Multi-loop quasi-diagonal recursive neural network PID control method for municipal solid waste incineration process
By constructing a multi-loop quasi-diagonal recurrent neural network PID controller, the problem of multivariable coupling in the process of urban solid waste incineration was solved, realizing synchronous control of multiple variables and parameter self-tuning, and improving the stability and adaptability of the controller.
Patent Information
- Application Number
- CN202111557331.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-18
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2041-12-18
AI Technical Summary
In the process of urban solid waste incineration, multiple variables are coupled and difficult to control synchronously. Existing single-loop PID controllers cannot meet the control requirements of multiple controlled variables, and the parameter tuning of multi-loop PID controllers is complicated.
A multi-loop quasi-diagonal recurrent neural network PID controller for urban solid waste incineration processes was designed. A multi-input multi-output Takagi-Sugeno fuzzy neural network controlled object model was constructed through data-driven construction, and parameter self-tuning was performed by combining the quasi-diagonal recurrent neural network model to construct a multi-loop PID controller with self-feedback channels and interconnection channels.
It achieves synchronous control of multiple variables during the incineration of urban solid waste, improves the self-tuning capability and stability of the controller, and is suitable for multivariable control of nonlinear systems.
Smart Images

Figure CN114527646B_ABST
Abstract
Description
Technical Field
[0001] In view of the problem that multi-variable coupling and synchronization control are difficult in municipal solid waste incineration, the present invention designs a multi-loop quasi-diagonal recursive neural network PID controller for the municipal solid waste incineration process. The controller analyzes the influencing factors of the model and extracts the key controlled quantities and operation quantities, constructs a multi-input and multi-output Takagi-Sugeno fuzzy neural network controlled object model based on data drive, constructs a quasi-diagonal recursive neural network model with self-feedback channels and interconnection channels and uses it for parameter self-tuning of multi-loop PID control, thereby solving the multi-variable control problem in the municipal solid waste incineration process. Background Art
[0002] Municipal solid waste incineration has become one of the main ways to dispose of municipal solid waste because of its advantages of harmlessness, resource utilization and volume reduction. Municipal solid waste incineration is a typical multi-input and multi-output industrial process with complex mechanism reaction. There is a serious coupling between multiple operating quantities and controlled quantities, making it difficult to explore control rules. Studying the control method of the municipal solid waste incineration process is of great significance to improving its process intelligence, improving its operating efficiency, reducing pollutant emissions and promoting the recycling of municipal solid waste resources.
[0003] Research shows that more than 90% of the current controls in the field of industrial control are implemented using PID control, which has the advantages of simple structure, good robustness, and mature parameter setting technology. Some scholars have improved the PID controller based on actual applications to enhance its ability to solve complex industrial control problems. Although the above PID controllers have good performance in dynamic response and set value tracking, the problems they target are mainly single-variable control problems in the industrial field, and the control target can be achieved by constructing a single-loop PID control. However, when facing the urban solid waste incineration process with many uncertain characteristics such as multiple objects, multiple variables, and strong coupling, it is impossible to meet the control requirements of synchronously controlling multiple controlled quantities using only a single-loop PID controller.
[0004] In response to the multivariable control problem of complex industrial processes, some scholars have built multi-loop PID controllers based on PID control. Multi-loop PID controllers meet the multivariable control needs of complex industries. However, with the increase of loops, the parameters that need to be adjusted for multiple controllers also increase, which puts higher requirements on the parameter adjustment of the system. In recent years, with the development of artificial intelligence theory and technology, the methods used in industrial control are also being innovated. The emergence of artificial neural networks provides a better solution for the self-tuning of parameters of multi-loop PID controllers. When facing complex industrial control problems, multi-loop PID controllers based on neural networks can not only self-tune the parameters of multiple controllers through their own computing power and nonlinear approximation capabilities, but also improve the stability of controllers in the face of highly uncertain industrial processes through their own learning ability.
