Ammonia injection amount adjustment method for SCR system considering boiler combustion state

Abstract

The present invention provides a method for adjusting the ammonia injection amount of an SCR system considering the boiler combustion state. First, process parameter data characterizing the unit combustion state and within the SCR system are acquired. Based on MIC, the delay time between each process parameter and the NOx emission at the SCR outlet is calculated. An initial prediction model for the NOx emission at the SCR outlet is established based on ELM, and the initial predicted value of the NOx emission at the SCR outlet at the current moment is obtained. A dynamic error correction model is established based on ELM to obtain the error correction value of the NOx emission at the SCR outlet at the current moment. The prediction results of the two models are superimposed, and then a dynamic prediction model for the NOx emission at the SCR outlet is established. Considering the goals such as the NOx emission at the SCR outlet meeting the national emission standards and reducing the ammonia injection cost of the SCR system, an intelligent method for adjusting the ammonia injection amount is designed based on NSGA-II, and then the optimization adjustment of the ammonia injection amount is completed. This method can reduce the NOx emission concentration at the SCR system outlet while reducing the ammonia injection cost of the power plant's denitration system, which is of great significance for guiding the operation of the power plant.

Description

Technical Field

[0001] The present invention belongs to the technical field of thermal power generation, and particularly relates to a method for adjusting the ammonia injection amount of an SCR system considering the boiler combustion state. Background Art

[0002] At present, new energy powers such as wind energy and solar energy have developed rapidly. After large-scale new energy powers with randomness and volatility are connected to the power grid, a deep peak shaving demand is generated, and thus flexible operation of coal-fired units is required. The deep peak shaving of coal-fired units means that the unit load fluctuates rapidly within a large range, which affects the combustion state of the unit and causes the nitrogen oxide emissions generated by combustion to fluctuate violently, increasing the difficulty of accurately controlling the ammonia injection amount of the Selective Catalytic Reduction (SCR) system. The SCR denitration system is an important technical means for flue gas treatment in power plants. Realizing the accurate dynamic modeling of the SCR denitration system and the intelligent optimization regulation of the ammonia injection amount is of great significance for the ultra-low emission and economic operation of nitrogen oxides in power plants.

[0003] The SCR denitration reaction process is complex and affected by various factors such as unit load and flue gas temperature, resulting in characteristics such as non-linearity and strong disturbance of the SCR system; and due to factors such as the measurement delay of the Continuous Emission Monitoring System (CEMS) and the dynamic change of the denitration reaction time, the SCR system has the characteristic of dynamic time delay. The above characteristics make the dynamic modeling of the SCR system and the optimization of the ammonia injection amount a challenging problem. In the current research methods for intelligent regulation of the ammonia injection amount, the optimization setting value of the ammonia injection amount under the actual operating conditions is mainly solved by a single-objective optimization algorithm, and usually only the single optimization objective of reducing the NOx emission concentration at the outlet of the SCR system is considered, without considering the consumption cost of the ammonia injection amount, which has certain limitations for guiding the economic operation of the denitration system. Summary of the Invention

[0004] Based on the above problems, the present invention provides a method for adjusting the ammonia injection amount of an SCR system considering the boiler combustion state, including:

[0005] Step 1: Collect a process parameter data set D representing the combustion state of the unit and the operating state of the SCR system; where D = {X t , Y t}, Y t = {y t} ∈ R t×1 , t = t1, t2,..., t N , t represents the sampling time, t1 is the initial sampling time, t NLet \(t\) be the termination sampling time, \(N\) be the number of modeling samples, \(i = 1, 2, 3, \ldots, m\), where \(m\) represents the number of process parameters, and \(X\) t is the set of \(m\) process parameters, is the sample value of the \(i\)-th process parameter at time \(t\), and \(y\) t is the NOx emission concentration at the outlet of the SCR system at time \(t\);

