A multi-index model predictive control method for cement burning denitrification system

Abstract

The present invention discloses a multi-index model predictive control method for a cement firing denitration system, belonging to the field of flue gas denitration control in the cement firing process, including establishing an accurate prediction model for predicting the flue gas emissions in the cement firing process by using historical data in a database; using the error between the predicted value of the prediction model and the real-time NO emission concentration of the denitration system to perform feedback correction on the prediction model; setting the NO concentration value according to the requirements of the environmental protection department for the NO emission concentration, and performing softening treatment on the set NO concentration value to form a reference trajectory; using the differential evolution algorithm to solve the value of the control variable, and according to the prediction model, solving the input value of the prediction model that makes the output value close to the reference trajectory, and bringing the first solved control quantity value into the denitration system to complete the control of the denitration system in the cement production process. The present invention realizes stable control of NO emissions in the cement firing process while reducing ammonia escape, achieving the purpose of reducing energy consumption and pollution. x emission concentration to perform feedback correction on the prediction model; according to the requirements of the environmental protection department for NO x emission concentration, set the NO x concentration value, and perform softening treatment on the set NO x concentration value to form a reference trajectory; use the differential evolution algorithm to solve the value of the control variable, and according to the prediction model, solve the input value of the prediction model that makes the output value close to the reference trajectory, and bring the first solved control quantity value into the denitration system to complete the control of the denitration system in the cement production process. The present invention realizes stable control of NO x emissions while reducing ammonia escape, achieving the purpose of reducing energy consumption and pollution.

Description

Technical Field

[0001] The invention relates to the field of flue gas denitration control in a cement burning process, and in particular to a multi-index model predictive control method for a cement burning denitration system. Background Art

[0002] The cement industry is an indispensable raw material industry for my country's economic development and production construction. The cement burning process is the most important link in cement production, and this process consumes the most energy and produces the most pollutants. Flue gas nitrogen oxides (NO x ) emission concentration is an important indicator for measuring pollutant emissions during cement burning. x Precise control of emission concentration can reduce pollutant emissions and ammonia escape. Therefore, it is of great significance to accurately control flue gas denitrification during cement burning. The cement burning process has complex nonlinear and time-varying delay characteristics, and it is difficult to control NO using traditional control methods. x The emission concentration is stably and accurately controlled.

[0003] In response to the above problems, some scholars have adopted different control methods to study the control of denitrification. For example:

[0004] A DMC-PID cascade controller for the selective catalytic reduction (SCR) denitrification system of a thermal power plant using a multi-model switching method is proposed. It is a local linear system model under different operating conditions, and the controller parameters under different operating conditions are designed according to the quadratic optimal adjustment principle. Compared with the traditional cascade PID, the SCR denitrification system is more stably controlled and reduces the use of ammonia.

[0005] A denitrification control strategy combining NARX neural network and improved dynamic matrix control algorithm is designed to cope with the impact of frequent peak load regulation of power grid on the combustion system of coal-fired power plant boilers and reduce the NO in the flue gas of coal-fired power plant boilers. x Fluctuations in concentration.

[0006] It can be seen that the existing research is all focused on the SCR denitrification system of thermal power plants; the cement burning process produces NO x The mechanism is more complicated and the selective non-catalytic reduction reaction (SNCR) method is used for denitration. Due to its own limitations, the above research is difficult to solve the long delay, nonlinearity and volatility problems in the cement burning denitration process. Therefore, it is necessary to develop a multi-index model predictive control method for the cement burning denitration system that can solve the above problems. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a multi-index model predictive control method for a cement burning denitrification system, which not only solves the problems of long delay, nonlinearity and volatility in the denitrification process of cement burning flue gas, but also realizes NO x The emission concentration is stably controlled while reducing ammonia slip emissions.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0009] A multi-index model predictive control method for a cement burning denitrification system comprises the following steps:

[0010] Step 1, using historical data in the database to establish an accurate prediction model for flue gas emissions during cement burning;

[0011] Step 2: Use the predicted value of the prediction model and the real-time NO x The error in emission concentration is used to provide feedback correction to the prediction model;

[0012] Step 3: According to the environmental protection department’s NO x The emission concentration requirement is set as NO x concentration value, and set the NO x The concentration value is softened to form a reference trajectory;

[0013] Step 4: Use the differential evolution algorithm to solve the value of the control variable, solve the prediction model input value that makes the output value close to the reference trajectory according to the prediction model, and bring the first control value solved into the denitrification system to complete the control of the denitrification system in the cement production process.

