Method for establishing precise ammonia injection fuzzy control model based on LSTM neural network

Abstract

The invention provides a method for establishing a precise ammonia injection fuzzy control model based on an LSTM neural network. The method comprises the following steps: S1, analyzing historical data, cleaning, designing a membership function and formulating a fuzzy rule table; s2, constructing a control model by using an LSTM neural network algorithm to obtain an error between a predicted value and an actual value and an error change rate; and S3, comparing a prediction result of the model with an actual output result. According to the invention, an advanced neural network technology and a flexible fuzzy control strategy are fused, so that accurate adjustment of ammonia spraying of the SCR denitration system is realized. A fuzzy expert control system is built, an error, a change rate range of the error and valve opening variation corresponding to different errors and error change rates are obtained by analyzing historical data of ammonia water adjustment, and a fuzzy discourse domain and a fuzzy rule table are designed in combination with a fuzzy control principle and a system working condition. And fuzzy reasoning and fuzzy judgment are carried out on real-time data of the system, and the increment of the ammonia water regulating valve under the current working condition is calculated and regulated.

Description

Technical Field

[0001] The present invention relates to the field of ammonia injection treatment, and in particular to a method for establishing a precise ammonia injection fuzzy control model based on LSTM neural network. Background Art

[0002] In terms of achieving ultra-low emissions of nitrogen oxides, steel plants currently use flue gas SCR denitrification technology to further remove NOx from flue gas, thereby achieving NOx emission standards. SCR technology is a process in which a reducing agent (NH3 or urea, etc.) reacts with NOx in flue gas through a catalyst to generate some environmentally friendly gases to achieve NOx removal. SCR devices generally have a simple structure, high denitrification efficiency, easy maintenance, and reliable stability.

[0003] SCR denitrification is carried out in the presence of a catalyst, and the activity of the catalyst for the reaction is closely related to the temperature. Due to the on-site process characteristics, the SCR catalyst has an optimal active temperature range. In this temperature range, the catalyst can efficiently catalyze the chemical reaction between ammonia and nitrogen oxides to achieve efficient denitrification. However, due to the unstable on-site gas pressure, the gas flow rate fluctuates greatly, which in turn causes drastic changes in the hot blast furnace temperature, the temperature after the furnace and the SCR inlet temperature. These temperature instabilities have a serious impact on the efficiency and performance of SCR denitrification. At present, there are still the following problems that need to be solved urgently:

[0004] 1. Difficult to adapt to operating condition fluctuations: Since the gas required in the process comes from the blast furnace, it is unstable. Therefore, frequent and drastic changes in gas pressure will lead to huge fluctuations in gas flow, which in turn cause drastic changes in the hot blast furnace temperature, the temperature after the furnace, and the SCR inlet temperature. Therefore, it is necessary to design a control method to achieve stable control of gas flow, and then stably control each temperature within the set range.

[0005] 2. The control method has hysteresis: Due to the large hysteresis in the data collection of the SCR denitration system, the traditional feedback control method cannot avoid this large delay. Therefore, it is still necessary to establish an intelligent algorithm model that can accurately predict the outlet NOx concentration in a future period of time in real time. On the basis of comparing the SCR inlet temperature and the target value to make feedback adjustments, the feedforward predictive control is used to timely reduce errors and eliminate interference, so that the temperature can be stably controlled above and below the target value. Summary of the invention

[0006] The purpose of the present invention is to solve the shortcomings of the prior art and propose a method for establishing a precise ammonia injection fuzzy control model based on LSTM neural network.