[0006] The process parameters are divided into output parameters and input parameters. Among them, the output parameter is the NOx emission concentration at the outlet of the SCR system, and the input parameters include 3 process parameters representing the combustion state of the unit, namely unit load, total air volume, and total coal consumption, and 12 process parameters representing the operating state of the SCR system, namely ammonia injection amount, inlet flue gas temperature, outlet flue gas temperature, inlet flue gas pressure, outlet flue gas pressure, inlet oxygen concentration, outlet oxygen concentration, inlet nitrogen oxide concentration, inlet carbon monoxide concentration, flue gas flow rate, dilution fan current, and ammonia slip concentration;

[0007] Step 2: Use the min-max normalization method to normalize the process parameter dataset \(D\) to obtain the normalized modeling dataset \(ND\). The formula for min-max normalization is as follows:

[0008]

[0009] where, is the minimum value of the \(i\)-th input parameter, is the maximum value of the \(i\)-th input parameter, is the normalized value of the \(i\)-th input parameter, is the minimum value of the \(i\)-th output parameter, is the maximum value of the \(i\)-th output parameter, is the normalized value of the \(i\)-th output parameter;

[0010] Step 3: Use the Maximal Information Coefficient (MIC) to calculate the delay time between each process parameter and the NOx emission concentration at the outlet of the SCR system, and reconstruct the modeling dataset according to the calculated delay time; including:

[0011] Step 3-1: Determine the delay time range \((0, K]\) of each process parameter, where \(K\) is the maximum delay time;

[0012] Step 3-2: For each process parameter in sample from the 1st moment before the sampling time to the \(k\)-th moment before the sampling time in the delay time range \((0, K]\) in turn to construct the calculation dataset of MIC

[0013]

[0014] wherein is the value of the i-th process variable at the moment 1 before the sampling moment t1, is the value of the i-th process variable at the moment k before the sampling moment tN,

[0015] Step 3-3: Calculate the MIC values between each column vector and the output vector in the data set using the MIC algorithm, and select the column vector with the largest MIC value as the delay moment k of the corresponding process parameter i (i = 1, 2,..., m), k i is the best delay moment corresponding to each process parameter Continuously repeat this process to obtain m best delay moments in total;

[0016] Step 3-4: Combine each group of process parameters corresponding to the best delay moment and the output variable Y t to form a reconstructed data set xc. Among them, the first column represents the data of the output variable, and each remaining column represents the value of each process parameter at the best delay moment;

[0017]

[0018] In the formula, is the value of the first process parameter at the corresponding delay moment k1, is the value of the m-th process parameter at the corresponding delay moment k m ;

[0019] Step 4: Use the Extreme Gradient Boosting (XGBoost) algorithm to screen the feature variables in the reconstructed data set xc as the modeling data SD; specifically stated as:

[0020] Combined with the analysis of the denitration reaction mechanism, the NOx concentration at the inlet of the SCR system and the ammonia injection amount are used as the feature variables for modeling. The importance of the remaining variables is determined using the Extreme Gradient Boosting algorithm. Calculate the importance of the remaining process parameters in the reconstructed data set xc except for the NOx concentration at the inlet of the SCR system and the ammonia injection amount relative to the NOx emission concentration at the outlet of the SCR system in the first column of the reconstructed data set xc. By setting an importance threshold, select the process parameters greater than the importance threshold as the feature variables for modeling. The data set after feature selection is SD = {S t , Y t t = t1,..., t N}, S tis the set of process parameters after feature selection, where sm ≤ m;

[0021] Step 5: Establish a prediction model M for the NOx emission concentration at the outlet of the SCR system based on the Extreme Learning Machine (ELM) h , specifically expressed as: the prediction model M h includes the initial prediction model M for the NOx emission concentration at the outlet of the SCR system based on ELM P and the dynamic error correction model M E . Using the modeling data set SD after feature selection, the initial prediction value y P (t) of the test sample is obtained by adopting the initial prediction model M p . Then, the error prediction value e E (t) of the current prediction sample is obtained by adopting the error correction model M p . By superimposing the prediction results of the two models, the final prediction value y h (t) of the NOx emission concentration at the outlet of the SCR system is finally obtained;

