Multi-objective predictive control method for SCR denitration process based on fuzzy neural network

Abstract

The present application relates to a SCR denitration process multi-objective predictive control method based on fuzzy neural network, aiming at the biomass cogeneration SCR denitration process, collecting sample data of the denitration process, and constructing a prediction model based on fuzzy neural network. And adopting a multi-objective ideal point method to define the objective function of the SCR denitration process, through rolling optimization of the objective function, the ammonia flow and the ammonia injection valve opening of each control cycle are calculated in real time, the ammonia injection amount is accurately controlled, and multi-objective predictive control is realized. Compared with the traditional SCR mechanism model, the present application does not need prior complex mechanism knowledge, the prediction model based on fuzzy neural network has higher accuracy and reliability; adopting a predictive control algorithm, the ammonia injection amount of the denitration process can be accurately controlled; considering multiple targets, while ensuring that the NOx and NH3 emission concentrations meet the standards, the economic performance of the SCR denitration process is also considered.

Description

Technical Field

[0001] This invention belongs to the field of automatic control of denitrification process in biomass cogeneration, and specifically relates to a multi-objective predictive control method for SCR denitrification process based on fuzzy neural network. Background Technology

[0002] Biomass combined heat and power (CHP) systems are low-carbon, high-efficiency integrated energy systems. With increasing environmental awareness, the nitrogen oxides (NOx) content of biomass CHP systems is rising. X The emissions issue has attracted widespread attention.

[0003] Currently, Selective Catalytic Reduction (SCR) is widely used in the denitrification process of biomass cogeneration. This method involves mixing ammonia (NH3) into the flue gas after combustion, where a selective reduction reaction occurs under the action of a catalyst, producing clean and harmless water and nitrogen, effectively reducing nitrogen oxide emissions. In this process, controlling the amount of ammonia injected is crucial; insufficient ammonia injection leads to incomplete reaction and increased nitrogen oxide (NOx) emissions. X Excessive emissions, particularly excessive ammonia injection, can lead to unnecessary ammonia escape, affecting subsequent system operation and polluting the environment. Therefore, precise control of ammonia injection is necessary to ensure that pollutant emissions meet standards.

[0004] The traditional method for controlling the SCR denitrification process is PID control. While this method is simple in principle and easy to implement, it cannot handle the constraints and time-delay characteristics inherent in the SCR denitrification process, resulting in poor control performance. Furthermore, the SCR denitrification process is a complex physicochemical process, influenced by multiple factors, exhibiting strong nonlinearity, uncertainty, and time-varying characteristics. Therefore, mechanistic models established based on physicochemical principles are difficult to accurately reflect the SCR denitrification process, and traditional control methods based on mechanistic models are no longer adequate for controlling the ammonia injection rate. Moreover, most current control methods for the SCR denitrification process only consider whether the nitrogen oxide (NOx) emission concentration meets the standards, paying less attention to the ammonia emission concentration and the economic performance of the denitrification process. Therefore, accurately controlling the ammonia injection rate in the SCR denitrification process to ensure that pollutant (NOx, NH3) emissions meet standards while achieving economical operation of the denitrification process is quite challenging. Summary of the Invention

[0005] This invention provides a multi-objective predictive control method for SCR denitrification process based on fuzzy neural network. This method takes into account multiple objectives such as the emission of pollutants, such as NOx and NH3, and the economic performance of the denitrification process. It overcomes the nonlinearity, uncertainty and time-varying nature of the biomass cogeneration denitrification process, and can accurately control the amount of ammonia injected, so as to achieve economic operation of the denitrification process while ensuring that pollutant emissions meet the standards.

[0006] The technical solution adopted by the present invention to solve its technical problem is: a multi-objective predictive control method for SCR denitrification process based on fuzzy neural network. The method collects sample data of biomass cogeneration denitrification process, establishes a prediction model based on fuzzy neural network, trains fuzzy neural network using the sample data, and updates the center, width and connection weight of membership function and subsequent network of fuzzy neural network.

[0007] Multiple objectives are selected to set performance functions, predict the state vector of the SCR denitrification process in the future prediction time domain, define the objective function of the SCR denitrification process using the multi-objective ideal point method, continuously optimize the objective function, and calculate the ammonia flow rate and ammonia injection valve opening in real time for each control cycle to achieve multi-objective predictive control.