[0005] Multivariable coupling is one of the key issues to be solved in building a multi-loop PID controller. Neural network PID control has the ability to self-tune parameters and is widely used in controller design. Among them, for the decoupling control of multiple variables, recursive neural networks have stronger expression ability and the ability to process transient information, and are more suitable for multivariable control problems of nonlinear systems. The recursive neural network feeds back the hidden layer nodes of the feedforward network to the nodes of the previous layer, or performs self-feedback on the nodes of the current layer. It overcomes the problem that the multi-layer feedforward network turns dynamic modeling into static modeling. It has a simple structure and less network calculation, and is more suitable for multivariable decoupling control. Research on neural networks in recent years has shown that there are tight and loose connections between neurons, and the diversification of connection methods implies more response methods. The quasi-diagonal recursive neural network is a local recursive neural network constructed by imitating the interconnection characteristics of biological neurons. It is a recursive neural network evolved on the basis of the diagonal recursive neural network. The quasi-diagonal recursive neural network not only constructs a self-feedback channel on the hidden layer neurons, but also adds an interconnection channel between the hidden layer neurons, which has a stronger information mining ability. Unlike fully connected recursive neural networks, quasi-diagonal recursive neural networks only establish connections between adjacent hidden layer neurons, and the recursive weight matrix they construct is a quasi-diagonal matrix. Quasi-diagonal recursive neural networks reduce the number of connection weights, thereby greatly reducing the amount of network calculations, speeding up the learning speed of the network, and being more suitable for real-time control. In summary, quasi-diagonal recursive neural networks have powerful mapping, memory, and dynamic response capabilities. The recursive channels they construct can capture the dynamic characteristics of the system in real time, and are more suitable for multivariable control problems of nonlinear systems.
[0006] According to the above analysis, the present invention designs a multi-loop quasi-diagonal recursive neural network PID controller for the urban solid waste incineration process based on the characteristics of the urban solid waste incineration process. First, the control process characteristics are analyzed and the key operating variables and controlled variables are selected; then, a multi-input and multi-output Takagi-Sugeno fuzzy neural network controlled object model is constructed through data driving; then, a multivariable PID controller based on a quasi-diagonal recursive neural network is constructed; finally, the effectiveness of the controller is verified through process data from a solid waste incineration plant in Beijing. Summary of the invention
[0007] The present invention obtains a multi-loop quasi-diagonal recursive neural network PID controller for the municipal solid waste incineration process. The controller analyzes the influencing factors of the model, constructs a multi-input and multi-output Takagi-Sugeno fuzzy neural network controlled object model based on data drive, constructs a quasi-diagonal recursive neural network model with self-feedback channels and interconnection channels and uses it for parameter self-tuning of multi-loop PID control, solving the multivariable control problem in the municipal solid waste incineration process.
[0008] The present invention adopts the following technical solutions and implementation steps:
[0009] The multi-loop quasi-diagonal recursive neural network PID controller method for municipal solid waste incineration process includes the following steps:
[0010] 1. A multi-loop quasi-diagonal recursive neural network PID control method for a municipal solid waste incineration process, characterized in that it comprises the following steps:
[0011] (1) Data collection of municipal solid waste incineration: The key operating variables of the municipal solid waste incineration process are collected through a one-way isolation system: drying grate air flow (left 1, right 1, left 2, right 2), combustion grate 1 section air flow (left 1, right 1, left 2, right 2), combustion grate 2 section air flow (left 1, right 1, left 2, right 2), burnout grate air flow (left, right), primary air total flow, secondary air flow, drying grate speed (left inner, right inner, left outer, right outer), combustion grate 1 section speed (left inner, right inner, left outer, right outer), combustion grate 2 section speed (left inner, right inner, left outer, right outer) and burnout grate speed (left inner, right inner), and the key controlled variables are collected: main steam flow, furnace temperature and flue gas oxygen content;
[0012] (2) Control characteristics analysis: using Pearson correlation coefficient ρ αβ For the manipulated variable α l and the controlled variable β l The correlation between them is evaluated and calculated as follows:
[0013]
[0014] In the formula, N is the total number of samples, and the total primary air flow, secondary air flow, and drying grate speed are selected as key operating variables based on the calculation results;
[0015] (3) Data preprocessing: The collected data is preprocessed by eliminating abnormal data and normalizing the data. The calculation steps are as follows:
[0016] ① Eliminate abnormal data: Use the 3σ criterion to eliminate abnormal data. First, set the dimension of sample data to q×K, where q is the number of controlled variables and K is the total number of samples. Indicates that is the sample mean, and its standard deviation σ s The calculation method is:
[0017]
[0018] when The corresponding residual error If the following conditions are met, Perform a culling operation, which is calculated as:
[0019]
[0020] ②Data normalization: Sample data Normalize and calculate the normalized data samples The calculation method is as follows:
[0021]
[0022] in, is the sample vector of variable s. After normalization, the range of each sample data is between [0,1];
[0023] (4) Design of controlled object model: A multi-input and multi-output Takagi-Sugeno fuzzy neural network controlled object model was constructed through data driving, which includes five layers: input layer, membership function layer, rule layer, consequent layer and output layer. Its mathematical description is as follows:
[0024] ① Input layer: This layer has n neurons, n is 3, and its function is to transmit the input value. When the kth sample enters, the output of the input layer is:
[0025] x i (k),i=1,2,...,n (5)
[0026] ② Membership function layer: This layer has a total of n×m neurons, m is 12, and the output of each node represents the membership value of the corresponding input quantity. The membership function is:
[0027]
[0028] In the formula, c ij (k) and δ ij (k) are the center and width of the membership function, and their initial values are random real numbers uniformly distributed in the range [0,2] generated by the rand random function;
[0029] ③Rule layer: This layer has a total of m neurons, using fuzzy multiplication operators as fuzzy logic rules. The output of the rule layer is:
[0030]
[0031] ④ Consequence layer: This layer has a total of m×q neurons, q is 3, each node performs the linear summation of TS-type fuzzy rules, and calculates the consequent parameters of the output corresponding to each rule The calculation process is as follows:
[0032]
[0033] In the formula, are the parameters of the fuzzy system, and their initial values are set to 0.3, x0(k), x1(k),…, x n (k) is the input variable, x0(k) is the input constant, which is used to enhance the network translation capability, and its value is 1;
[0034] ⑤ Output layer: This layer has q output nodes, each of which performs weighted summation on the input parameters. The calculation formula is as follows:
[0035]
[0036] ⑥Model parameter learning: Use the gradient descent algorithm to adjust the network parameters. First, define the error calculation method as follows:
[0037]
[0038] In the formula, y s (k) is the sth actual output corresponding to the kth input sample, is the sth calculated output corresponding to the kth input sample, e s (k) is the error between the two. Based on the error, the center, width and fuzzy system parameter update algorithm of the network are defined as follows:
[0039]
[0040]
[0041]
[0042] In the formula, η is the online learning rate, the value range of η is [0.01, 0, 05], c ij (k-1),δ ij (k-1) and are the center, width and fuzzy system parameters of the network membership function layer when the k-1th sample is input. After completing this parameter update, input the training sample data x i (k+1), repeat steps ① to ⑥ until all training samples are input and the number of training samples is 80% of the total number of samples K. Then the model is iteratively trained until the number of iterations reaches the maximum iteration value It max , It max is 500, at which point the modeling error is minimized, and the model at this point is established as the controlled object model;
[0043] (5) Multivariable controller design: This controller is used to control the controlled object model. It consists of a quasi-diagonal recurrent neural network self-tuning model, Jacobian information identification of the controlled object, PID parameter tuning mechanism and PID controller. The functions of each part are described as follows:
[0044] ① Quasi-diagonal recurrent neural network self-tuning model:
[0045] Input layer: This layer has ξ neurons, which are used to transmit input values. At time t, the input of the network can be expressed as:
[0046] I sr (t),r=1,2,...,ξ,s=1,2,...,q (14)
[0047] In the formula, ξ=3, q=3, and the input is [u s (t-1),y s (t),1], where u s (t-1) is the actual value of the manipulated variable at time t-1, y s (t) is the controlled variable, and the constant 1 is the bias term, which is used to enhance the translation ability of the network;
[0048] Regression layer: This layer has a total of ω neurons, ω is 7, and the input of the regression layer is G sg (t) includes input layer information, self-feedback information and interconnected neuron information, which can be expressed as:
[0049]
[0050] Where D sg (t-1) is the output of the regression layer neuron of model s at time t-1, is the connection weight between the input layer and the regression layer, is the recursive weight of the neurons in the regression layer, g = 1, 2, ..., ω,
[0051] Recursive weights of each neuron in the regression layer It can be expressed in the form of a quasi-diagonal matrix as shown below:
[0052]
[0053] According to formulas (15) and (16), each node of the regression layer is activated by the Sigmoid function f(), and the output of the regression layer is calculated as:
[0054]
[0055] Output layer: This layer has a single output node that performs a weighted sum of the input parameters. The output of the output layer can be expressed as:
[0056]
[0057] In the formula, is the connection weight between the regression layer and the output layer;