[0022] The construction process of the initial prediction model M P is as follows:

[0023] Step 5-1-1: Divide the modeling data into a training set and a test set;

[0024] Step 5-1-2: Use the training set to complete the training of the weight matrix between the hidden layer and the output layer of ELM. The trained model is called the initial prediction model M P ;

[0025] Step 5-1-3: Use the initial prediction model M P to predict the test sample, and obtain the initial prediction value y p (t) of the NOx emission concentration at the outlet of the SCR system;

[0026] The construction process of the dynamic error correction model M E is as follows:

[0027] Step 5-2-1: Obtain the initial prediction value y P (t) of the initial prediction model M p on the training set, and then subtract it from the actual measured value to obtain the error data sequence e p (t) of the training set. The calculation formula of e p (t) is as follows;

[0028] e p (t) = y p (t) - y m (t) (4)

[0029] Among them, y p (t) is the predicted value of the initial model M at the current time t, P y m (t) is the actual measured value of the NOx emission concentration at the outlet of the SCR system at the current time t;

[0030] Step 5-2-2: Use the historical errors e p (t-1), e p (t-2), e p (t-3) and the input S of the current initial prediction model t as the input, and the error e p (t) at the current time as the output to train the ELM network to obtain the error correction model M E ;

[0031] Step 5-2-3: Use the error correction model M E to complete the prediction of the test set samples and obtain the error correction values of the test set samples;

[0032] Finally, superimpose the prediction results of the initial model M P and the error correction model M E to finally obtain the final predicted value y h (t) of the NOx emission concentration at the outlet of the SCR system;

[0033] Step 6: Select the ammonia injection amount Q as the controllable variable, establish a multi-objective optimization function for the controllable variable, and use the non-dominated sorting genetic algorithm NSGA-II with an elite strategy to solve the multi-objective optimization function to obtain the optimal solution set that meets the optimization objectives; including:

[0034] Step 6-1: Establish the multi-objective optimization function expressed as:

[0035]

[0036] In the formula, f NOx (s) represents the deviation between the predicted NOx concentration value output by the prediction model M h and the expected value, f price (Q) represents the ammonia injection consumption of the SCR system, s is the input sample of the prediction model M h , sm is the dimension of the input sample, Q is the ammonia injection amount, s1, s2,..., s sm-1 are the process variables in the input sample except the ammonia injection amount, E NOx is the expected value of the NOx emission concentration at the outlet of the SCR system, M h (s) is the prediction model M hThe predicted value of the NOx emission concentration at the outlet of the SCR system, M h (s) less than 50 mg / Nm 3 ;

[0037] Step 6-2: Use the non-dominated sorting genetic algorithm NSGA-II to solve the multi-objective optimization function, find the Pareto front solution set of the error between the ammonia injection consumption and the NOx predicted value and the expected value output by the SCR hybrid prediction model, and obtain the optimal solution set that meets the optimization objectives. The operator can select the results in the optimal solution set for ammonia injection control.

[0038] The beneficial effects of the present invention are:

[0039] The present invention proposes a method for adjusting the ammonia injection amount of the SCR system considering the boiler combustion state. The maximum information coefficient is used to estimate the delay time of each variable and data reconstruction is performed, which can make up for the defect of large time lag in the original system; and the extreme gradient boosting algorithm is used to screen the modeling variables of the reconstructed data set, so as to realize the dynamic prediction of the NOx emission concentration at the outlet of the SCR system, which can overcome the limitation of the network structure of the traditional algorithm, extract the deep features of the data, and establish a prediction model and an error correction model at the same time, having the advantage of high prediction accuracy; considering reducing the ammonia injection cost while meeting the NOx emission concentration requirements, the non-dominated sorting genetic algorithm NSGA-II with an elite strategy is used to solve the multi-objective optimization function, and the solution set of ultra-low ammonia injection amount that meets the NOx emission concentration requirements is obtained. This method can reduce the NOx emission concentration at the outlet of the SCR system while reducing the ammonia injection cost of the power plant's denitration system, which is of great significance for guiding the operation of the power plant. Description of the Drawings

[0040] Figure 1 It is a flow chart of the method for adjusting the ammonia injection amount of the SCR system considering the boiler combustion state in the present invention.