[0008] Preferably, the sample data at time k consists of a control vector U(k-1) and a state vector X(k-1). The control vector U(k-1) includes ammonia flow rate sample data u1(k-1) and ammonia injection valve opening sample data u2(k-1). The state vector X(k-1) includes NOx emission concentration sample data x1(k-1) and NH3 emission concentration sample data x2(k-1).

[0009] In the prediction model, U(k-1) and X(k-1) are taken as inputs, and X(k) is taken as output.

[0010] That is, u1(k-1) and u2(k-1) together form the control vector U(k-1) = [u1(k-1), u2(k-1)] T The combination of x1(k-1) and x2(k-1) forms the state vector X(k-1) = [x1(k-1), x2(k-1)]. T .

[0011] In this invention, at time k, let the inputs to the prediction model to be built be U(k-1) and X(k-1), and the output be X(k), where U(k-1) = [u1(k-1), u2(k-1)]. T X(k-1) = [x1(k-1), x2(k-1)] T X(k) = [x1(k), x2(k)] Tu1(k-1) corresponds to the ammonia flow rate at time k-1, u2(k-1) corresponds to the ammonia injection valve opening at time k-1, x1(k-1) corresponds to the NOx emission concentration at time k-1, x1(k) corresponds to the NOx emission concentration at time k, x2(k-1) corresponds to the NH3 emission concentration at time k-1, x2(k) corresponds to the NH3 emission concentration at time k, and the superscript T indicates the transpose of the vector;

[0012] During the SCR denitrification process, there are constraints on the emission concentrations of NOx and NH3, ammonia flow rate, and ammonia injection valve opening, as shown in equations (1) and (2):

[0013] x 1,min ≤x1≤x 1,max ,x 2,min ≤x2≤x 2,max (1)

[0014] u 1,min ≤u1≤u 1,max ,u 2,min ≤u2≤u 2,max (2)

[0015] Wherein, equation (1) represents the state constraint, x 1,min Indicates the lower limit of NOx emission concentration, x 1,max Indicates the upper limit of NOx emission concentration, x 2,min Indicates the lower limit of NH3 emission concentration, x 2,max This represents the upper limit of NH3 emission concentration. Equation (2) is the control constraint, u 1,min u represents the lower limit of ammonia flow rate. 1,max u represents the upper limit of ammonia flow rate. 2,min This indicates the lower limit of the ammonia injection valve opening, u 2,max This indicates the upper limit of the ammonia injection valve opening.

[0016] Preferably, the antecedent network includes an input layer, a membership function layer, a rule layer, a consequent layer, and an output layer, and the consequent network includes an input layer and a hidden layer.

[0017] In this invention, the prediction model based on a fuzzy neural network combines the advantages of both neural networks and fuzzy logic, capable of representing both fuzzy and qualitative rules while possessing strong learning capabilities. This prediction model is a black-box model, reflecting the mapping relationship between input and output quantities in the SCR denitrification process.

[0018] In this invention, the input layer of the preamble network has a total of l neurons. When sample data k is input, the input is directly passed to the input layer as the output, as shown in equation (3):

[0019] v i (k), i = 1, 2, ..., l (3)

[0020] Among them, v i The output of the i-th input layer neuron.

[0021] V(k) = [v1(k), v2(k), ..., v l (k)] T =[X(k-1),U(k-1)] T =[x1(k-1),x2(k-1),u1(k-1),u2(k-1)] T V(k) is the output column vector of the input layer of the preamble network, and l is a positive integer representing the number of neurons in the input layer, with a value of l = 4.

[0022] The membership function layer of the preamble network contains l×s neurons. Each neuron calculates the membership degree value corresponding to each input, and the output is given by equation (4):

[0023]

[0024] Where, μ ij (k) represents v i (k) corresponds to s membership values, exp represents the exponential function, and c ij (k) and b ij (k) represent the center and width of the membership function, respectively, and s is a positive integer, representing the number of neurons in the regular layer;

[0025] The rule layer of the preceding network has a total of s neurons. The output of each rule layer neuron is calculated by the multiplication operator as shown in equation (5):

[0026]

[0027] Where, γ j (k) represents the output of the j-th neuron in the rule layer. After defuzzifying the output rule, the normalized output is obtained as equation (6):

[0028]

[0029] Where, θ j (k) represents the normalized output of the j-th neuron in the rule layer;

[0030] The consequent layer of the antecedent network has s×h neurons. This layer passes the consequent parameters obtained from the consequent network to the output layer. (Output is used...) Let h be a positive integer representing the number of neurons in the output layer, and g = 1, 2, ..., h.