[0007] To achieve the above purpose, the present invention adopts the following technical solution: a method for establishing a precise ammonia injection fuzzy control model based on LSTM neural network, and the specific steps of the method are as follows:

[0008] S1: Analyze historical data and clean it up, design membership functions and formulate fuzzy rule tables;

[0009] Conduct a detailed analysis and cleanup of the historical data of the SCR denitrification system, annotate the operating conditions, and establish a complete historical data set to provide basic data support; based on the analysis of historical data and the application of operating condition identification technology, determine the initial value of the fuzzy control input and output domain, and preliminarily design the membership function and fuzzy rule table for different operating conditions;

[0010] S2: Use the LSTM neural network algorithm to build a control model and obtain the error and error change rate between the predicted value and the actual value;

[0011] Clean and normalize the data to prepare an input format suitable for the LSTM model; design the LSTM neural network structure and optimize the number of layers and parameters of the network; extract features from the historical time series data in step S1 to improve the prediction ability of the model; use the processed data to train the LSTM model to ensure that the model can effectively capture the dynamic characteristics of the SCR denitrification system; use the trained LSTM model to predict the future state of the SCR denitrification system; calculate the error between the predicted value and the actual value and the error change rate;

[0012] S3: Compare the model’s prediction results with the actual output results;

[0013] The future state of the SCR denitrification system is predicted through the LSTM model trained in step S2; the prediction results are generated through the LSTM model and compared with the actual output to obtain the future error and its error change rate; the fuzzy control system is used to perceive the current error and error change rate, update the controller parameters, and correct and compensate the controller through the fuzzy control system to achieve the purpose of advance regulation, so that the system can respond to future changes in advance and reduce the adverse effects of hysteresis.

[0014] Preferably, in step S1, the historical data of the SCR denitrification system includes outlet flue NOx concentration, ammonia flow rate, inlet flue NOx concentration, ammonia delivery pump frequency, inlet flue flue gas flow rate, outlet flue flue gas flow rate, SCR inlet catalyst temperature, ammonia evaporator temperature, outlet flue gas O2 concentration, second layer catalyst temperature and booster fan operating current frequency feedback data.

[0015] Preferably, in step S1, the basis for determining the initial value of the fuzzy control input and output domain is based on the historical data of the SCR denitrification system after removing the abnormal operating conditions of shutdown calibration and backwashing, assuming that the input and output data conform to the normal distribution, calculating the mean and standard deviation, and using the mean plus or minus 3 times the standard deviation to obtain the upper and lower limits.

[0016] Preferably, in step S1, the membership function is a mathematical tool used to describe the degree of relationship between the input variable and the fuzzy set; the fuzzy rule table is a set of rules that describe the fuzzy relationship between the input variable and the output variable in natural language or symbolic form.

[0017] Preferably, in step S2, the model deep learning uses Adam optimization algorithm to perform model training.

[0018] Preferably, in step S3, the correction is performed by calculating the error between the actual output value and the set value, using the LSTM model to predict future errors, and calculating the error change rate. The adjustment factor is generated by combining the current error, the predicted error and the error change rate, and finally the correction value is calculated for adjusting the control parameters.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: the method of the present invention integrates advanced neural network technology and flexible fuzzy control strategy to achieve precise regulation of ammonia injection in the SCR denitrification system. A fuzzy expert control system is built, and the error and its change rate range and the valve opening change corresponding to different errors and error change rates are obtained by analyzing the historical data of ammonia regulation. A two-dimensional fuzzy controller is selected, and a fuzzy domain and fuzzy rule table are designed in combination with fuzzy control principles and system operating conditions, and a suitable membership function is selected. Then, by performing fuzzy reasoning and fuzzy judgment on the real-time data of the system, the increment of the ammonia regulating valve under the current operating conditions is calculated and adjusted. Based on the high-precision prediction of the LSTM neural network model, the system can accurately predict the future change trend of NOx on the basis of real-time performance, providing a reliable data basis for feedforward control. The model predictive control rolling optimization mechanism is used to update the fuzzy controller parameters, so that the system can dynamically adapt to changing operating conditions during operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a schematic diagram of the present invention. DETAILED DESCRIPTION

[0021] In order to provide a further understanding of the purpose, structure, features, and functions of the present invention, the following detailed description is given in conjunction with the embodiments.