[0041] Figure 2 It is a flow chart of the method for establishing a prediction model of the NOx emission concentration at the outlet of the SCR system in the present invention.

[0042] Figure 3 It is the predicted value of the NOx emission concentration at the outlet of the SCR system on the test set in the present invention.

[0043] Figure 4 It is a comparison chart of the ammonia injection consumption and the outlet NOx emission concentration before and after the optimization adjustment in the present invention. Detailed Embodiments

[0044] The present invention will be further described below in conjunction with the drawings and specific implementation examples.

[0045] For the precise optimization and adjustment of the ammonia injection amount in the SCR system, the present invention provides a multi-objective optimization and adjustment method for the ammonia injection amount in the SCR system considering the boiler combustion state. The method includes: obtaining the process parameter data characterizing the unit combustion state and the SCR system through the plant-level monitoring information system (SIS) of the thermal power plant; calculating the delay time between each process parameter and the NOx emission concentration at the outlet of the SCR system based on the maximum information coefficient (MIC); establishing an initial prediction model for the NOx emission concentration at the outlet of the SCR system based on the extreme learning machine, obtaining the initial predicted value of the NOx emission concentration at the outlet of the SCR system at the current moment, and establishing a dynamic error correction model based on ELM to obtain the error correction value of the NOx emission concentration at the outlet of the SCR system at the current moment, and superimposing the prediction results of the two models to further establish a dynamic prediction model for the NOx emission concentration at the outlet of the SCR system; considering the objectives of meeting the national emission standards for the NOx emission concentration at the outlet of the SCR system and reducing the ammonia injection cost of the SCR system, designing an intelligent adjustment method for the ammonia injection amount based on the non-dominated sorting genetic algorithm with elitist strategy (NSGA-II), and then completing the optimization and adjustment of the ammonia injection amount.

[0046] As Figure 1 shown, a method for adjusting the ammonia injection amount in the SCR system considering the boiler combustion state includes:

[0047] Step 1: Collecting a process parameter data set D characterizing the unit combustion state and the SCR system operation state through the plant-level monitoring information system; where D = {X t , Y t}, Y t = {y t} ∈ R, t = t1, t2, …, t N , t represents the sampling time, t1 is the initial sampling time, t N is the termination sampling time, N is the number of modeling samples, i = 1, 2, 3, …, m, m represents the number of process parameters, X t is a set of m-dimensional process parameters, is the sample value of the i-th process parameter at time t, y t is the NOx emission concentration at the outlet of the SCR system at time t;

[0048] In this embodiment, a data set containing 16 parameters is collected. Among them, the output parameter is the NOx emission concentration at the outlet of the SCR system. The output parameter includes 3 process parameters characterizing the combustion state of the unit, namely unit load, total air volume, and total coal quantity, and 12 process parameters characterizing the operation state of the SCR system, namely ammonia injection quantity, inlet flue gas temperature, outlet flue gas temperature, inlet flue gas pressure, outlet flue gas pressure, inlet oxygen concentration, outlet oxygen concentration, inlet nitrogen oxide concentration, inlet carbon monoxide concentration, flue gas flow rate, dilution fan current, and ammonia slip concentration. The process parameter data from the start time of 06:25:10 on April 2, 2016 to the end time of 08:00:00 on April 3, 2016 is collected at intervals of 10s, with a total of 9210 groups of data.