[0014] A further improvement of the technical solution of the present invention is that: in step 1, firstly, NO x The generation mechanism of NO was analyzed and 15 x The process variables related to emission concentration and ammonia slip are used as the input of the prediction model. The variables are arranged in chronological order into a time series as the input of the prediction model. In order to achieve multi-objective prediction of flue gas emissions, the MIMO strategy is used to achieve multi-index output of the prediction model and realize simultaneous prediction of NO x Emission concentration and ammonia slip magnitude;

[0015] By analyzing the influence of NO in the whole cement burning process x The factors that cause and eliminate the problem, combined with the experience and knowledge of the field engineers, the 15 variables finally selected are: Group 1 to Group 9 spray gun ammonia flow rate X 1 ~X 9 、Smoke Room O 2 FeedbackX 10 、Smoke room NOx FeedbackX 11 、Smoke chamber CO feedback X 12 , Decomposition furnace temperature feedback X 13 , coal feeding amount of decomposition furnace X 14 , O 2 FeedbackX 15 .

[0016] A further improvement of the technical solution of the present invention is that in step 1, the prediction model is a multi-objective prediction model of flue gas emission concentration in the cement burning process combined with a bidirectional long short-term memory network of a time series, and the specific steps of constructing the prediction model include:

[0017] 1.1, build a bidirectional long short-term memory network;

[0018] 1.2, introduce the time series into the input layer of the bidirectional long short-term memory network structure to avoid the influence of delay and build a prediction model;

[0019] 1.3, training the prediction model, inputting the data to be predicted into the prediction model, and obtaining the prediction results of each indicator.

[0020] The further improvement of the technical solution of the present invention is that in step 1.2, specifically:

[0021] The input variable data is integrated into a time series of m-step time periods and introduced into the input layer of the bidirectional long short-term memory network structure. The time series X(t) is expressed as:

[0022] X 1 =[X 1 (t),X 1 (t+1),…,X 1 (t+m)] (1)

[0023] X(t)=[X 1 (t)+X 2 (t)+…+X 15 (t)] (2)

[0024] In the formula, X 1 is the time series corresponding to the ammonia flow rate variable of the first group of spray guns, X 1 (t)~X 15 (t) represents 15 kinds of NO x The time series corresponding to the relevant process variables at time t, X(t) is the time series of the input layer of the bidirectional LSTM model, m is the width of the time series, and t is the current time;

[0025] The bidirectional long short-term memory network uses two independent LSTM hidden layers to represent the feature information extracted by the network in the forward time sequence and the feature information extracted in the reverse time sequence respectively, and links them to obtain the final prediction model output. The hidden state H of the prediction model at time t is t Contains forward and backward

[0026]

[0027]

[0028]

[0029] In the formula, T is the sequence length, h t-1 is the hidden state of the LSTM layer at the previous moment, h t+1 is the hidden state of the LSTM layer at the next moment, x t is the input data at the current moment, c t-1 is the cell state at the previous moment, c t+1 is the cell state at the next moment.

[0030] A further improvement of the technical solution of the present invention is that after determining the network structure by analyzing the denitrification data set, the optimal parameters of the prediction model are: the unit cells of the two LSTM hidden layers are 200 respectively, the forgetting probability of the random inactivation layer is 0.1, the number of training times is 50 times, the optimization function is the adaptive momentum estimation method, and the loss functions are the mean absolute error, the root mean square error, and the symmetric mean absolute percentage error.

[0031] The further improvement of the technical solution of the present invention is that: in step 2, the actual output is compared with the predicted value at each step of the control to correct the uncertainty of the prediction model. When the denitration system has time-varying, model mismatch and interference factors, feedback correction can correct the predicted value in time, so that the optimization is based on a more accurate prediction, so as to improve the robustness of the control system;

[0032] The feedback correction is to correct the prediction model by calculating the error between the prediction value of the prediction model and the real-time flue gas emission concentration of the denitrification system, thereby improving the stability and accuracy of the denitrification process control:

[0033] e(t+1)=y(t)-y m (t+1) (13)

[0034] y p (t+1)=y m (t)+h*e(t+1) (14)

[0035] Where t is the current time, e(t+1) is the error between the predicted value of the calculation prediction model and the actual flue gas emission value of the denitrification system, and y m (t+1) is the predicted value of the prediction model at the next moment, y(t) is the actual flue gas concentration of the denitrification system, h is the feedback coefficient, and y m (t) is the predicted value of the prediction model at the current moment, y p (t+1) is the output value of the flue gas emission concentration at the next moment after feedback correction.