[0031] In this invention, the input layer of the consequent network has a total of l+1 neurons. The input of this layer is the input of the input layer of the antecedent network plus a constant value of 1, denoted by v0(k) = 1, which is used to represent the constant term of the consequent part of the fuzzy rule.

[0032] The hidden layer of the consequent network has a total of s×h neurons, and the layer outputs the consequent parameters of the fuzzy rule, as shown in equation (7):

[0033]

[0034] in, This represents the connection weights from the input layer to the hidden layer of the consequent network.

[0035] In this invention, the output layer of the antecedent network has a total of h neurons. This layer weights and sums the outputs of the consequent layer to obtain the output of the entire network as shown in equation (8):

[0036]

[0037] in, This represents the output of the g-th neuron in the output layer.

[0038] In this invention, the fuzzy neural network uses the state variable X(k) (the emission concentrations of NOx and NH3) as the output variable, then h = 2, and is expressed in the form of a state vector.

[0039] Based on the input-output mapping relationship of the fuzzy neural network, the prediction model for the SCR denitrification process in biomass cogeneration can be obtained as Equation (9):

[0040] X(k)=f(X(k-1),U(k-1)) (9)

[0041] Where f represents a prediction model based on a fuzzy neural network.

[0042] In this invention, a fuzzy neural network is trained using sample data from the SCR denitrification process to update the center c of the membership function. ij (k), width b ij (k) and the connection weights of the consequent network

[0043] Sample data consists of input-output pairs selected according to given rules, which include:

[0044] a) The selected sample data should reach a certain scale. Too little data will affect the fitting accuracy of the fuzzy neural network; too much data will lead to a long training time.

[0045] b) The selected sample should include data generated under various operating conditions of the biomass cogeneration system, cover the known fluctuation range as much as possible, and eliminate extreme abnormal data;

[0046] Furthermore, the gradient descent algorithm is used to train the fuzzy neural network, update the network parameters, and define the error function as equation (10):

[0047]

[0048] Among them, y g (k) is the g-th actual output corresponding to the k-th sample data. It is the output of the g-th fuzzy neural network corresponding to the k-th sample data;

[0049] Membership function center c ij (k) and width b ij The update algorithm for (k) is given by equation (11).

[0050]

[0051] Where η is the learning rate of the gradient descent algorithm, and c ij (k) represents the membership function centers corresponding to the k-th sample data, b ij (k) is the width of each membership function corresponding to the k-th sample data, c ij (k-1) represents the center of each membership function corresponding to the (k-1)th sample data, b ij (k-1) is the width of each membership function corresponding to the (k-1)th sample data. Indicates the partial differential symbol;

[0052] Connection weights of the consequent network The update algorithm is given by equation (12):

[0053]

[0054] in, These are the connection weights of the consequent network corresponding to the k-th sample data. These are the connection weights of the consequent network corresponding to the (k-1)th sample data.

[0055] Preferably, the multiple targets include NOx emission concentration, NH3 emission concentration, and the economic performance of the denitrification process; the emission concentrations of NOx and NH3 are used to represent the pollutant emission targets of the biomass cogeneration SCR denitrification process, and the ammonia consumption and the throttling loss of the ammonia injection valve are used to represent the economic performance targets of the SCR denitrification process.

[0056] Preferably, a performance function and a function vector L are set; o = 4 performance functions and function vectors L are set, as shown in equations (13) to (17) respectively:

[0057] L1(X,U)=(x1-r) 2 (13)

[0058] L2(X,U)=x2 (14)

[0059] L3(X,U)=u1 (15)

[0060] L4(X,U)=-u2 (16)

[0061] L(X,U)=[L1(X,U),L2(X,U),L3(X,U),L4(X,U)] T (17)

[0062] Where r is the setpoint for NOx emission concentration in the SCR denitrification process of biomass cogeneration, L1(X,U) represents the target of achieving the setpoint r for NOx emission concentration in the SCR denitrification process, L2(X,U) represents minimizing the NH3 emission concentration in the SCR denitrification process, L3(X,U) represents minimizing the ammonia consumption, L4(X,U) represents minimizing the throttling loss of the ammonia injection valve, which decreases as the opening of the ammonia injection valve increases, and L(X,U) is a function vector.