[0049] Step 2: Perform a normalization operation on the process parameter data set D using the min-max normalization method to obtain the normalized modeling data set ND. The formula for min-max normalization is as follows:

[0050]

[0051] Where, is the minimum value of the i-th input parameter, is the maximum value of the i-th input parameter, is the normalized value of the i-th input parameter, is the minimum value of the i-th output parameter, is the maximum value of the i-th output parameter, is the normalized value of the i-th output parameter;

[0052] Step 3: Calculate the delay time between each process parameter and the NOx emission concentration at the outlet of the SCR system using the Maximal Information Coefficient (MIC), and reconstruct the modeling data set according to the calculated delay time; including:

[0053] Step 3-1: According to the on-site unit operation experience, determine the maximum delay time range (0, K] of each process parameter, where K is the maximum delay time; assume that the delay times of each process parameter are d1, d2,..., d m , where d i is the delay time of the i-th process parameter . According to the actual operation situation of the selected unit and relevant operation experience, set the maximum delay time range of the process parameter to 600s. Assume that the delay times of 15 process parameters are d1, d2,..., d 15 . Since the sampling interval is 10s, the maximum delay moment k of each process parameter is 60.

[0054] Step 3-2: For each process parameter Within the delay time range (0, K], samples are taken in sequence from the 1st moment before the sampling time to the kth moment before, to construct the MIC calculation data set

[0055]

[0056] wherein, is the value of the ith process variable at the 1st moment before the sampling moment t1, is the value of the ith process variable at the kth moment before the sampling moment tN, k = 60, N = 9210, to obtain the variable reconstructed data matrix, where each column vector represents the reconstructed data sequence with different delays within the delay moment range (0, 60].

[0057] Step 3 - 3: Analyze the reconstructed data matrix using the maximum information coefficient MIC for the nonlinear correlation between each row vector and the output vector in, and select the column vector with the largest MIC value as the corresponding process parameter delay moment k i (i = 1, 2,..., m), the delay moment k corresponding to this column vector i is the process parameter optimal delay moment process parameter optimal delay time T is the sampling interval; calculate the delay times of all process parameters in sequence according to the above method to complete the solution of the delay times of each process parameter and the reconstruction of the overall modeling data set;

[0058] The specific calculation formula of MIC is as follows:

[0059]

[0060]

[0061] where I(x, y) is the mutual information value between variables x and y, p(x) and p(y) are the marginal probability distributions of variables x and y respectively, p(x, y) is the joint probability distribution of variables x and y, and B(n) is the 0.6th power of the sample data volume n;

[0062] Step 3 - 4: According to the optimal delay moment of each variable Select the data corresponding to the optimal delay moment of each variable as the starting data and the output data for rearrangement. The reconstructed data set is xc, where the first column represents the output data, and each of the remaining columns is a process parameter reconstructed according to the optimal delay moment;

[0063]

[0064] In the formula, is the value of the first process parameter at the corresponding delay moment k1, is the value of the m-th process parameter at the corresponding delay moment k m ;

[0065] The delay times of the calculated process parameters are shown in Table 1.

[0066] Table 1 Delay time table of each process parameter

[0067]

[0068] Step 4: Use the Extreme Gradient Boosting (XGBoost) algorithm to screen the characteristic variables in the reconstructed data set xc as the modeling data; specifically expressed as:

[0069] Use the Extreme Gradient Boosting algorithm to determine the importance of each process parameter in the reconstructed data set relative to the NOx emission concentration at the outlet of the SCR system. By setting an importance threshold, select the process parameters greater than the importance threshold as the characteristic variables for modeling. At the same time, combined with the analysis of the denitration reaction mechanism, select the NOx at the inlet of the SCR and the ammonia injection amount as the characteristic variables for modeling. The data set after feature selection is SD = {S t ,Y t |t = t1,…,t N}, S t is the set of characteristic variables with a dimension of sm after feature selection, where sm ≤ m;

[0070] In this embodiment, the importance threshold is set to 0.2. Use XGBoost to calculate the importance of each process parameter and normalize the calculation results. Select the process parameters with a normalized importance value greater than 0.2 as the input parameters of the prediction model. The calculation results of XGBoost are shown in Table 2. Considering that the denitration reaction is the selective catalytic reduction reaction of NOx at the inlet and ammonia in the SCR system reactor, the NOx concentration at the inlet is used as the input parameter of the prediction model. At the same time, variables such as the ammonia injection amount, outlet flue gas temperature, inlet flue gas temperature, outlet oxygen concentration, unit load, and total air volume with a normalized importance greater than 0.2 are used as the input parameters of the prediction model. The input data set S after feature selectiont The characteristic dimension sm is 7.