[0036] A further improvement of the technical solution of the present invention is that: in step 3, in order to avoid a sharp change in the input and output of the denitration system, the output of the multi-index model predictive control process is made to reach the set value along an expected gentle curve, and the actual output value and the set value at the current moment are subjected to a first-order exponential transformation to obtain a softened smooth reference trajectory;

[0037] The reference trajectory used is a first-order exponential change form, using the actual NO x Emission concentration and NO x The emission concentration set value is transformed into a softened reference trajectory by first-order exponential transformation to obtain the reference value at the future moment:

[0038] y r (k+i)=α j y(k)+(1-α j )y r ,j=1,2,…p (15)

[0039]

[0040] Where k is the current time, j is the control time domain size, α is the softening coefficient, T is the sampling period, τ is the time constant, and y(k) is the actual NO at time k. x Emission concentration value, y r NO x Emission concentration set value, y r (k+j) is the reference trajectory after softening.

[0041] The further improvement of the technical solution of the present invention is that: in step 4, the differential evolution algorithm includes population initialization, mutation, crossover, selection, boundary absorption, and then generates new individuals. The prediction model has a total of 15 input variables, including 9 groups of spray gun ammonia flow u 1 ~u 9 As the control variables of the control system, these 9 variables need to be optimized and solved;

[0042] The differential evolution algorithm is used to solve the value of the control variable. The specific algorithm is as follows:

[0043] 4.1, Randomly generate the initial population:

[0044] {X i (0)|X i (0) = [x i,1 ,x i,2 ,...,x i,9 ],i=1,2,...NP} (17)

[0045] x i,j =a j +rand·(b j -a j ) (18)

[0046] In the formula, j = 1, 2, ..., 9, NP represents the population size, X i (0) represents the i-th individual in the initial population, x i,j represents the jth component of the ith individual, a j and b j Respectively represent x ij The upper and lower bounds of the value range, rand represents a random number uniformly distributed in the interval [0,1];

[0047] 4.2, mutation operation:

[0048] The DE algorithm implements mutation operation through the difference method. The basic method is to randomly select two different individuals in the current population, scale their difference vectors, and perform vector operations with other individuals to be mutated to generate new individuals:

[0049] V i (g+1)=X r1 (g)+F·(X r2 (g)-X r1 (g)) (19)

[0050] Where i = 1, 2, ..., NP, i ≠ r 1 ≠r 2 ≠r 3 , r 1 ,r 2 ,r 3 are all random integers in the interval [1, NP], NP is the population size, F is the scaling factor, g represents the evolutionary generation, X i (g) represents the i-th individual in the g-th generation population. After mutation, the g-th generation population generates a new intermediate population V i (g+1);

[0051] 4.3 Crossover Operation:

[0052] For the g-th generation population X i(g) and its variant intermediate population V i (g+1) performs crossover operation between individuals:

[0053]

[0054] Where i = 1, 2, ..., NP, j = 1, 2, ..., 9, U i (g+1)=[u i,1 ,u i,2 ,…,u i,9 ] represents the new population of the g+1th generation, u i,j (g+1) and v i,j (g+1) are populations U i (g+1) and V i The jth component in (g+1), CR represents the crossover probability, j rand is a random integer in the interval [1,9], x ij represents the jth component of the i-th individual;

[0055] 4.4, Select operation:

[0056] In order to ensure the effectiveness of the solution in the algorithm implementation, the upper and lower limits of the denitrification system variable data are determined according to the actual situation of the cement production plant for constraints; if the individual exceeds the constraint range, a new individual is generated by population initialization to replace it; the selection operation is to select the individual entering the new population by calculating the size of the objective function:

[0057]

[0058] Where i = 1, 2, ..., NP, f is the objective function, U i (g+1) is the intermediate population of the g+1th generation, X i (g) is the population of the gth generation, X i (g+1) is the new generation population generated after the g-th generation population is selected;

[0059] Repeat the above four steps until the conditions are met, and solve the control quantity by minimizing the objective function.

[0060] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0061] 1. The multi-objective prediction model of harmful flue gas emission concentration in cement burning process constructed by the present invention adopts MIMO strategy to realize multi-index output of the prediction model in order to realize multi-objective prediction of flue gas emission, and realizes the ability to simultaneously predict NO x Emission concentration and ammonia slip size.

[0062] 2. In order to solve the time delay problem, the present invention arranges the variables into a time series in chronological order as the input layer of Bi-LSTM, and the prediction accuracy of the prediction model is higher than that of the existing model.

