invention

[0006] The purpose of the present invention is to provide a smoke exhaust NO based on LSTM model real-time calibration x Concentration prediction method is proposed to solve the above technical problems.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The present invention discloses a smoke exhaust port NO based on LSTM model real-time calibration x A concentration prediction method, the method comprising the following steps:

[0009] Step 1: Determine depth-related factors: According to the SCR denitrification principle, select multiple influencing factors, including inlet flue gas flow, inlet NO x concentration, inlet oxygen volume, outlet oxygen volume, inlet flue gas temperature, ammonia concentration, ammonia escape, unit load, and the influencing factors are compared with the flue gas outlet NO x The correlation analysis was performed on the concentrations, and the factors with strong correlation were regarded as deep correlation factors;

[0010] Step 2: Data collection: Collect and obtain the NO at the flue gas outlet through the historical data of flue gas monitoring. x Historical data on concentrations and their corresponding depth-related factors;

[0011] Step 3: Data preprocessing: preprocess the historical data collected in step 2, including data cleaning, missing value processing, and normalization processing;

[0012] Step 4: Data division: Use the preprocessed data as training data and divide the training data into training set and test set;

[0013] Step 5: Build an LSTM model: Build an LSTM model, set up two LSTM layers, one dropout layer, and finally add a fully connected layer; use the output of the last time step of the LSTM model as input, connect a fully connected layer to generate a prediction result for the export NO in the next time step x Concentration is predicted;

[0014] Step 6: Model training and accuracy evaluation: Use the training set to train the constructed LSTM model, take the depth-related factors as the model input, and convert the corresponding smoke outlet NO x The concentration is used as the model output. During the training process, the back propagation algorithm is used to calculate the error gradient layer by layer from the output layer to the input layer of the LSTM model to solve the partial derivatives of the neural network parameters to achieve the optimization of the network parameters and the minimization of the loss function. The weights and biases of the model are continuously adjusted to minimize the error between the predicted value and the actual value.

[0015] The test set data is used to evaluate the accuracy of the LSTM model, and the mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R) are calculated. 2 ,If the evaluation criteria threshold range is met, the model training is completed;

[0016] Step 7, smoke outlet NO x Concentration prediction: The trained LSTM model is used to access the actual production system to predict the concentration of NO in the flue gas outlet of the actual production system. x The concentration is predicted; that is, the depth-related factor data of the actual production system is pre-processed and input into the trained model, and the model outputs the corresponding flue gas outlet NO x Concentration data, the concentration data is denormalized to obtain the NO x Concentration prediction results;

[0017] Step 8, real-time model calibration: In the process of using the model for actual prediction, the model prediction results are compared with the actual monitoring results at intervals of a certain period of time. If the relative error of the continuous comparison data exceeds the set threshold within the set time, the model is dynamically updated, that is, the original training data is replaced with the new production monitoring data, and the model is retrained to achieve the purpose of dynamic update of the model and real-time calibration.

[0018] Furthermore, in step 1, each influencing factor is respectively associated with the NO of the flue gas outlet. x The correlation analysis of concentration is as follows: Correlation analysis is performed using the Spearman coefficient, and the calculation formula of the correlation coefficient is:

[0019]

[0020] Where, ρ is the correlation coefficient; d i is the difference between the individual value of the influencing factor and the average value of the influencing factor; n is the number of influencing factors; among them, the correlation coefficient |ρ|<0.4 indicates a weak correlation; 0.4≤|ρ|≤0.7 indicates a medium correlation; |ρ|>0.7 indicates a strong correlation.

[0021] Furthermore, the depth-related factors in step 1 include inlet flue gas flow, inlet NO x concentration, outlet oxygen volume, inlet flue gas temperature, and ammonia concentration.

[0022] Furthermore, the smoke outlet NO collected in step 2 x The historical data of concentration and each depth-related factor shall not be less than 20,000.

[0023] Furthermore, the data cleaning described in step 3 refers to treating the data exceeding 50% of the mean in each data as outliers and removing the outliers;

[0024] The missing value processing refers to filling in the missing value according to the mean value of the data at the 5 moments before the blank value;

[0025] The normalization process is because the data have different dimensions, so the data should be normalized to make each indicator in the same order of magnitude. The normalization formula is:

[0026]

[0027] In the formula, x * represents the normalized data; x represents the original data; min represents the minimum value of the data; max represents the maximum value of the data.