[0063] Preferably, according to the prediction model (9), the state vector of the SCR denitrification process at time k+j in the future is given by equation (18):

[0064] X(j|k)=f(X(j-1|k),U(j-1|k)),j=1,2…,N (18)

[0065] Where N is the prediction time domain of the SCR denitrification process, and X(j|k)=[x1(j|k),x2(j|k)] T Let U(j-1|k) represent the state vector predicted at time k+j, where U(j-1|k) = [u1(j-1|k), u2(j-|k)]. T , represents the control vector acting on the SCR denitrification process at time k+j-1;

[0066] Preferably, the objective function of the SCR denitrification process is: The objective function for the SCR denitrification process is constructed using the multi-objective ideal point method.

[0067] Combining performance functions (13) to (16) with state constraints (1) and control constraints (2), we construct optimization problem (19) and establish the steady-state ideal point. The performance index function L for the SCR denitrification process is defined using the multi-objective ideal point method. e (X,U), distinguished from the other four performance functions by subscripts, see equation (20), which is a self-balancing multi-objective performance function that integrates multiple objectives from performance function equations (13) to (16); based on equation (20), the objective function for the SCR denitrification process in the prediction time domain is constructed to satisfy equation (21).

[0068]

[0069] L e (X,U)=||L(X,U)-L * ||2 (20)

[0070]

[0071] Where X = f(X,U) represents the condition under which the SCR denitrification control process reaches stability, which is the state condition under which the denitrification process reaches stability; Represents a state sequence. This indicates a control sequence.

[0072] Preferably, the objective function (21) of the SCR denitrification process is optimized by rolling, and the ammonia flow rate and ammonia injection valve opening are obtained in each sampling control cycle. Combining the objective function (21) of the SCR denitrification process, the state prediction equation (18), and the state and control constraints (1) and (2), the following multi-objective optimization problem (22) is solved in each sampling control cycle.

[0073]

[0074] stX(j|k)=f(X(j-1|k),U(j-1|k)) (22b)

[0075] x 1,min ≤x1(j|k)≤x 1,max ,x 2,min ≤x2(j|k)≤x 2,max (22c)

[0076] u 1,min ≤u1(j|k)≤u 1,max ,u 2,min ≤u2(j|k)≤u 2,max (22d)

[0077] X(0|k)=X(k) (22e)

[0078] in, X(0|k) = X(k) represents the optimal solution to the optimization problem (22), and X(0|k) = X(k) represents the initial state conditions of each sampling control cycle.

[0079] In this invention, the ammonia injection rate in the biomass cogeneration SCR denitrification process is controlled by adjusting the ammonia flow rate and the opening of the ammonia injection valve.

[0080] Preferably, the optimal solution at time k is obtained. Afterwards, The first element U * (0|k) acts on the biomass cogeneration SCR denitrification process to adjust the ammonia flow rate and the opening of the ammonia injection valve; in the next sampling control cycle, the emission concentrations of NOx and NH3 are remeasured as the initial state condition X(k) at that moment, and the multi-objective optimization problem (22) is solved again, and so on.

[0081] This invention provides a multi-objective predictive control method for SCR denitrification processes based on fuzzy neural networks. For biomass cogeneration SCR denitrification processes, sample data of the denitrification process is collected, and a predictive model is constructed based on a fuzzy neural network. The objective function of the SCR denitrification process is defined using the multi-objective ideal point method. By continuously optimizing this objective function, the ammonia flow rate and ammonia injection valve opening are calculated in real time for each control cycle, precisely controlling the ammonia injection quantity and achieving multi-objective predictive control. This method ensures that NOx and NH3 emission concentrations meet standards while achieving economical operation of the denitrification process.

[0082] The beneficial effects of this invention are mainly reflected in:

[0083] (1) Compared with the traditional SCR mechanism model, it does not require prior complex mechanism knowledge. The prediction model based on fuzzy neural network has higher accuracy and reliability.

[0084] (2) The predictive control algorithm can accurately control the amount of ammonia injected during the denitrification process;

[0085] (3) Multiple objectives were considered: while ensuring that NOx and NH3 emission concentrations meet the standards, the economic performance of the SCR denitrification process was also considered. Attached Figure Description

[0086] Figure 1 This is a schematic diagram of the SCR denitrification process in biomass cogeneration.