[0058] ② Jacobian information identification of the controlled object:
[0059] First, the error between the system output and the network output is used as the adjustment signal of the identifier, and the identification error of the controlled object system is defined as:
[0060] e ms (t) = y s (t)-y ms (t) (19)
[0061] In the formula, y s (t) is the actual value of the controlled quantity, y ms (t) is the estimated value of the controlled quantity calculated by the model;
[0062] Define the performance index J of the controlled object s (t) are as follows:
[0063]
[0064] The least mean square algorithm is used to update the parameters of the quasi-diagonal recurrent neural network model:
[0065]
[0066]
[0067]
[0068] In the formula, and The learning rates of the input layer, regression layer, and output layer are 4^10 respectively. -4 , the derivation process of the loss function to the weight is as follows:
[0069]
[0070] In the formula, f′(G sg (t)) is the derivative of the Sigmoid function of the regression layer;
[0071]
[0072]
[0073] According to the above derivation, the Jacobian information of the controlled object is calculated as:
[0074]
[0075] In the formula, u s (t) is the actual value of the manipulated variable at time t, is the input layer weight corresponding to the operation variable;
[0076] 3.PID controller:
[0077] The controller input error is defined as follows:
[0078] e s (t) = r s (t)-y s (t) (28)
[0079] In the formula, r s (t) is the set value of the controlled quantity;
[0080] The proportion, integral and differential of the error are defined as The calculation is as follows:
[0081]
[0082] Where T is the sampling time interval, the sampling interval is 1 second, is the time point within the sampling period;
[0083] The output of the PID controller is expressed as:
[0084]
[0085] In the formula, and They are the proportional coefficient, integral coefficient and differential coefficient of PID respectively;
[0086] ④PID parameter tuning mechanism:
[0087] To tune the parameters of the multivariable PID controller, first, define the PID performance index as:
[0088]
[0089] According to the performance index, calculate the proportional coefficient of PID Integration coefficient Differential coefficient The adjustment formula is as follows:
[0090]
[0091]
[0092]
[0093] In the formula, and They are the learning rates corresponding to the proportional coefficient, integral coefficient and differential coefficient of PID, and their values are 5^10 respectively. -5 , 3^10 -5 , 1^10 -5 , is the Jacobian information of the controlled object, which is obtained by identification using formula (27).
[0094] The creativity of the present invention is mainly reflected in:
[0095] (1) The present invention analyzes the characteristics of the municipal solid waste incineration process and extracts the key operating variables and controlled variables that can reflect the combustion state in combination with the air distribution and material distribution process;
[0096] (2) The present invention adopts a data-driven method to establish a controlled object model based on a multi-input and multi-output Takagi-Sugeno fuzzy neural network. The constructed controlled object model has multi-task learning capabilities, can use the complementary information between multiple tasks to accurately fit multiple controlled quantities at the same time, and update the network parameters online;
[0097] (3) The present invention constructs a multi-loop PID controller, designs a quasi-diagonal recursive neural network model with self-feedback channels and interconnected channels, and uses its ability to process transient information to adjust PID parameters, thereby achieving synchronous control of multiple variables; BRIEF DESCRIPTION OF THE DRAWINGS
[0098] Figure 1 This is the control flow chart of the municipal solid waste incineration process of the present invention
[0099] Figure 2 This is a multi-input multi-output controlled object model diagram based on the Takagi-Sugeno type fuzzy neural network of the present invention.
[0100] Figure 3 This is a quasi-diagonal recursive neural network PID multi-loop control strategy diagram for the municipal solid waste incineration process of the present invention.
[0101] Figure 4 This is the quasi-diagonal recursive neural network model structure diagram of the present invention
[0102] Figure 5 This is the main steam flow setting value control tracking result diagram of the present invention
[0103] Figure 6 This is the furnace temperature setting value control tracking result diagram of the present invention
[0104] Figure 7 This is the flue gas oxygen content setting value control tracking result diagram of the present invention
[0105] Figure 8 This is the primary air flow self-tuning process diagram of the present invention
[0106] Fig. 9 This is the secondary air flow self-tuning process diagram of the present invention
[0107] Fig.10 This is the drying grate speed self-tuning process diagram of the present invention DETAILED DESCRIPTION
[0108] The present invention obtains a multi-loop quasi-diagonal recursive neural network PID controller for the municipal solid waste incineration process. The controller analyzes the influencing factors of the model and extracts the key controlled quantities and operating quantities. Based on data drive, a multi-input and multi-output Takagi-Sugeno fuzzy neural network controlled object model is constructed. A quasi-diagonal recursive neural network model with self-feedback channels and interconnection channels is constructed and used for parameter self-tuning of multi-loop PID control, solving the multivariable control problem in the municipal solid waste incineration process.