[0028] Furthermore, the division of the training data into a training set and a test set in step 4 is specifically as follows: the training data is divided according to the historical time when it is actually generated, the data of the first 80% period is used as the training set, and the data of the last 20% period is used as the test set.

[0029] Furthermore, the parameters for constructing the LSTM model described in step 5 are set as follows: hyperparameter settings for the first layer LSTM, 64 neurons, activation function relu, second layer LSTM, 64 neurons, activation function tanh, time step timestep=5, batch_size is 128, and training epochs are 100; the output of the last time step is connected to the fully connected layer.

[0030] Furthermore, the mean square error MSE, mean absolute error MAE, root mean square error RMSE and determination coefficient R in step 6 are 2 The calculation formulas are:

[0031]

[0032]

[0033]

[0034] R 2 =1 - (SSR / SST) (6)

[0035]

[0036]

[0037] In the formula, SSR is the regression sum of squares; SST is the total deviation sum of squares; m represents the number of data; Indicates the actual value; y i represents the predicted value; Represents the predicted mean.

[0038] Furthermore, the evaluation criteria threshold ranges in step 6 are: MAE<0.4, RMSE<0.5, MSE<0.2, R 2 >0.9.

[0039] Furthermore, the interval period for comparing the model prediction results with the actual monitoring results in step 8 is 30 minutes, the set time for the relative error of the continuous comparison data within the set time to exceed the set threshold is 10 minutes, and the set threshold is 10%.

[0040] The beneficial effect of the present invention is that the method of the present invention utilizes the advantages of the LSTM model in processing time-series, nonlinear and complex data, and can better capture NO x Dynamic change law of emission concentration to achieve NO x Accurately predict the concentration; by training the LSTM model, we can learn the long-term dependencies and patterns in historical data, improve the accuracy of prediction, and solve the problem of NOx at the export port. x The problem of large lag in concentration measurement is solved to achieve accurate ammonia injection, reduce the consumption of denitrification reducing agent, and reduce the outlet NO x The probability of environmental protection events of instantaneous exceeding the standard and hourly average exceeding the standard occurs; and the method of the present invention can be applied to flue gas outlets of different regions and different sizes, has good generalization ability, strong applicability, and is easy to promote in practical applications; compared with traditional statistical models, the method of the present invention can more accurately predict NO x Emission concentration provides a more reliable decision-making basis for environmental protection and pollution control.

[0041] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the process of the method of the present invention;

[0043] Figure 2 This is a schematic diagram of the correlation analysis results of various influencing factors;

[0044] Figure 3 Schematic diagram of model loss changing with the number of iterations. DETAILED DESCRIPTION

[0045] The present invention discloses a smoke exhaust port NO based on LSTM model real-time calibration x Concentration prediction methods, such as Figure 1As shown, the method comprises the following steps:

[0046] Step 1: Determine depth-related factors: According to the SCR denitrification principle, select the inlet flue gas flow rate, inlet NO x concentration, inlet oxygen volume, outlet oxygen volume, inlet flue gas temperature, ammonia concentration, ammonia escape, unit load and other influencing factors. x The correlation analysis was performed with the concentrations.

[0047] Specifically, the Spearman coefficient is used for correlation analysis. The Spearman coefficient is a non-parametric method for measuring the correlation between two variables. It measures the degree of association between the two variables by calculating the Pearson product-moment correlation coefficient between the ranks of the two variables. The Spearman coefficient does not require the data to follow a normal distribution, so it is very suitable for evaluating the correlation of nonlinear relationships or ordered data. The calculation formula for the correlation coefficient is:

[0048]

[0049] Where, ρ is the correlation coefficient; d i is the difference between the individual value of the influencing factor and the average value of the influencing factor; n is the number of influencing factors. Among them, the correlation coefficient |ρ|<0.4 indicates a weak correlation; 0.4≤|ρ|≤0.7 indicates a medium correlation; |ρ|>0.7 indicates a strong correlation.