[0022] Please refer to Figure 1The present invention provides a method for establishing a precise ammonia injection fuzzy control model based on LSTM neural network, and the specific steps of the method are as follows:

[0023] S1: Analyze historical data and clean it up, design membership functions and formulate fuzzy rule tables;

[0024] The historical data of the SCR denitrification system is analyzed and cleaned in detail, and the operating conditions are marked to establish a complete historical data set to provide basic data support; based on the historical data analysis and the application of operating condition identification technology, the initial values ​​of the fuzzy control input and output domains are determined, and the membership functions and fuzzy rule tables for different operating conditions are preliminarily designed.

[0025] The cleaning process of historical data is as follows: first fill in missing data values, and use the mean value when reading null values; then process outliers to remove measurement errors and instrument calibration data; finally, use an exponential coefficient of 0.9 to smooth the data.

[0026] The operating conditions mainly include calibration and backblowing of the flue inlet / outlet, flameout of the hot blast furnace, production shutdown, and large fluctuations in gas supply. The above operating conditions are marked according to the time series data characteristics such as inlet CEMS flue gas NOx content, inlet CEMS flue gas flow, inlet CEMS flue gas O2 content, outlet CEMS flue gas NOx content, outlet CEMS flue gas flow, outlet CEMS flue gas O2 content, hot blast furnace post-furnace temperature, catalyst inlet temperature, gas valve opening, gas pressure, etc.

[0027] The basis for determining the initial value of the fuzzy control input and output domain is as follows: Based on the historical time series data in step S1, after removing abnormal operating conditions such as shutdown calibration and backwashing, it is assumed that the input and output data conform to the normal distribution, and the mean and standard deviation are calculated. The upper and lower limits are obtained by adding or subtracting 3 times the standard deviation from the mean.

[0028] Including the mean NOX concentration at the inlet at 1min, 5min, and 10min, as well as the variance of NOX concentration at the inlet at 1min, 5min, and 10min.

[0029] The membership function is a mathematical tool used to describe the degree of relationship between the input variables and the fuzzy sets. The input variables are usually divided into multiple fuzzy sets, and each fuzzy set corresponds to a membership function.

[0030] This function defines the membership value corresponding to the input variable under different values, that is, the degree to which the input value belongs to a certain fuzzy set. The value range of the membership function is usually between [0,1], where 0 means that it does not belong to the fuzzy set at all, 1 means that it completely belongs to the fuzzy set, and the intermediate value indicates the degree to which it partially belongs. [It can be simply understood as the probability that a certain value belongs to each of the divided fuzzy sets. For example, if the NOx outlet concentration error is 5, what is the membership (probability) of "large error"? What is the membership (probability) of "very large error"? ]

[0031] A fuzzy rule table is a set of rules that describe the fuzzy relationship between input variables and output variables in natural language or symbolic form. In a fuzzy control system, since the input and output variables are often fuzzy and imprecise, a fuzzy rule table is needed to map these fuzzy inputs to fuzzy outputs, and then obtain specific control quantities through defuzzification.

[0032] The rules in the fuzzy rule table usually take the form of "IF-THEN", that is, "if the input satisfies a certain fuzzy condition, the output is a certain fuzzy value." Specifically, a fuzzy rule can be expressed as:

[0033] IF input 1 is A1 AND input 2 is A2...AND input n is An

[0034] THEN the output is B

[0035] Among them, A1, A2, ..., An are fuzzy sets of input variables, and B is a fuzzy set of output variables. These fuzzy sets are defined by membership functions, which indicate the degree to which the input or output variable belongs to a fuzzy set. [For example, if "error" is "large" and "error derivative" is "large" then "proportional coefficient Kd" is "large", which can be understood as the outlet NOx is close to the target value, increasing the damping].