[0109] The present invention adopts the following technical solutions and implementation steps:
[0110] The multi-loop quasi-diagonal recursive neural network PID controller method for municipal solid waste incineration process includes the following steps:
[0111] 1. A multi-loop quasi-diagonal recursive neural network PID control method for a municipal solid waste incineration process, characterized in that it comprises the following steps:
[0112] (1) Data collection of municipal solid waste incineration: The key operating variables of the municipal solid waste incineration process are collected through a one-way isolation system: drying grate air flow (left 1, right 1, left 2, right 2), combustion grate 1 section air flow (left 1, right 1, left 2, right 2), combustion grate 2 section air flow (left 1, right 1, left 2, right 2), burnout grate air flow (left, right), primary air total flow, secondary air flow, drying grate speed (left inner, right inner, left outer, right outer), combustion grate 1 section speed (left inner, right inner, left outer, right outer), combustion grate 2 section speed (left inner, right inner, left outer, right outer) and burnout grate speed (left inner, right inner), and the key controlled variables are collected: main steam flow, furnace temperature and flue gas oxygen content;
[0113] (2) Control characteristics analysis: using Pearson correlation coefficient ρ αβ For the manipulated variable α l and the controlled variable β l The correlation between them is evaluated and calculated as follows:
[0114]
[0115] In the formula, N is the total number of samples, and the total primary air flow, secondary air flow, and drying grate speed are selected as key operating variables based on the calculation results;
[0116] (3) Data preprocessing: The collected data is preprocessed by eliminating abnormal data and normalizing the data. The calculation steps are as follows:
[0117] ① Eliminate abnormal data: Use the 3σ criterion to eliminate abnormal data. First, set the dimension of sample data to q×K, where q is the number of controlled variables and K is the total number of samples. Indicates that is the sample mean, and its standard deviation σ s The calculation method is:
[0118]
[0119] when The corresponding residual error If the following conditions are met, Perform a culling operation, which is calculated as:
[0120]
[0121] ②Data normalization: Sample data Normalize and calculate the normalized data samples The calculation method is as follows:
[0122]
[0123] in, is the sample vector of variable s. After normalization, the range of each sample data is between [0,1];
[0124] (4) Design of controlled object model: A multi-input and multi-output Takagi-Sugeno fuzzy neural network controlled object model was constructed through data driving, which includes five layers: input layer, membership function layer, rule layer, consequent layer and output layer. Its mathematical description is as follows:
[0125] ① Input layer: This layer has n neurons, n is 3, and its function is to transmit the input value. When the kth sample enters, the output of the input layer is:
[0126] x i (k),i=1,2,...,n (5)
[0127] ② Membership function layer: This layer has a total of n×m neurons, m is 12, and the output of each node represents the membership value of the corresponding input quantity. The membership function is:
[0128]
[0129] In the formula, c ij (k) and δ ij (k) are the center and width of the membership function, and their initial values are random real numbers uniformly distributed in the range [0,2] generated by the rand random function;
[0130] ③Rule layer: This layer has a total of m neurons, using fuzzy multiplication operators as fuzzy logic rules. The output of the rule layer is:
[0131]
[0132] ④ Consequence layer: This layer has a total of m×q neurons, q is 3, each node performs the linear summation of TS-type fuzzy rules, and calculates the consequent parameters of the output corresponding to each rule The calculation process is as follows:
[0133]
[0134] In the formula, are the parameters of the fuzzy system, and their initial values are set to 0.3, x0(k), x1(k),…, x n(k) is the input variable, x0(k) is the input constant, which is used to enhance the network translation capability, and its value is 1;
[0135] ⑤ Output layer: This layer has q output nodes, each of which performs weighted summation on the input parameters. The calculation formula is as follows:
[0136]
[0137] ⑥Model parameter learning: Use the gradient descent algorithm to adjust the network parameters. First, define the error calculation method as follows:
[0138]
[0139] In the formula, y s (k) is the sth actual output corresponding to the kth input sample, is the sth calculated output corresponding to the kth input sample, e s (k) is the error between the two. Based on the error, the center, width and fuzzy system parameter update algorithm of the network are defined as follows:
[0140]
[0141]
[0142]
[0143] In the formula, η is the online learning rate, the value range of η is [0.01, 0, 05], c ij (k-1),δ ij (k-1) and are the center, width and fuzzy system parameters of the network membership function layer when the k-1th sample is input. After completing this parameter update, input the training sample data x i (k+1), repeat steps ① to ⑥ until all training samples are input and the number of training samples is 80% of the total number of samples K. Then the model is iteratively trained until the number of iterations reaches the maximum iteration value It max , It max is 500, at which point the modeling error is minimized, and the model at this point is established as the controlled object model;
[0144] (5) Multivariable controller design: This controller is used to control the controlled object model. It consists of a quasi-diagonal recurrent neural network self-tuning model, Jacobian information identification of the controlled object, PID parameter tuning mechanism and PID controller. The functions of each part are described as follows:
[0145] ① Quasi-diagonal recurrent neural network self-tuning model:
[0146] Input layer: This layer has ξ neurons, which are used to transmit input values. At time t, the input of the network can be expressed as:
[0147] I sr (t),r=1,2,...,ξ,s=1,2,...,q (14)
[0148] In the formula, ξ=3, q=3, and the input is [u s (t-1),y s (t),1], where u s (t-1) is the actual value of the manipulated variable at time t-1, y s (t) is the controlled variable, and the constant 1 is the bias term, which is used to enhance the translation ability of the network;
[0149] Regression layer: This layer has a total of ω neurons, ω is 7, and the input of the regression layer is G sg (t) includes input layer information, self-feedback information and interconnected neuron information, which can be expressed as:
[0150]
[0151] Where D sg (t-1) is the output of the regression layer neuron of model s at time t-1, is the connection weight between the input layer and the regression layer, is the recursive weight of the neurons in the regression layer, g = 1, 2, ..., ω,
[0152] Recursive weights of each neuron in the regression layer It can be expressed in the form of a quasi-diagonal matrix as shown below:
[0153]
[0154] According to formulas (15) and (16), each node of the regression layer is activated by the Sigmoid function f(), and the output of the regression layer is calculated as:
[0155]
[0156] Output layer: This layer has a single output node that performs a weighted sum of the input parameters. The output of the output layer can be expressed as:
[0157]
[0158] In the formula, is the connection weight between the regression layer and the output layer;
[0159] ② Jacobian information identification of the controlled object:
[0160] First, the error between the system output and the network output is used as the adjustment signal of the identifier, and the identification error of the controlled object system is defined as:
[0161] e ms (t) = y s (t)-y ms (t) (19)
[0162] In the formula, y s (t) is the actual value of the controlled quantity, y ms (t) is the estimated value of the controlled quantity calculated by the model;
[0163] Define the performance index J of the controlled object s (t) are as follows:
[0164]
[0165] The least mean square algorithm is used to update the parameters of the quasi-diagonal recurrent neural network model:
[0166]
[0167]
[0168]
[0169] In the formula, and The learning rates of the input layer, regression layer, and output layer are 4^10 respectively. -4 , the derivation process of the loss function to the weight is as follows:
[0170]
[0171] In the formula, f′(G sg (t)) is the derivative of the Sigmoid function of the regression layer;
[0172]
[0173]
[0174] According to the above derivation, the Jacobian information of the controlled object is calculated as:
[0175]
[0176] In the formula, u s (t) is the actual value of the manipulated variable at time t, is the input layer weight corresponding to the operation variable;
[0177] 3.PID controller:
[0178] The controller input error is defined as follows:
[0179] e s (t) = r s (t)-y s (t) (28)
[0180] In the formula, r s (t) is the set value of the controlled quantity;
[0181] The proportion, integral and differential of the error are defined as The calculation is as follows:
[0182]
[0183] Where T is the sampling time interval, the sampling interval is 1 second, is the time point within the sampling period;
[0184] The output of the PID controller is expressed as:
[0185]
[0186] In the formula, and They are the proportional coefficient, integral coefficient and differential coefficient of PID respectively;
[0187] ④PID parameter tuning mechanism:
[0188] To tune the parameters of the multivariable PID controller, first, define the PID performance index as:
[0189]
[0190] According to the performance index, calculate the proportional coefficient of PID Integration coefficient Differential coefficient The adjustment formula is as follows:
[0191]
[0192]
[0193]
[0194] In the formula, and They are the learning rates corresponding to the proportional coefficient, integral coefficient and differential coefficient of PID, and their values are 5^10 respectively. -5 , 3^10-5 , 1^10 -5 , is the Jacobian information of the controlled object, which is obtained by identification using formula (27).
[0195] Figure 1 Control process for municipal solid waste incineration; Figure 2 It is a multi-input and multi-output controlled object model based on Takagi-Sugeno type fuzzy neural network; Figure 3 Quasi-diagonal recursive neural network PID multi-loop control strategy for municipal solid waste incineration process; Figure 4 It is a quasi-diagonal recurrent neural network model structure; Figure 5 The tracking result of the main steam flow set value control, X-axis: time, unit is s, Y-axis: main steam flow, unit is t / h, the black dotted line is the set value, and the gray solid line is the tracking result; Figure 6 It is the tracking result of furnace temperature set value control, X axis: time, unit is s, Y axis: furnace temperature, unit is ℃, dotted line is set value, black line is tracking result; Figure 7 Figure 8 is the control tracking result of the flue gas oxygen content set value, X axis: time, unit is s, Y axis: flue gas oxygen content, unit is %, the dotted line is the set value, and the black line is the tracking result; Figure 9 is the primary air flow self-tuning process, X axis: time, unit is s, Y axis: primary air flow, unit is km 3 N / h; Fig. 9 For the secondary air flow self-tuning process, X-axis: time, unit is s, Y-axis: secondary air flow, unit is km 3 N / h; Fig.10 This is the drying grate speed self-tuning process, X-axis: time, unit is s, Y-axis: drying grate speed, unit is %; The results show the effectiveness of the controller for multivariable control of municipal solid waste incineration process.