[0050] Through the correlation analysis of the above-mentioned influencing factors, the factors with strong correlation are selected as the deep correlation factors. x concentration, inlet flue gas flow, inlet NO x concentration, inlet oxygen volume, outlet oxygen volume, inlet flue gas temperature, ammonia concentration, ammonia slip, and unit load are shown in Table 1.

[0051] Table 1

[0052] serial number Related factors 0 <![CDATA[NO concentration at the flue gas discharge port x > 1 Inlet flue gas flow 2 <![CDATA[Inlet NO x Concentration]]> 3 Inlet oxygen volume 4 Outlet oxygen volume 5 Inlet flue gas temperature 6 Ammonia concentration 7 Ammonia slip 8 Unit load

[0053] The correlation analysis results of the above-mentioned influencing factors are as follows: Figure 2 As shown in the figure, the darker the color, the stronger the correlation between these factors. According to the correlation analysis results, the inlet flue gas flow rate and inlet NO x Concentration, outlet oxygen volume, inlet flue gas temperature, and ammonia concentration are used as deep correlation factors for subsequent model training.

[0054] Step 2: Data collection: Collect and obtain the NO at the flue gas outlet through the historical data of flue gas monitoring. x The historical data of concentration and its corresponding depth-related factors, and the collected NOx The historical data of concentration and each depth-related factor shall not be less than 20,000 to improve the accuracy of subsequent predictions.

[0055] Step 3: Data preprocessing: Preprocess the historical data collected in step 2. Preprocessing includes data cleaning, missing value processing, and normalization processing.

[0056] Data cleaning means treating data that exceeds 50% of the mean as outliers and removing them. The cleaned data can better reflect the real situation and improve the accuracy of the prediction.

[0057] Missing value processing refers to filling in the blank value with the mean value of the data at the 5 moments before the blank value.

[0058] Normalization is done because the data have different dimensions, which will affect the results of data analysis. Therefore, the data should be normalized. After the original data is normalized, each indicator is at the same order of magnitude, which is suitable for comprehensive comparative evaluation. The normalization formula is:

[0059]

[0060] In the formula, x * represents the normalized data; x represents the original data; min represents the minimum value of the data; max represents the maximum value of the data.

[0061] Step 4, data division: The preprocessed data is used as training data, and the training data is divided into a training set and a test set. Specifically, the training data is divided according to the actual historical time when it was generated, with the first 80% of the time period as the training set and the last 20% of the time period as the test set, to ensure that the model has good generalization ability in future predictions.

[0062] Step 5: Construct LSTM model: Construct LSTM (Long Short-Term Memory Neural Network) model to perform NO x Concentration prediction. Specifically, two layers of LSTM and one dropout layer are set to prevent overfitting, and finally a fully connected layer is added. The hyperparameter settings are as follows: the first layer of LSTM, the number of neurons is 64, the activation function is relu, the second layer of LSTM, the number of neurons is 64, the activation function is tanh, the time step is timestep=5, the batch_size is 128, and the number of training epochs is 100. The output of the last time step is connected to the fully connected layer: In this case, the output of the last time step of the LSTM model is used as input, and a fully connected layer is connected to generate a prediction result for the outlet NO in the next time step. x Concentration prediction.

[0063] Step 6: Model training and accuracy evaluation: Use the training set to train the constructed LSTM model, take the depth-related factors as the model input, and convert the corresponding smoke outlet NO x The concentration is used as the model output. During the training process, the back propagation algorithm is used, that is, by calculating the error gradient layer by layer from the output layer to the input layer of the LSTM model, the partial derivatives of the neural network parameters are solved to optimize the network parameters and minimize the loss function. The weights and biases of the model are continuously adjusted to minimize the error between the predicted value and the actual value. The training loss of the model, that is, the gap between the predicted result of the model and the actual result, generally approaches 0 as the number of iterations increases, such as Figure 3 shown.

[0064] The mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R) are used. 2 The model accuracy is evaluated and the calculation formulas are:

[0065]

[0066]

[0067]

[0068] R 2 =1 - (SSR / SST) (6)

[0069]

[0070]

[0071] In the formula, SSR is the regression sum of squares; SST is the total deviation sum of squares; m represents the number of data; Indicates the actual value; y i represents the predicted value; Represents the predicted mean.