[0063] 3. The present invention establishes a model predictive controller for the denitrification system of cement burning process based on a neural network predictive model. The accurate predictive model provides guidance for the controller, and the differential evolution algorithm performs rolling optimization and solves the problem, thus achieving the goal of NO x The emission concentration is stably controlled while reducing the emission of ammonia slip. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative labor.

[0065] Figure 1 It is a schematic diagram of the structure of a multi-objective prediction model for harmful flue gas emission concentration in a cement burning process based on MT-BiLSTM in an embodiment of the present invention;

[0066] Figure 2 is a flow chart of a differential evolution algorithm in an embodiment of the present invention;

[0067] Figure 3 The present invention provides a flowchart of a multi-index model predictive control method for a cement burning denitrification system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or apparatus.

[0069] The embodiment of the present application provides a multi-index model predictive control method for a cement burning denitrification system, which solves the problems of long delay, nonlinearity and volatility in the cement burning denitrification process that are difficult to solve in the prior art. The general idea is: first, 15 input variables related to NOx emission concentration and ammonia escape size are selected from the database of the cement burning system to establish an accurate flue gas emission concentration prediction model. The prediction model is a bidirectional long short-term memory network (MT-BiLSTM) prediction model combined with a time series; the predicted value of the prediction model is combined with the real-time NO of the system.x The prediction model is corrected based on the error of emission concentration; then the control quantity of the denitrification system is solved through the optimization algorithm. It can not only solve the problems of long delay, nonlinearity and volatility in the denitrification process of cement burning flue gas, but also achieve NO x While the emission concentration is stably controlled, the emission of ammonia escape is reduced, achieving the goal of reducing consumption and pollution.

[0070] The cement industry is the most important part of building material production. The cement burning process is a high energy consumption and high pollution process. x Emission concentration is an important indicator for measuring pollutant emissions. Accurate control in the denitrification system can not only reduce pollutant emissions but also reduce ammonia consumption. Therefore, accurate control of the denitrification process can achieve NO x The stable control of emission concentration is very important. Because the cement burning process has complex nonlinear and time-delay characteristics, it is difficult to control NO using traditional control methods. x The emission concentration can be stably and accurately controlled, and it is easy to cause excessive ammonia escape emissions. In order to solve the problem of modeling difficulties caused by the nonlinearity and time delay characteristics of the denitrification process of the cement burning system, the present invention innovatively proposes a cement burning process flue gas NO combined with a bidirectional long short-term memory network of time series. x Emission concentration prediction model; feedback correction and differential evolution algorithm are used to solve the values ​​of various control quantities in the denitrification system to achieve stable control of the denitrification system.

[0071] The present invention is further described in detail below with reference to the accompanying drawings and embodiments:

[0072] like Figure 1-3 As shown, the present invention proposes a multi-index model predictive control method for a cement burning denitrification system, which is a flue gas denitrification control method for a cement burning process based on a neural network model predictive control, comprising the following steps:

[0073] Step 1: Select NO from the database of cement burning system x Fifteen input variables related to emission concentration are used to construct a multi-objective prediction model for flue gas emission concentration in the cement burning process based on a bidirectional long short-term memory network (MT-BiLSTM) combined with time series.

[0074] Specifically include:

[0075] First, the influence of NO on the whole cement burning process is analyzed. x Factors that affect the generation and elimination of NO x The relevant process variables are used as input variables, such as Figure 1As shown in the input layer, the 15 input variables are respectively the ammonia flow rate of the spray guns in groups 1 to 9, X 1 ~X 9 、Smoke Room O 2 FeedbackX 10 、Smoke room NO x FeedbackX 11 、Smoke chamber CO feedback X 12 , Decomposition furnace temperature feedback X 13 , coal feeding amount of decomposition furnace X 14 , O 2 FeedbackX 15 The flue gas emission concentration detection point is at the chimney at the end of the kiln. The denitrification reaction is carried out in the decomposition furnace and preheater. The humidification and dust removal processes in the middle cause the input model input variables to be different from NO x There is a long time delay between emission concentrations. The present invention arranges the input variables into time series data of a certain length as the input layer of Bi-LSTM, fully extracts the time delay information between the data, and eliminates the influence of the time delay on the prediction accuracy.