[0071] Table 2 Results of importance calculation based on XGBoost

[0072]

[0073] Step 5: Establish a prediction model M for the NOx emission concentration at the outlet of the SCR system based on the Extreme Learning Machine (ELM) h , including: establishing an initial prediction model M for the NOx emission concentration at the outlet of the SCR system based on the Extreme Learning Machine P and a dynamic error correction model M E , using the modeling data set after feature selection, and adopting the initial prediction model M P to obtain the initial predicted value y p (t) of the test sample, and then adopting the error correction model M E to obtain the error predicted value e p (t) of the current prediction sample, and superimposing the prediction results of the two models to finally establish a hybrid prediction model M h for the NOx emission concentration at the outlet of the SCR system, and obtaining the final predicted value y h (t);

[0074] The construction process of the initial prediction model M P is as follows:

[0075] Step 5-1-1: Divide the modeling data into a training set and a test set; select the first 7500 groups of data as the training set, and the remaining 1710 groups of data as the test set.

[0076] Step 5-1-2: Use the training set to complete the training of the weight matrix between the hidden layer and the output layer of the ELM. Set the number of neurons in the hidden layer to 25, randomly initialize the weight matrix between the input layer and the hidden layer and the biases of the neurons in the hidden layer, and solve the output weight matrix between the hidden layer and the output layer through the training set to complete the training process of the ELM. The trained model is called the initial prediction model M P ;

[0077] Step 5-1-3: Use the initial prediction model M P to predict the 1710 groups of test samples, and obtain the initial predicted value y p (t) of the NOx emission concentration at the outlet of the SCR system;

[0078] Regarding the dynamic error correction problem of the initial prediction model M P for the NOx emission concentration at the outlet of the SCR system, use the ELM to mine Mp Based on the regularity of the prediction error, a dynamic error correction model M is established based on ELM E The construction process is as follows:

[0079] Step 5-2-1: Obtain the initial prediction model M P The initial predicted value on the training set, and then take the difference with the actual measured value to obtain the error data sequence e p (t) of the training set, and the calculation formula of e p (t) is as follows;

[0080] e p (t) = y p (t) - y m (t) (4)

[0081] Among them, y p (t) is the predicted value of the initial model M at the current time t P , and y m (t) is the actual measured value of the NOx emission concentration at the outlet of the SCR system at the current time t;

[0082] Step 5-2-2: Use the historical errors e(t - 1), e(t - 2), e(t - 3) of the previous three moments and the input S of the current initial prediction model t as the input, and the error e(t) at the current moment as the output to train the ELM network to obtain the error correction model M E ;

[0083] Step 5-2-3: Use the error correction model M E to complete the prediction of the test set samples and obtain the error correction values of the test set samples;

[0084] Step 5-2-4: The predicted value of the NOx emission concentration at the outlet of the SCR system on the test set is as Figure 2 shown. Use the initial prediction model M P to obtain the initial predicted value y p (t) of the test sample, and then use the error correction model M E to obtain the error predicted value e p (t) of the current predicted sample. Superimpose the prediction results of the two models to finally establish the hybrid prediction model M h of the NOx emission concentration at the outlet of the SCR system, and obtain the final predicted value y h (t) of the NOx emission concentration at the outlet of the SCR system. The specific calculation formula is as follows:

[0085] y h (t) = y p (t) + e p (t)