[0036] S2: Use the LSTM neural network algorithm to build a control model and obtain the error and error change rate between the predicted value and the actual value;

[0037] Clean and normalize the data to prepare an input format suitable for the LSTM model; design the LSTM neural network structure and optimize the number of network layers and parameters; extract features from the historical time series data in step S1 to improve the model's prediction ability; use the processed data to train the LSTM model to ensure that the model can effectively capture the dynamic characteristics of the SCR denitrification system; use the trained LSTM model to predict the future state of the SCR denitrification system; calculate the error between the predicted value and the actual value and the error change rate.

[0038] Error formula = current NOx real-time feedback value - NOx target value:

[0039] e(t)=y predicted -y stepoint

[0040] Error change rate = NOx error feedback value at n moments in the future - NOx error feedback value at the current moment:

[0041] Δe(t)=e(t+n)-e(t)

[0042] A regulating factor α is introduced into the fuzzy controller, which reflects the weight of the error and the error change rate in the control. The output control quantity of the fuzzy control system based on LSTM neural network prediction is for:

[0043]

[0044] In order to make the control system have better anti-interference ability, a controller correction link is added. The correction link has three inputs, namely the actual output value at the current moment, the set value and the predicted value at the next moment estimated by the LSTM neural network model; the output is the correction value u2 of the main controller and the adjustment factor α.

[0045] The calibration is as follows:

[0046] Calculate the error between the actual output value and the set value:

[0047] e(t)=y predicted -y stepoint

[0048] The difference between the value predicted by the LSTM model for a period of time in the future and the set value:

[0049] e predicted (t+18)=y predicted -y stepoint

[0050] The error change rate indicates the change in error during this period of time:

[0051] Δe(t)=e(t+18)-e(t)

[0052] The adjustment factor is generated by combining the current error, the predicted error, and the error change rate:

[0053] α′=sigmod(K d ·Δe(t)+K f ·e predicted (t+18)

[0054] Where Kd and Kf are the gain coefficients of error and error change rate, respectively, and are initialized and set according to the on-site operating conditions;

[0055] Calculate the correction value using the calculated adjustment factor:

[0056]

[0057] S3: Compare the model’s prediction results with the actual output results;

[0058] The future state of the SCR denitrification system is predicted through the LSTM model trained in step S2; the prediction results are generated through the LSTM model and compared with the actual output to obtain the future error and its error change rate; the fuzzy control system is used to perceive the current error and error change rate, update the controller parameters, and correct and compensate the controller through the fuzzy control system to achieve the purpose of advance regulation, so that the system can respond to future changes in advance and reduce the adverse effects of hysteresis.

[0059] Through the combination of LSTM neural network model and fuzzy control system, this study realizes the advance regulation of NOx emissions in steel plants. The LSTM model provides future prediction information, and the fuzzy control system dynamically optimizes the control parameters according to the prediction error and current error, generates correction signals and adjusts the actuators, and finally realizes real-time control of NOx emissions, reduces the adverse effects of hysteresis, and improves control accuracy and stability.

[0060] Preferably, in step S1, the historical data of the SCR denitrification system includes outlet flue NOx concentration, ammonia flow rate, inlet flue NOx concentration, ammonia delivery pump frequency, inlet flue flue gas flow rate, outlet flue flue gas flow rate, SCR inlet catalyst temperature, ammonia evaporator temperature, outlet flue gas O2 concentration, second layer catalyst temperature and booster fan operating current frequency feedback data.

[0061] Preferably, in step S1, the membership function is a mathematical tool used to describe the degree of relationship between the input variable and the fuzzy set; the fuzzy rule table is a set of rules that describe the fuzzy relationship between the input variable and the output variable in natural language or symbolic form.

[0062] Preferably, in step S2, the model deep learning uses Adam optimization algorithm to perform model training.

[0063] Preferably, in step S3, the correction is performed by calculating the error between the actual output value and the set value, using the LSTM model to predict future errors, and calculating the error change rate. The adjustment factor is generated by combining the current error, the predicted error and the error change rate, and finally the correction value is calculated for adjusting the control parameters.