[0076] In step 1, time delay processing and bidirectional long short-term memory network are involved:

[0077] Due to the response lag caused by the long process flow of the denitrification system, the measurement of flue gas emission concentration is not only related to the state of the input variables at the current moment, but also affected by the state of the variables in the m-step time period. The time series is integrated into the input layer of the MT-BiLSTM network structure to avoid the significant impact of time delay on the prediction results. The time series X(t) can be expressed as:

[0078] X 1 =[X 1 (t),X 1 (t+1),…,X 1 (t+m)] (1)

[0079] X(t)=[X 1 (t)+X 2 (t)+…+X 6 (t)] (2)

[0080] In the formula, X 1 is the time series corresponding to the ammonia flow rate variable of the first group of spray guns, X 1 (t)~X 15 (t) represents 15 kinds of NO x The time series corresponding to the relevant process variables at time t, X(t) is the time series of the input layer of the bidirectional LSTM model, m is the width of the time series, and t is the current time.

[0081] The core of the MT-BiLSTM network can be understood as two independent LSTM hidden layers that represent the feature information extracted by the network in the forward time sequence and the feature information extracted in the reverse time sequence, respectively. They are linked to obtain the final prediction model output. The hidden state H of MT-BiLSTM at time t t Contains forward and backward

[0082]

[0083]

[0084]

[0085] In the formula, T is the sequence length, h t-1 is the hidden state of the LSTM layer at the previous moment, h t+1 is the hidden state of the LSTM layer at the next moment, x t is the input data at the current moment, c t-1 is the cell state at the previous moment, c t+1 is the cell state at the next moment.

[0086] After analyzing the denitrification data set to determine the network structure, the optimal parameters of the prediction model are: the number of unit cells in the two LSTM hidden layers is 200, the forgetting probability of the random inactivation layer is 0.1, and the number of training times is 50. The optimization function is the adaptive momentum estimation method (Adam), and the loss functions are the mean absolute error (MAE), the root mean square error (RMSE), and the symmetric mean absolute percentage error (SMAPE).

[0087] Adaptive Moment Estimation, English: Adaptive Moment Estimation, English abbreviation: Adam;

[0088] Mean absolute error, English: Mean Absolute Error, English abbreviation;

[0089] Root Mean Square Error, English: Root Mean Square Error, English abbreviation;

[0090] Symmetric Mean Absolute Percentage Error, English: Symmetric Mean Absolute Percentage Error, English abbreviation: SMAPE.

[0091] The training process of the prediction model uses the back propagation through time (BPTT) to train the model. BPTT uses the deviation between the output value and the true value as the cost function for continuous optimization, and uses the gradient descent algorithm to continuously minimize the cost function to complete. Commonly used gradient descent methods include: SGDM, RMSProp and Adam, etc. Here, the Adam algorithm is used for model gradient descent.

[0092] The Adam algorithm combines the momentum idea of ​​SGDM and the adaptive learning rate idea of ​​RMSProp. The mean of the gradient and the square of the mean are jointly analyzed to finally calculate the speed and learning rate. The Adam algorithm is discussed as follows:

[0093] First, set the initial velocity to v, the learning rate to η, the training cycle to k, and the initial learning rate of the network to α. Initialize the time step t and weight w, and set t = 0. Randomly sample the training sample data of the denitrification system:

[0094] {(x (1) ,y (1) ),(x (2) ,y (2) ),…,(x (m) ,y (m) )} (6)

[0095] Where n is the number of randomly sampled denitrification data training samples, X = {x (1) ,x (2) ,…,x (n)} is the denitrification data set variable input to the model, Y = {y (1) ,y (2) ,…,y (n)} is the corresponding denitrification data output set.

[0096] Secondly, calculate the gradient of the current sampling data:

[0097]

[0098] In the formula, g(x;w) is the learner of the Adam algorithm. Update the current speed v t and η t :

[0099] v t =β 1 v t-1 +(1-β 1 )grad (8)

[0100] η t =β 2 η t-1 +(1-β 2 )grad 2(9)

[0101] In the formula, β 1 and β 2 is the exponential decay rate. Update the current training times k=k+1.

[0102] After updating the parameters, the current speed and current learning rate are corrected:

[0103]

[0104]

[0105] Finally, calculate the parameter w i The update formula is as follows:

[0106]

[0107] In order to prevent the overflow of sample data of the denitrification system, ξ=10 -8 .

[0108] In the present invention, the prediction model based on the MT-BiLSTM network uses the BPTT method to perform gradient calculation, and the Adam algorithm is used to perform gradient descent, so that the parameter adjustment of the network is completed.

[0109] Step 2: Use the predicted value of the prediction model and the real-time NO x The error in emission concentration is used to provide feedback correction to the prediction model.

[0110] The feedback correction is to correct the prediction model by calculating the error between the prediction value of the prediction model and the real-time flue gas emission concentration of the denitration system, thereby improving the stability and accuracy of the denitration process control.