[0086] Step 6: To reduce the NOx emission concentration at the outlet of the SCR system and the cost of ammonia injection consumption, and considering the controllability of relevant operating parameters within the SCR system, the ammonia injection rate Q is selected as the controllable variable, a multi-objective optimization function for the controllable variable is established, and the non-dominated sorting genetic algorithm NSGA-II with an elitist strategy is used to solve the multi-objective optimization function to obtain the optimal solution set that meets the optimization objectives; including:

[0087] Step 6-1: Establish the multi-objective optimization function as follows:

[0088]

[0089] In the formula, f NOx (s), f price (Q) are respectively two constructed optimization objective functions, which respectively refer to minimizing the error between the predicted NOx value output by the prediction model M h and the expected value, and minimizing the ammonia injection consumption within the SCR system. s is the input sample of the hybrid prediction model M h , sm is the dimension of the input sample, Q is the ammonia injection rate, s1, s2, …, s sm-1 are the non-adjustable process variables in the input sample except the ammonia injection rate, E NOx is the expected value of the NOx emission concentration at the outlet of the SCR system, M h (s) is the actual value of the NOx emission concentration at the outlet of the SCR system obtained by the hybrid prediction model M h , satisfying that M h (s) is less than 50 mg / Nm 3 ;

[0090] After feature selection, the dimension sm of the input sample s is 7, where Q is the ammonia injection rate, s1, s2, …, s sm-1 are the process variables in the input sample except the ammonia injection rate, namely variables such as the outlet flue gas temperature, inlet flue gas temperature, outlet oxygen concentration, unit load, total air volume, etc. Considering the model prediction error and the actual operation of the unit SCR system, to reduce the NOx emission concentration at the outlet, E NOx is set to 30 mg / Nm 3 .

[0091] Step 6-2: Use the non-dominated sorting genetic algorithm NSGA-II with an elitist strategy to solve the multi-objective optimization function and obtain the optimal solution set that meets the optimization objectives;

[0092] The NSGA-II algorithm is used to solve the above determined objective function to obtain the Pareto optimal solution set that meets the optimization objectives, and then determine the optimized ammonia injection rate and the NOx emission concentration at the outlet of the SCR system after the ammonia injection rate is optimized and adjusted. The prediction results are as followsFigure 3 As shown, 10 groups of samples are selected for optimization, and the optimization results are as Figure 4 shown. It can be seen that the method of the present invention can effectively reduce the ammonia injection consumption and the NOx emission concentration at the outlet of the SCR system by optimizing and adjusting the ammonia injection amount on the premise of meeting the national pollutant discharge standards.

Claims

1. A method for adjusting the ammonia injection amount of an SCR system considering the boiler combustion state, characterized in that, Including: Step 1: Collect the process parameter dataset D characterizing the combustion state of the unit and the operating state of the SCR system; Step 2: Perform normalization operation on the modeling dataset D using the min-max normalization method to obtain the normalized modeling dataset ND; Step 3: Calculate the delay time between each process parameter and the NOx emission at the SCR outlet using the maximum information coefficient, and reconstruct the modeling dataset according to the calculated delay time; Step 4: Screen the reconstructed feature variables using the extreme gradient boosting algorithm as the modeling data; Step 5: Establish a prediction model M for the NOx emission concentration at the outlet of the SCR system based on the extreme learning machine h , and obtain the final predicted value y h (t) of the NOx emissions at the SCR outlet; Step 6: Select the ammonia injection quantity Q as the controllable variable, establish a multi-objective optimization function for the controllable variable, and use the non-dominated sorting genetic algorithm NSGA-II with elitist strategy to solve the multi-objective optimization function to obtain the optimal solution set that meets the optimization objectives; The multi-objective optimization function is expressed as: minf NOx (s) = M h (s) - E NOx minf price (Q) = Q s = [Q, s1, s2, …, s sm-1 ​ s.t.M h (s) < 50 Wherein, f NOx (s), f price (Q) are two constructed optimization objective functions, respectively referring to minimizing the error between the NOx predicted value output by the SCR hybrid prediction model M h and the expected value, and minimizing the ammonia injection consumption in the SCR system. s is the input sample of the hybrid prediction model M h , sm is the dimension of the input sample, Q is the ammonia injection amount, s1, s2, …, s sm-1 are uncontrollable process variables in the input sample except the ammonia injection amount, E NOx is the expected value of the NOx emission at the SCR system outlet, M h (s) is the actual value of the NOx emission at the SCR outlet obtained by the hybrid prediction model M h .