[0087] Figure 2 This is a structural diagram of a fuzzy neural network;

[0088] Figure 3 This is a flowchart of a multi-objective predictive control algorithm for denitrification processes based on fuzzy neural networks. Detailed Implementation

[0089] The main execution part of this invention is implemented and run on the process control computer of the biomass cogeneration SCR denitrification process.

[0090] Figure 1 The diagram shows the SCR denitrification process in a biomass cogeneration plant. The solid lines indicate the flow direction of the flue gas after combustion, the dashed lines indicate the flow direction of ammonia, and the square arrows indicate the flow direction of the gas mixture after mixing with ammonia. The SCR process can be summarized as follows: Flue gas from the biomass burner passes through the economizer and arrives upstream of the SCR reactor. Liquid ammonia in the ammonia tank evaporates and vaporizes, then mixes with the flue gas through valve regulation. Together, they pass through the catalyst layer and undergo a selective catalytic reduction reaction to produce nitrogen and water, which are finally discharged into the atmosphere with the resulting flue gas.

[0091] The application process of this invention is divided into three stages:

[0092] Parameter settings, including fuzzy neural network parameters and predictive controller parameters: In the import interface of the control computer model, import the NOx emission concentration setpoint r of the biomass cogeneration SCR denitrification process and the sample data of the input and output of the denitrification process, and divide the fuzzy variables of the fuzzy neural network input variables; In the predictive controller parameter setting interface, input the prediction time domain N and the sampling control period T;

[0093] Offline debugging: Based on the gradient descent algorithm, a fuzzy neural network is trained using sample data of the input and output of the denitrification process to obtain the optimal membership function and connection weights.

[0094] The system operates online, starting the CPU of the control computer and reading the debugged fuzzy neural network parameters, predictive controller parameters, and NOx emission concentration setpoints. It executes the corresponding algorithm program, calculates the optimal control quantity for each sampling control cycle, and applies it to the SCR denitrification process. The control system adjusts the ammonia flow rate and ammonia injection valve opening in real time based on the calculation results. This process repeats continuously, achieving multi-objective automatic control of the biomass cogeneration SCR denitrification process.

[0095] This invention relates to a multi-objective predictive control method for SCR denitrification processes based on fuzzy neural networks, and the specific implementation includes the following steps:

[0096] 1) Collect sample data on the denitrification process of biomass cogeneration:

[0097] Collect ammonia flow rate sample data u1(k-1), ammonia injection valve opening sample data u2(k-1), NOx emission concentration sample data x1(k-1), and NH3 emission concentration sample data x2(k-1) at time k.

[0098] The sample data at time k consists of the control vector U(k-1) and the state vector X(k-1), where the control vector U(k-1) = [u1(k-1), u2(k-1)]. T The state vector X(k-1) = [x1(k-1), x2(k-1)] T Where u1(k-1) corresponds to the ammonia flow rate at time k-1, u2(k-1) corresponds to the ammonia injection valve opening at time k-1, x1(k-1) corresponds to the NOx emission concentration at time k-1, x2(k-1) corresponds to the NH3 emission concentration at time k-1, and the superscript T indicates the transpose of the vector;

[0099] During the SCR denitrification process, there are constraints on the emission concentrations of NOx and NH3, ammonia flow rate, and ammonia injection valve opening, as shown in equations (1) and (2):

[0100] x 1,min ≤x1≤x 1,max ,x 2,min ≤x2≤x 2,max (1)

[0101] u 1,min ≤u1≤u 1,max ,u 2,min ≤u2≤u 2,max (2)

[0102] Wherein, equation (1) represents the state constraint, x 1,min Indicates the lower limit of NOx emission concentration, x 1,max Indicates the upper limit of NOx emission concentration, x 2,min Indicates the lower limit of NH3 emission concentration, x 2,max This represents the upper limit of NH3 emission concentration. Equation (2) is the control constraint, u 1,min u represents the lower limit of ammonia flow rate. 1,max u represents the upper limit of ammonia flow rate. 2,min This indicates the lower limit of the ammonia injection valve opening, u 2,max Indicates the upper limit of the ammonia injection valve opening;

[0103] 2) Establish a prediction model f based on a fuzzy neural network:

[0104] like Figure 2 As shown, the antecedent network includes an input layer, a membership function layer, a rule layer, a consequent layer, and an output layer. The consequent network includes an input layer and a hidden layer. The antecedent network matches the antecedent part of the fuzzy rule, while the consequent network corresponds to the consequent part of the fuzzy rule. This network structure combines the advantages of both neural networks and fuzzy logic, capable of representing both fuzzy and qualitative rules, and possessing strong learning capabilities.