[0111] e(t+1)=y(t)-y m (t+1) (13)

[0112] y p (t+1)=y m (t)+h*e(t+1) (14)

[0113] Where t is the current time, e(t+1) is the error y between the predicted value of the calculation model and the actual flue gas emission value of the denitrification system m (t+1) is the predicted value of the prediction model at the next moment, y(t) is the actual flue gas concentration of the denitrification system, h is the feedback coefficient, and y m (t) is the predicted value of the prediction model at the current moment, y p (t+1) is the output value of the flue gas emission concentration at the next moment after feedback correction.

[0114] Step 3: According to the environmental protection department’s NOx The emission concentration requirement is set as NO x concentration value; secondly, in order to smoothly reach the set value and reduce the fluctuation of the denitrification system, the set value is softened to form a reference trajectory.

[0115] Specifically, in order to avoid the sudden change of input and output of the denitrification system, the output of the multi-index model predictive control process reaches the set value along an expected gentle curve, and the actual output value and the set value at the current moment are subjected to a first-order exponential transformation to obtain a softened smooth reference trajectory;

[0116] The reference trajectory used is a first-order exponential change form, using the actual NO x Emission concentration and NO x The emission concentration set value is transformed into a softened reference trajectory by first-order exponential transformation to obtain the reference value at the future moment:

[0117] y r (k+i)=α j y(k)+(1-α j )y r ,j=1,2,…p (15)

[0118]

[0119] Where k is the current time, j is the control time domain size, α is the softening coefficient, T is the sampling period, τ is the time constant, and y(k) is the actual NO at time k. x Emission concentration value, y r NO x Emission concentration set value, y r (k+j) is the reference trajectory after softening.

[0120] Step 4: Use the differential evolution algorithm to solve the value of the control variable, solve the model input value that makes the output value close to the reference trajectory according to the prediction model, and bring the first solved control value into the denitrification system to complete the stable control of the denitrification system in the cement production process.

[0121] The differential evolution algorithm includes population initialization, mutation, crossover, selection, boundary absorption, and then generates new individuals. The prediction model has 15 input variables, including 9 groups of spray gun ammonia flow u 1 ~u 9 As the control variables of the control system, these 9 variables need to be optimized and solved.

[0122] 4.1, Randomly generate the initial population:

[0123] {X i (0)|X i(0) = [x i,1 ,x i,2 ,...,x i,9 ],i=1,2,...NP} (17)

[0124] x i,j =a j +rand·(b j -a j ) (18)

[0125] Where j = 1, 2, ..., 9, NP represents the population size, X i (0) represents the i-th individual in the initial population, x i,j represents the jth component of the ith individual, a j and b j Respectively represent x ij The upper and lower bounds of the value range, rand represents a random number uniformly distributed in the interval [0,1].

[0126] 4.2, mutation operation:

[0127] The DE algorithm implements mutation operation through the difference method. The basic method is to randomly select two different individuals in the current population, scale their difference vectors, and perform vector operations with other individuals to be mutated to generate new individuals:

[0128] V i (g+1)=X r1 (g)+F·(X r2 (g)-X r1 (g)) (19)

[0129] Where i = 1, 2, ..., NP, i ≠ r 1 ≠r 2 ≠r 3 , r 1 ,r 2 ,r 3 are all random integers in the interval [1, NP], NP is the population size, F is the scaling factor, g represents the evolutionary generation, X i (g) represents the i-th individual in the g-th generation population. After mutation, the g-th generation population generates a new intermediate population V i (g+1).

[0130] 4.3, Crossover operation:

[0131] For the g-th generation population X i (g) and its variant intermediate population V i (g+1) performs crossover operation between individuals:

[0132]

[0133] Where i = 1, 2, ..., NP, j = 1, 2, ..., 9, U i (g+1)=[u i,1 ,u i,2 ,…,u i,9 ] represents the new population of the g+1th generation, u i,j (g+1) and v i,j (g+1) are populations U i (g+1) and V i The jth component in (g+1), CR represents the crossover probability, j rand is a random integer in the interval [1,9], x ij represents the jth component of the ith individual.

[0134] 4.4, Select operation:

[0135] In order to ensure the effectiveness of the solution in the algorithm implementation, the upper and lower limits of the denitrification system variable data are determined according to the actual situation of the cement production plant. If the individual exceeds the constraint range, a new individual is generated by population initialization to replace it. The selection operation is to select individuals to enter the new population by calculating the size of the objective function:

[0136]

[0137] Where i = 1, 2, ..., NP, f is the objective function, U i (g+1) is the intermediate population of the g+1th generation, X i (g) is the population of the gth generation, X i (g+1) is the new generation population generated after the g-th generation population is selected.