2. The method for adjusting the ammonia injection amount of the SCR system considering the boiler combustion state according to claim 1, characterized in that, The formula for min-max normalization in Step 2 is as follows: wherein, is the minimum value of the i-th process parameter, is the maximum value of the i-th process parameter, is the sample value of the i-th process parameter at time t, is the normalized value of the i-th process parameter.

3. The method for adjusting the ammonia injection amount of the SCR system considering the boiler combustion state according to claim 1, characterized in that, Step 3 includes: Step 3-1: Determine the maximum delay time range [0, K] for each process parameter, where K is the maximum delay time; Step 3-2: For each process parameter, sample from the 1st moment before the sampling time to the kth moment before the sampling time in sequence within the delay time range (0, K], and construct a calculation data set for the maximum information coefficient MIC Step 3-3: Calculate the MIC values between each column vector in the dataset and the output vector Y t Select the column vector with the largest MIC value as the optimal delay moment of the process parameter, and calculate the optimal delay moments of m process parameters; Step 3-4: Combine the process parameters and the output variable Y corresponding to the m optimal delay times t into a reconstructed data set xc, where the first column represents the elements of the output variable, and each of the remaining columns represents the values of the respective process parameters at the optimal delay times.

4. A method for adjusting the ammonia injection amount of an SCR system considering the boiler combustion state according to claim 1, characterized in that, Step 4 is specifically described as: The extreme gradient boosting algorithm is used to determine the importance of each process parameter in the reconstructed dataset relative to the SCR outlet NOx emissions. By setting an importance threshold, the process parameters greater than the importance threshold are selected as the characteristic variables for modeling. At the same time, combined with the analysis of the denitration reaction mechanism, the SCR inlet NOx and ammonia injection amount are used as the characteristic variables for modeling. The dataset after feature selection is SD = {S t ,Y t |t=t1,…,t N}, S t is the set of characteristic variables with the dimension of sm after feature selection, where sm ≤ m.

5. A method for adjusting the ammonia injection amount of an SCR system considering the boiler combustion state according to claim 1, characterized in that, Step 5 includes: an initial prediction model M of the NOx emission concentration at the outlet of the SCR system based on ELM P and a dynamic error correction model M E , using the modeling data set after feature selection, and obtaining the initial predicted value y P (t) of the test sample by adopting the initial prediction model M p , and then obtaining the error predicted value e E (t) of the current prediction sample by adopting the error correction model M p , superposing the prediction results of the two models, and finally obtaining the final predicted value y h (t) of the NOx emission concentration at the outlet of the SCR system; The initial prediction model M P is constructed as follows: Step 5-1-1: Divide the modeling data into a training set and a test set; Step 5-1-2: Use the training set to complete the training of the weight matrix between the hidden layer and the output layer of the extreme learning machine. The trained model is called the initial prediction model M P ; Step 5-1-3: Using the initial prediction model M P Predict the test samples to obtain the initial predicted value y p (t) of the NOx emissions at the SCR outlet; The construction process of the dynamic error correction model M E is as follows: Step 5-2-1: Obtain the initial prediction model M P The initial predicted value y p (t) on the training set is subtracted from the actual measured value to obtain the error data sequence e p (t); Step 5-2-2: Use the historical errors e(t-1), e(t-2), e(t-3) at the previous three moments and the input S of the current initial prediction model t as inputs, and the error e(t) at the current moment as the output. Train the extreme learning machine network to obtain the error correction model M E ; Step 5-2-3: Use the error correction model M E Complete the prediction of the test set samples to obtain the error correction values of the test set samples.

Images