[0105] The input layer of the preamble network has a total of l neurons. When sample data k is input, the input is directly passed to the input layer as the output, see equation (3):

[0106] v i (k), i = 1, 2, ..., l (3)

[0107] Among them, v i The output of the i-th input layer neuron.

[0108] V(k) = [v1(k), v2(k), ..., v l (k)] T =[X(k-1),U(k-1)] T =[x1(k-1),x2(k-1),u1(k-1),u2(k-1)] T V(k) is the output column vector of the input layer of the preamble network, and l is a positive integer representing the number of neurons in the input layer, with a value of l = 4.

[0109] The membership function layer of the preamble network contains l×s neurons. Each neuron calculates the membership degree value corresponding to each input, and the output is given by equation (4):

[0110]

[0111] Where, μ ij (k) represents v i (k) corresponds to s membership values, exp represents the exponential function, and c ij (k) and b ij (k) represent the center and width of the membership function, respectively, and s is a positive integer, representing the number of neurons in the regular layer;

[0112] The rule layer of the preceding network has a total of s neurons. The output of each rule layer neuron is calculated by the multiplication operator as shown in equation (5):

[0113]

[0114] Where, γ j (k) represents the output of the j-th neuron in the rule layer. After defuzzifying the output rule, the normalized output is obtained as equation (6).

[0115]

[0116] Where, θ j (k) represents the normalized output of the j-th neuron in the rule layer;

[0117] The consequent layer of the antecedent network has s×h neurons. This layer passes the consequent parameters obtained from the consequent network to the output layer. (Output is used...) Indicates that h is a positive integer, representing the number of neurons in the output layer, and g = 1, 2, ..., h;

[0118] The input layer of the consequent network has a total of l+1 neurons. The input of this layer is the input of the input layer of the antecedent network plus a constant value of 1, denoted by v0(k) = 1, which is used to represent the constant term of the consequent part of the fuzzy rule.

[0119] The hidden layer of the consequent network has a total of s×h neurons. This layer outputs the consequent parameters of the fuzzy rule, as shown in equation (7):

[0120]

[0121] in, This represents the connection weights from the input layer to the hidden layer of the consequent network;

[0122] The output layer of the antecedent network has a total of h neurons. This layer sums the weighted outputs of the consequent layer to obtain the output of the entire network as shown in equation (8):

[0123]

[0124] in, This represents the output of the g-th neuron in the output layer. In this invention, the fuzzy neural network uses the state variable X(k) (the emission concentrations of NOx and NH3) as the output variable, so h = 2, and is represented in the form of a state vector.

[0125] Based on the input-output mapping relationship of the fuzzy neural network, the prediction model for the SCR denitrification process in biomass cogeneration can be obtained as Equation (9):

[0126] X(k)=f(X(k-1),U(k-1)) (9)

[0127] Where f represents the prediction model based on the fuzzy neural network;

[0128] 3) Train a fuzzy neural network using sample data from the SCR denitrification process to update the center c of the membership function. ij (k), width b ij (k) and the connection weights of the consequent network

[0129] The fuzzy neural network is trained using the gradient descent algorithm, and the network parameters are updated as follows:

[0130] The error function is defined as equation (10):

[0131]

[0132] Among them, yg (k) is the g-th actual output corresponding to the k-th sample data. It is the output of the g-th fuzzy neural network corresponding to the k-th sample data;

[0133] Membership function center c ij (k) and width b ij The update algorithm for (k) is given by equation (11).