Claims
1. A multi-loop quasi-diagonal recursive neural network PID control method for municipal solid waste incineration process, characterized in that: The following steps are involved: (1) Data collection of municipal solid waste incineration: The key operating variables of the municipal solid waste incineration process are collected through a one-way isolation system: drying grate air flow, combustion grate stage 1 air flow, combustion grate stage 2 air flow, burnout grate air flow, primary air total flow, secondary air flow, drying grate speed, combustion grate stage 1 speed, combustion grate stage 2 speed and burnout grate speed; the key controlled variables are collected: main steam flow, furnace temperature and flue gas oxygen content; The drying grate air flow rate includes the left 1 drying grate air flow rate, the right 1 drying grate air flow rate, the left 2 drying grate air flow rate and the right 2 drying grate air flow rate; The air flow rate of combustion grate section 1 includes the air flow rate of combustion grate section 1 on the left 1, the air flow rate of combustion grate section 1 on the right 1, the air flow rate of combustion grate section 1 on the left 2, and the air flow rate of combustion grate section 1 on the right 2; the air flow rate of combustion grate section 2 includes the air flow rate of combustion grate section 2 on the left 1, the air flow rate of combustion grate section 2 on the right 1, the air flow rate of combustion grate section 2 on the left 2, and the air flow rate of combustion grate section 2 on the right 2; The burnout grate air flow rate includes the left burnout grate air flow rate and the right burnout grate air flow rate; the drying grate speed includes the left inner drying grate speed, the right inner drying grate speed, the left outer drying grate speed and the right outer drying grate speed; The combustion grate stage 1 speed includes the left inner combustion grate stage 1 speed, the right inner combustion grate stage 1 speed, the left outer combustion grate stage 1 speed and the right outer combustion grate stage 1 speed; the combustion grate stage 2 speed includes the left inner combustion grate stage 2 speed, the right inner combustion grate stage 2 speed, the left outer combustion grate stage 2 speed and the right outer combustion grate stage 2 speed; the burnout grate speed includes the left inner burnout grate speed and the right inner burnout grate speed; (2) Control characteristics analysis: using Pearson correlation coefficient ρ αβ For the manipulated variable α l and the controlled variable β l The correlation between them is evaluated and calculated as follows: In the formula, N is the total number of samples, and the total primary air flow, secondary air flow, and drying grate speed are selected as key operating variables based on the calculation results; (3) Data preprocessing: The collected data is preprocessed by eliminating abnormal data and normalizing the data. The calculation steps are as follows: ① Eliminate abnormal data: Use the 3σ criterion to eliminate abnormal data. First, set the dimension of sample data to q×K, where q is the number of controlled variables and K is the total number of samples. Indicates, where s = 1, 2, ..., q, k = 1, 2, ..., K; is the sample mean, and its standard deviation σ s The calculation method is: when The corresponding residual error If the following conditions are met, Perform a culling operation, which is calculated as: ②Data normalization: Sample data Normalize and calculate the normalized data samples The calculation method is as follows: in, is the sample vector of variable s. After normalization, the range of each sample data is between [0,1]; (4) Design of controlled object model: A multi-input and multi-output Takagi-Sugeno fuzzy neural network controlled object model was constructed through data driving, which includes five layers: input layer, membership function layer, rule layer, consequent layer and output layer. Its mathematical description is as follows: ① Input layer: This layer has n neurons, n is 3, and its function is to transmit the input value. When the kth sample enters, the output of the input layer is: x i (k),i=1,2,…,n (5) ② Membership function layer: This layer has a total of n×m neurons, m is 12, and the output of each node represents the membership value of the corresponding input quantity. The membership function is: In the formula, c ij (k) and δ ij (k) are the center and width of the membership function, and their initial values are random real numbers uniformly distributed in the range [0,2] generated by the rand random function; ③Rule layer: This layer has a total of m neurons, using fuzzy multiplication operators as fuzzy logic rules. The output of the rule layer is: ④ Consequence layer: This layer has a total of m×q neurons, q is 3, each node performs the linear summation of TS-type fuzzy rules, and calculates the consequent parameters of the output corresponding to each rule The calculation process is as follows: In the formula, are the parameters of the fuzzy system, and their initial values are set to 0.3, x0(k), x1(k),…, x n (k) is the input variable, x0(k) is the input constant, which is used to enhance the network translation capability, and its value is 1; ⑤ Output layer: This layer has q output nodes, each of which performs weighted