[0072] The test set data is used to evaluate the accuracy of the LSTM model, and the mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R) are calculated. 2 If the evaluation criteria threshold range (generally set according to the experience value in this field) is met, that is, MAE<0.4, RMSE<0.5, MSE<0.2, R 2 >0.9, it indicates that the model has achieved a good fit, that is, the model training is complete.

[0073] Step 7, smoke outlet NO xConcentration prediction: The trained LSTM model is used to access the actual production system to predict the concentration of NO in the flue gas outlet of the actual production system. x Specifically, the depth-related factor data of the actual production system is preprocessed and then input into the trained model, and the model outputs the corresponding NO concentration at the flue gas outlet. x Concentration data, the concentration data is denormalized to obtain the NO x Concentration prediction results. Based on the prediction results, the denitrification process of flue gas emissions can be finely controlled to achieve the purpose of energy saving and environmental protection.

[0074] Step 8, real-time model calibration: In the process of using the model for actual prediction, the model prediction results are compared with the corresponding actual monitoring results (the actual monitoring result lag time is 2 minutes) every 30 minutes. If the relative error of the compared data for 10 consecutive minutes exceeds 10%, the model is dynamically updated, that is, the original training data is replaced with the new production monitoring data, and the model is retrained to achieve the purpose of dynamic update of the model for real-time calibration.

[0075] The method of the present invention is based on real-time calibration of the LSTM model, which can make full use of historical data information to adjust the NO x Compared with traditional statistical methods or traditional neural network models, this method has better prediction performance and generalization ability, and can accurately predict NO concentration at flue gas outlets. x It provides strong support for concentration control and precise management.

[0076] Embodiment 1

[0077] The above method has been applied in many power plants such as the thermal power station of the production control center of Chalco Guangxi Huasheng New Materials Co., Ltd. and the Liupanshan thermal power plant of Chalco Ningxia Energy Group Co., Ltd., and achieved good results. The following is a comparison list of some predicted data and actual monitoring data of the thermal power station of the production control center of Chalco Guangxi Huasheng New Materials Co., Ltd. According to the results analysis in Table 3, the relative error between the predicted value and the monitored value is basically below 5%, achieving a good prediction accuracy.

[0078] Table 3

[0079]

[0080]

[0081] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred arrangement scheme, a person skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A smoke outlet NOx detector based on real-time calibration of LSTM model x The concentration prediction method is characterized in that The method comprises the following steps: Step 1: Determine depth-related factors: According to the SCR denitrification principle, select multiple influencing factors, including inlet flue gas flow, inlet NO x concentration, inlet oxygen volume, outlet oxygen volume, inlet flue gas temperature, ammonia concentration, ammonia escape, unit load, and the influencing factors are compared with the flue gas outlet NO x The correlation analysis was performed on the concentrations, and the factors with strong correlation were regarded as deep correlation factors; Step 2: Data collection: Collect and obtain the NO at the flue gas outlet through the historical data of flue gas monitoring. x Historical data on concentrations and their corresponding depth-related factors; Step 3: Data preprocessing: preprocess the historical data collected in step 2, including data cleaning, missing value processing, and normalization processing; Step 4: Data division: Use the preprocessed data as training data and divide the training data into training set and test set; Step 5: Build an LSTM model: Build an LSTM model, set up two LSTM layers, one dropout layer, and finally add a fully connected layer; use the output of the last time step of the LSTM model as input, connect a fully connected layer to generate a prediction result for the export NO in the next time step x Concentration is predicted; Step 6: Model training and accuracy evaluation: Use the training set to train the constructed LSTM model, take the depth-related factors as the model input, and convert the corresponding smoke outlet NO x The concentration is used as the model output. During the training process, the back propagation algorithm is used to calculate the error gradient layer by layer from the output layer to the input layer of the LSTM model to solve the partial derivatives of the neural network parameters to achieve the optimization of the network parameters and the minimization of the loss function. The weights and biases of the model are continuously adjusted to minimize the error between the predicted value and the actual value. The test set data is used to evaluate the accuracy of the LSTM model, and the mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R) are calculated. 2 ,If the evaluation criteria threshold range is met, it indicates that the model training is completed; Step 7, smoke outlet NO x Concentration prediction: The trained LSTM model is used to access the actual production system to predict the concentration of NO in the flue gas outlet of the actual production system. x The concentration is predicted; that is, the depth-related factor data of the actual production system is pre-processed and input into the trained model, and the model outputs the corresponding flue gas outlet NO x Concentration data, the concentration data is denormalized to obtain the NO x Concentration prediction results; Step 8, real-time model calibration: In the process of using the model for actual prediction, the model prediction results are compared with the actual monitoring results at intervals of a certain period of time. If the relative error of the continuous comparison data exceeds the set threshold within the set time, the model is dynamically updated, that is, the original training data is replaced with the new production monitoring data, and the model is retrained to achieve the purpose of dynamic update of the model and real-time calibration.