[0064] The present invention has been described by the above-mentioned relevant embodiments, however, the above-mentioned embodiments are only examples for implementing the present invention. It must be pointed out that the disclosed embodiments do not limit the scope of the present invention. On the contrary, changes and modifications made without departing from the spirit and scope of the present invention are all within the scope of patent protection of the present invention.

Claims

1. A method for establishing a precise ammonia injection fuzzy control model based on LSTM neural network, characterized in that: The specific steps of the method are as follows: S1: Analyze historical data and clean it up, design membership functions and formulate fuzzy rule tables; Conduct a detailed analysis and cleanup of the historical data of the SCR denitrification system, annotate the operating conditions, and establish a complete historical data set to provide basic data support; based on the analysis of historical data and the application of operating condition identification technology, determine the initial value of the fuzzy control input and output domain, and preliminarily design the membership function and fuzzy rule table for different operating conditions; S2: Use the LSTM neural network algorithm to build a control model and obtain the error and error change rate between the predicted value and the actual value; Clean and normalize the data to prepare an input format suitable for the LSTM model; design the LSTM neural network structure and optimize the number of layers and parameters of the network; extract features from the historical time series data in step S1 to improve the model's prediction ability; use the processed data to train the LSTM model to ensure that the model can effectively capture the dynamic characteristics of the SCR denitrification system; Apply the trained LSTM model to predict the future state of the SCR denitrification system; Calculate the error between the predicted value and the actual value and the error change rate; S3: Compare the model’s prediction results with the actual output results; The future state of the SCR denitration system is predicted by the LSTM model trained in step S2; the prediction result is generated by the LSTM model and compared with the actual output to obtain the future error and its error change rate; Use the fuzzy control system to sense the current error and error change rate, update the controller parameters, and correct and compensate the controller through the fuzzy control system to achieve the purpose of advance regulation, so that the system can respond to future changes in advance and reduce the adverse effects of hysteresis.

2. The method for establishing a precise ammonia injection fuzzy control model based on an LSTM neural network as claimed in claim 1, characterized in that: In step S1, the historical data of the SCR denitrification system includes the outlet flue NO X concentration, ammonia flow, inlet flue NO X Concentration, ammonia delivery pump frequency, inlet flue gas flow rate, outlet flue gas flow rate, SCR inlet catalyst temperature, ammonia evaporator temperature, outlet flue gas O2 concentration, second layer catalyst temperature and booster fan operating current frequency feedback data.

3. The method for establishing a precise ammonia injection fuzzy control model based on an LSTM neural network as claimed in claim 1, characterized in that: In step S1, the basis for determining the initial value of the fuzzy control input and output domain is based on the historical data of the SCR denitrification system after removing the abnormal operating conditions of the shutdown calibration backwash, assuming that the input and output data conform to the normal distribution, calculating the mean and standard deviation, and using the mean plus or minus 3 times the standard deviation to obtain the upper and lower limits.

4. The method for establishing a precise ammonia injection fuzzy control model based on an LSTM neural network as claimed in claim 1, characterized in that: In step S1, the membership function is a mathematical tool used to describe the degree of relationship between the input variable and the fuzzy set; the fuzzy rule table is a set of rules that describe the fuzzy relationship between the input variable and the output variable in natural language or symbolic form.

5. The method for establishing a precise ammonia injection fuzzy control model based on an LSTM neural network as claimed in claim 1, characterized in that: In step S2, the model deep learning uses the Adam optimization algorithm to perform model training.

6. The method for establishing a precise ammonia injection fuzzy control model based on an LSTM neural network as claimed in claim 1, characterized in that: In step S3, the correction is performed by calculating the error between the actual output value and the set value, predicting the future error using the LSTM model, and calculating the error change rate. The adjustment factor is generated by combining the current error, the predicted error and the error change rate, and finally the correction value is calculated for adjusting the control parameters.

Images