[0134]

[0135] Where η is the learning rate of the gradient descent algorithm, and c ij (k) represents the membership function centers corresponding to the k-th sample data, b ij (k) is the width of each membership function corresponding to the k-th sample data, c ij (k-1) represents the center of each membership function corresponding to the (k-1)th sample data, b ij (k-1) is the width of each membership function corresponding to the (k-1)th sample data. Indicates the partial differential symbol;

[0136] Connection weights of the consequent network The update algorithm is given by equation (12):

[0137]

[0138] in, These are the connection weights of the consequent network corresponding to the k-th sample data. These are the connection weights of the consequent network corresponding to the (k-1)th sample data;

[0139] 4) Set performance functions based on multiple objectives, including NOx emission concentration, NH3 emission concentration, and the economic performance of the denitrification process:

[0140] The emission concentrations of NOx and NH3 are used to represent the pollutant emission targets of the SCR denitrification process in biomass cogeneration, and the ammonia consumption and throttling loss of the ammonia injection valve are used to represent the economic performance targets of the SCR denitrification process. Four performance functions and a function vector L are set, as shown in equations (13) to (17):

[0141] L1(X,U)=(x1-r) 2 (13)

[0142] L2(X,U)=x2 (14)

[0143] L3(X,U)=u1 (15)

[0144] L4(X,U)=-u2 (16)

[0145] L(X,U)=[L1(X,U),L2(X,U),L3(X,U),L4(X,U)] T (17)

[0146] Where r is the setpoint for NOx emission concentration in the SCR denitrification process of biomass cogeneration, L1(X,U) represents the target of achieving the setpoint r for NOx emission concentration in the SCR denitrification process, L2(X,U) represents minimizing the NH3 emission concentration in the SCR denitrification process, L3(X,U) represents minimizing the ammonia consumption, L4(X,U) represents minimizing the throttling loss of the ammonia injection valve, the throttling loss decreases as the opening of the ammonia injection valve increases, and L(X,U) is a function vector;

[0147] 5) Predict the state vector of the SCR denitrification process at time k+j in the future:

[0148] Let N be the predicted time domain of the SCR denitrification process. According to the prediction model (9), the state vector at the future time k+j can be obtained, see equation (18):

[0149] X(j|k)=f(X(j-1|k),U(j-1|k)),j=1,2…,N (18)

[0150] Where X(j|k)=[x1(j|k),x2(j|k)] T Let U(j-1|k) represent the state vector predicted at time k+j, where U(j-1|k) = [u1(j-1|k), u2(j-1|k)]. T , represents the control vector acting on the SCR denitrification process at time k+j-1;

[0151] 6) Construct the objective function for the SCR denitrification process using the multi-objective ideal point method.

[0152] Combining performance functions (13) to (16) with state constraints (1) and control constraints (2), we construct optimization problem (19) and calculate the steady-state ideal point.

[0153]

[0154] Where X = f(X,U) represents the condition under which the SCR denitrification control process reaches stability;

[0155] Define the performance index function L for the SCR denitrification process. e (X,U), see equation (20):

[0156] L e (X,U)=||L(X,U)-L *||2 (20)

[0157] Based on equation (20), the objective function for the SCR denitrification process in the prediction time domain is constructed as equation (21):

[0158]

[0159] in, Represents a state sequence. Indicates a control sequence;

[0160] 7) The objective function (21) of the SCR denitrification process is optimized by rolling, and the ammonia flow rate and ammonia injection valve opening are obtained for each sampling control cycle:

[0161] Combining the objective function (21) of the SCR denitrification process, the state prediction equation (18), and the state and control constraints (1) and (2), the following multi-objective optimization problem (22) is solved in each sampling control cycle:

[0162]

[0163] stX(j|k)=f(X(j-1|k),U(j-1|k)) (22b)

[0164] x 1,min ≤x1(j|k)≤x 1,max ,x 2,min ≤x2(j|k)≤x 2,max (22c)

[0165] u 1,min ≤u1(j|k)≤u 1,max ,u 2,min ≤u2(j|k)≤u 2,max (22d)

[0166] X(0|k)=X(k) (22e)

[0167] in, X(0|k) = X(k) represents the optimal solution to the optimization problem (22), and X(0|k) = X(k) represents the initial state conditions of each sampling control cycle.

[0168] Obtain the optimal solution at time k. Afterwards, The first element U * (0|k) acts on the biomass cogeneration SCR denitrification process to adjust the ammonia flow rate and the opening of the ammonia injection valve; in the next sampling control cycle, the emission concentrations of NOx and NH3 are remeasured as the initial state condition X(k) at that moment, and the multi-objective optimization problem (22) is solved again, and so on.