[0138] Repeat the above four steps until the conditions are met, and solve the control quantity by minimizing the objective function. Finally, the control quantity obtained by rolling optimization is used to complete the process control of the denitrification system in cement calcination.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-index model predictive control method for cement burning denitrification system, Features: The following steps are involved: Step 1, using historical data in the database to establish an accurate prediction model for flue gas emissions during cement burning; The prediction model is a multi-objective prediction model of flue gas emission concentration in the cement burning process combined with a bidirectional long short-term memory network of a time series. The specific steps of constructing the prediction model include: 1.1, build a bidirectional long short-term memory network; 1.2, introduce the time series into the input layer of the bidirectional long short-term memory network structure to avoid the influence of delay and build a prediction model; 1.3, training the prediction model, inputting the data to be predicted into the prediction model, and obtaining the prediction results of each indicator; Step 2: Use the predicted value of the prediction model and the real-time NO x The error in emission concentration is used to provide feedback correction to the prediction model; Step 3: According to the environmental protection department’s NO x The emission concentration requirement is set as NO x concentration value, and set the NO x The concentration value is softened to form a reference trajectory; Step 4: Use the differential evolution algorithm to solve the value of the control variable, solve the prediction model input value that makes the output value close to the reference trajectory according to the prediction model, and bring the first control value solved into the denitrification system to complete the control of the denitrification system in the cement production process.

2. According to claim 1, a multi-index model predictive control method for a cement burning denitrification system, Features: In step 1, the NO in the cement burning process is first x The generation mechanism of NO was analyzed and 15 x The process variables related to emission concentration and ammonia slip are used as the input of the prediction model. The variables are arranged in chronological order into a time series as the input of the prediction model. In order to achieve multi-objective prediction of flue gas emissions, the MIMO strategy is used to achieve multi-index output of the prediction model and realize simultaneous prediction of NO x Emission concentration and ammonia slip magnitude; By analyzing the influence of NO in the whole cement burning process x The factors that cause and eliminate the problem, combined with the experience and knowledge of the field engineers, the 15 variables finally selected are: Group 1 to Group 9 spray gun ammonia flow rate X 1 ~X 9 、Smoke Room O 2 FeedbackX 10 、Smoke room NO x FeedbackX 11 、Smoke chamber CO feedback X 12 , Decomposition furnace temperature feedback X 13 , coal feeding amount of decomposition furnace X 14 , O 2 FeedbackX 15 .

3. A multi-index model predictive control method for a cement burning denitrification system according to claim 1, It is characterized in that In step 1.2, specifically: The input variable data is integrated into a time series of m-step time periods and introduced into the input layer of the bidirectional long short-term memory network structure. The time series X(t) is expressed as: X 1 =[X 1 (t),X 1 (t+1),L,X 1 (t+m)] (1) X(t)=[X 1 (t)+X 2 (t)+L+X 15 (t)] (2) In the formula, X 1 is the time series corresponding to the ammonia flow rate variable of the first group of spray guns, X 1 (t)~X 15 (t) represents 15 kinds of NO x The time series corresponding to the relevant process variables at time t, X(t) is the time series of the input layer of the bidirectional LSTM model, m is the width of the time series, and t is the current time; The bidirectional long short-term memory network uses two independent LSTM hidden layers to represent the feature information extracted by the network in the forward time sequence and the feature information extracted in the reverse time sequence respectively, and links them to obtain the final prediction model output. The hidden state H of the prediction model at time t is t Contains forward and backward In the formula, T is the sequence length, h t-1 is the hidden state of the LSTM layer at the previous moment, h t+1 is the hidden state of the LSTM layer at the next moment, x t is the input data at the current moment, c t-1 is the cell state at the previous moment, c t+1 is the cell state at the next moment.

4. A multi-index model predictive control method for a cement burning denitrification system according to claim 3, Features: After analyzing the denitrification data set to determine the network structure, the optimal parameters of the prediction model are: the number of unit cells in the two LSTM hidden layers is 200, the forgetting probability of the random inactivation layer is 0.1, the number of training times is 50, the optimization function is the adaptive momentum estimation method, and the loss functions are the mean absolute error, the root mean square error, and the symmetric mean absolute percentage error.