2. A smoke outlet NO based on LSTM model real-time calibration according to claim 1 x The concentration prediction method is characterized in that In step 1, each influencing factor is respectively associated with the NO x The correlation analysis of concentration is as follows: Correlation analysis is performed using the Spearman coefficient, and the calculation formula of the correlation coefficient is: Where, ρ is the correlation coefficient; d i is the difference between the individual value of the influencing factor and the average value of the influencing factor; n is the number of influencing factors; among them, the correlation coefficient |ρ|<0.4 indicates a weak correlation; 0.4≤|ρ|≤0.7 indicates a medium correlation; |ρ|>0.7 indicates a strong correlation.

3. A smoke outlet NO based on LSTM model real-time calibration according to claim 1 x The concentration prediction method is characterized in that The depth-related factors in step 1 include inlet flue gas flow, inlet NO x concentration, outlet oxygen volume, inlet flue gas temperature, and ammonia concentration.

4. A smoke outlet NO based on LSTM model real-time calibration according to claim 1 x The concentration prediction method is characterized in that The smoke outlet NO collected in step 2 x The historical data of concentration and each depth-related factor shall not be less than 20,000.

5. A smoke outlet NO based on LSTM model real-time calibration according to claim 1 x The concentration prediction method is characterized in that The data cleaning described in step 3 refers to treating the data exceeding 50% of the mean as outliers and removing the outliers; The missing value processing refers to filling in the missing value according to the mean value of the data at the 5 moments before the blank value; The normalization process is because the data have different dimensions, so the data should be normalized to make each indicator in the same order of magnitude. The normalization formula is: In the formula, x * represents the normalized data; x represents the original data; min represents the minimum value of the data; max represents the maximum value of the data.

6. A smoke outlet NO based on LSTM model real-time calibration according to claim 1 x The concentration prediction method is characterized in that The step 4 of dividing the training data into a training set and a test set is specifically as follows: the training data is divided according to the historical time when it is actually generated, with the data of the first 80% period as the training set and the data of the last 20% period as the test set.

7. A smoke outlet NO based on LSTM model real-time calibration according to claim 1 x The concentration prediction method is characterized in that The parameters for constructing the LSTM model described in step 5 are set as follows: hyperparameter settings for the first layer LSTM, 64 neurons, activation function relu, second layer LSTM, 64 neurons, activation function tanh, time step timestep = 5, batch_size size is 128, training epochs are 100; the output of the last time step is connected to the fully connected layer.

8. A smoke outlet NO based on LSTM model real-time calibration according to claim 1 x The concentration prediction method is characterized in that The mean square error MSE, mean absolute error MAE, root mean square error RMSE and determination coefficient R mentioned in step 6 2 The calculation formulas are: R 2 =1 - (SSR / SST) (6) In the formula, SSR is the regression sum of squares; SST is the total deviation sum of squares; m represents the number of data; Indicates the actual value; y i represents the predicted value; Represents the predicted mean.

9. A smoke outlet NO based on LSTM model real-time calibration according to claim 1 x The concentration prediction method is characterized in that The evaluation criteria threshold ranges described in step 6 are: MAE<0.4, RMSE<0.5, MSE<0.2, R 2 >0.

9.

10. A smoke outlet NO based on LSTM model real-time calibration according to claim 1 x The concentration prediction method is characterized in that The interval period for comparing the model prediction results with the actual monitoring results in step 8 is 30 minutes, the set time for the relative error of the continuous comparison data to exceed the set threshold within the set time is 10 minutes, and the set threshold is 10%.

Images