[0169] like Figure 3 The flowchart shown is the algorithm flowchart of this invention. First, a fuzzy neural network is trained using sample data from the SCR denitrification process to establish a prediction model. Then, parameters such as N, T, and r are initialized. Based on the established prediction model, future NOx and NH3 emission concentrations are predicted. A multi-objective ideal point method is used to construct the objective function of the SCR denitrification process. This objective function is then continuously optimized, and the ammonia flow rate and ammonia injection valve opening for the current control cycle are calculated in real time and applied to the SCR denitrification process. In the next control cycle, the NOx and NH3 emission concentrations are remeasured, and this process is repeated continuously.

Claims

1. A multi-objective predictive control method for SCR denitrification process based on fuzzy neural network, characterized in that: The method collects sample data of the denitrification process in biomass cogeneration, establishes a prediction model based on a fuzzy neural network, trains the fuzzy neural network using the sample data, and updates the center and width of the membership function and the connection weights of the consequent network of the fuzzy neural network. Select multiple targets, set performance functions, and configure the performance functions and function vectors. , ， ， ， ， ， in, These are the state vector and the control vector, respectively. The setpoint for NOx emission concentration during the SCR denitrification process in biomass cogeneration. and These are sample data for NOx emission concentration and NH3 emission concentration, respectively. and These are sample data for ammonia flow rate and sample data for ammonia injection valve opening, respectively. This indicates that the NOx emission concentration during the SCR denitrification process has reached the set value. The goal, This represents minimizing the NH3 emission concentration during the SCR denitrification process. This represents minimizing the consumption of ammonia. This indicates minimizing the throttling loss of the ammonia injection valve; Predicting the state vector of the SCR denitrification process in the future time domain; the future state vector of the SCR denitrification process. The state vector at time t is Where N is the prediction time domain for the SCR denitrification process, , indicating in The first time predicted The state vector at time t, This indicates the action taken during the SCR denitrification process. Control vector at time; The objective function of the SCR denitrification process is defined using the multi-objective ideal point method. Establish steady-state ideal point Define the performance index function of the SCR denitrification process. , ,satisfy ， in, Represents a state sequence. Indicates a control sequence; ， in, This indicates that the SCR denitrification control process has reached a stable condition. This indicates the lower limit of NOx emission concentration. This indicates the upper limit of NOx emission concentration. This indicates the lower limit of NH3 emission concentration. This indicates the upper limit of NH3 emission concentration. This indicates the lower limit of ammonia flow rate. This indicates the upper limit of ammonia flow rate. This indicates the lower limit of the ammonia injection valve opening. Indicates the upper limit of the ammonia injection valve opening; The objective function is continuously optimized, and the ammonia flow rate and ammonia injection valve opening are calculated in real time for each control cycle to achieve multi-objective predictive control.

2. The multi-objective predictive control method for SCR denitrification process based on fuzzy neural network according to claim 1, characterized in that: The The sample data at time t is the control vector. and state vector Control vector Including ammonia flow rate sample data Sample data on ammonia injection valve opening State vector Includes NOx emission concentration sample data and NH3 emission concentration sample data ; In the prediction model, with and Input is, output is .

3. The multi-objective predictive control method for SCR denitrification process based on fuzzy neural network according to claim 1, characterized in that: The antecedent network of the fuzzy neural network includes an input layer, a membership function layer, a rule layer, an consequent layer, and an output layer. The consequent network includes an input layer and a hidden layer.

4. The multi-objective predictive control method for SCR denitrification process based on fuzzy neural network according to claim 2, characterized in that: Multiple targets include NOx emission concentration, NH3 emission concentration, and the economic performance of the denitrification process.

5. The multi-objective predictive control method for SCR denitrification process based on fuzzy neural network according to claim 1, characterized in that: Rolling optimization of the objective function: ， ， ， ， ， in, This represents the optimal solution to the optimization problem. This indicates the initial state conditions for each sampling control cycle.

6. The multi-objective predictive control method for SCR denitrification process based on fuzzy neural network according to claim 5, characterized in that: Get the current optimal solution at time 1 Afterwards, The first element This process is applied to the SCR denitrification process in biomass cogeneration, adjusting the ammonia flow rate and the opening of the ammonia injection valve; in the next sampling control cycle, the emission concentrations of NOx and NH3 are remeasured as the initial state conditions at that moment. Then, the multi-objective optimization problem is solved again, and the process is repeated cyclically.

Images