5. According to claim 1, a multi-index model predictive control method for a cement burning denitrification system, Features: In step 2, the actual output is compared with the predicted value at each control step to correct the uncertainty of the prediction model. When the denitrification system has time-varying, model mismatch and interference factors, feedback correction can correct the predicted value in time, so that the optimization is based on a more accurate prediction, thereby improving the robustness of the control system. The feedback correction is to correct the prediction model by calculating the error between the prediction value of the prediction model and the real-time flue gas emission concentration of the denitrification system, thereby improving the stability and accuracy of the denitrification process control: e(t+1)=y(t)-y m (t+1) (13) y p (t+1)=y m (t)+h*e(t+1) (14) Where t is the current time, e(t+1) is the error between the predicted value of the calculation prediction model and the actual flue gas emission value of the denitrification system, and y m (t+1) is the predicted value of the prediction model at the next moment, y(t) is the actual flue gas concentration of the denitrification system, h is the feedback coefficient, and y m (t) is the predicted value of the prediction model at the current moment, y p (t+1) is the output value of the flue gas emission concentration at the next moment after feedback correction.

6. A multi-index model predictive control method for a cement burning denitrification system according to claim 1, Features: In step 3, in order to avoid the sudden change of input and output of the denitration system, the output of the multi-index model predictive control process reaches the set value along an expected gentle curve, and the actual output value and the set value at the current moment are subjected to a first-order exponential transformation to obtain a softened smooth reference trajectory; The reference trajectory used is a first-order exponential change form, using the actual NO x Emission concentration and NO x The emission concentration set value is transformed into a softened reference trajectory by first-order exponential transformation to obtain the reference value at the future moment: y r (k+i)=α j y(k)+(1-α j )y r ,j=1,2,…p (15) Where k is the current time, j is the control time domain size, α is the softening coefficient, T is the sampling period, τ is the time constant, and y(k) is the actual NO at time k. x Emission concentration value, y r NO x Emission concentration set value, y r (k+j) is the reference trajectory after softening.

7. A multi-index model predictive control method for a cement burning denitrification system according to claim 1, Features: In step 4, the differential evolution algorithm includes population initialization, mutation, crossover, selection, boundary absorption, and then generates new individuals. The prediction model has 15 input variables, including 9 groups of spray gun ammonia flow u 1 ~u 9 As the control variables of the control system, these 9 variables need to be optimized and solved; The differential evolution algorithm is used to solve the value of the control variable. The specific algorithm is as follows: 4.1, Randomly generate the initial population: {X i (0)|X i (0)=[x i,1 ,x i,2 ,...,x i,9 ],i=1,2,NP} (17) x i,j =a j +rand·(b j -a j ) (18) In the formula, j = 1, 2, ..., 9, NP represents the population size, X i (0) represents the i-th individual in the initial population, x i，j represents the jth component of the ith individual, a j and b j Respectively represent x ij The upper and lower bounds of the value range, rand represents a random number uniformly distributed in the interval [0,1]; 4.2, mutation operation: The DE algorithm implements mutation operation through the difference method. The basic method is to randomly select two different individuals in the current population, scale their difference vectors, and perform vector operations with other individuals to be mutated to generate new individuals: V i (g+1)=X r1 (g)+F·(X r2 (g)-X r1 (g)) (19) Where i = 1, 2, ..., NP, i ≠ r 1 ≠r 2 ≠r 3 , r 1 , r 2 , r 3 are all random integers in the interval [1, NP], NP is the population size, F is the scaling factor, g represents the evolutionary generation, X i (g) represents the i-th individual in the g-th generation population. After mutation, the g-th generation population generates a new intermediate population V i (g+1); 4.3 Crossover Operation: For the g-th generation population X i (g) and its variant intermediate population V i (g+1) performs crossover operation between individuals: in, i=1,2,...,NP, j=1,2,...,9, U i (g+1)=[u i,1 ,u i,2 ,...,u i,9 ] represents the g+1th generation new population, u i,j (g+1) and v i，j (g+1) are populations U i (g+1) and V i The jth component in (g+1), CR represents the crossover probability, j rand is a random integer in the interval [1,9], x ij represents the jth component of the i-th individual; 4.4, Select operation: In order to ensure the effectiveness of the solution in the algorithm implementation, the upper and lower limits of the denitrification system variable data are determined according to the actual situation of the cement production plant for constraints; if the individual exceeds the constraint range, a new individual is generated by population initialization to replace it; the selection operation is to select the individual entering the new population by calculating the size of the objective function: Where i = 1, 2, ..., NP, f is the objective function, U i (g+1) is the intermediate population of the g+1th generation, X i (g) is the population of the gth generation, X i (g+1) is the new generation population generated after the g-th generation population is selected; Repeat the above four steps until the conditions are met, and solve the control quantity by minimizing the objective function.

Images