summation on the input parameters. The calculation formula is as follows: ⑥Model parameter learning: Use the gradient descent algorithm to adjust the network parameters. First, define the error calculation method as follows: In the formula, y s (k) is the sth actual output corresponding to the kth input sample, is the sth calculated output corresponding to the kth input sample, e s (k) is the error between the two. Based on the error, the center, width and fuzzy system parameter update algorithm of the network are defined as follows: In the formula, η is the online learning rate, the value range of η is [0.01, 0, 05], c ij (k-1),δ ij (k-1) and are the center, width and fuzzy system parameters of the network membership function layer when the k-1th sample is input. After completing this parameter update, input the training sample data x i (k+1), repeat steps ① to ⑥ until all training samples are input and the number of training samples is 80% of the total number of samples K. Then the model is iteratively trained until the number of iterations reaches the maximum iteration value It max , It max is 500, at which point the modeling error is minimized, and the model at this point is established as the controlled object model; (5) Multivariable controller design: This controller is used to control the controlled object model. It consists of a quasi-diagonal recurrent neural network self-tuning model, Jacobian information identification of the controlled object, PID parameter tuning mechanism and PID controller. The functions of each part are described as follows: ① Quasi-diagonal recurrent neural network self-tuning model: Input layer: This layer has ξ neurons, which are used to transmit input values. At time t, the input of the network can be expressed as: I sr (t),r=1,2,…,ξ,s=1,2,...,q (14) In the formula, ξ=3, q=3, and the input is [u s (t-1),y s (t),1], where u s (t-1) is the actual value of the manipulated variable at time t-1, y s (t) is the controlled variable, and the constant 1 is the bias term, which is used to enhance the translation ability of the network; Regression layer: This layer has a total of ω neurons, ω is 7, and the input of the regression layer is G sg (t) includes input layer information, self-feedback information and interconnected neuron information, which can be expressed as: Where D sg (t-1) is the output of the regression layer neuron of model s at time t-1, is the connection weight between the input layer and the regression layer, is the recursive weight of the neurons in the regression layer, g = 1, 2, ..., ω, Recursive weights of each neuron in the regression layer It can be expressed in the form of a quasi-diagonal matrix as shown below: According to formulas (15) and (16), each node of the regression layer is activated by the Sigmoid function f( ), and the output of the regression layer is calculated as: Output layer: This layer has a single output node that performs a weighted sum of the input parameters. The output of the output layer can be expressed as: In the formula, is the connection weight between the regression layer and the output layer; ② Jacobian information identification of the controlled object: First, the error between the system output and the network output is used as the adjustment signal of the identifier, and the identification error of the controlled object system is defined as: e ms (t)=y s (t)-y ms (t) (19) In the formula, y s (t) is the actual value of the controlled quantity, y ms (t) is the estimated value of the controlled quantity calculated by the model; Define the performance index J of the controlled object s (t) are as follows: The least mean square algorithm is used to update the parameters of the quasi-diagonal recurrent neural network model: In the formula, and The learning rates of the input layer, regression layer, and output layer are 4^10 respectively. -4 , the derivation process of the loss function to the weight is as follows: In the formula, f′(G sg (t)) is the derivative of the Sigmoid function of the regression layer; According to the above derivation, the Jacobian information of the controlled object is calculated as: In the formula, u s (t) is the actual value of the manipulated variable at time t, is the input layer weight corresponding to the operation variable; 3.PID controller: The controller input error is defined as follows: e s (t)=r s (t)-y s (t) (28) In the formula, r s (t) is the set value of the controlled quantity; The proportion, integral and differential of the error are defined as The calculation is as follows: Where T is the sampling time interval, the sampling interval is 1 second, is the time point within the sampling period; The output of the PID controller is expressed as: In the formula, and They are the proportional coefficient, integral coefficient and differential coefficient of PID respectively; ④PID parameter tuning mechanism: To tune the parameters of the multivariable PID controller, first, define the PID performance index as: According to the performance index, calculate the proportional coefficient of PID Integration coefficient Differential coefficient The adjustment formula is as follows: In the formula, and They are the learning rates corresponding to the proportional coefficient, integral coefficient and differential coefficient of PID, and their values are 5^10 respectively. -5 , 3^10 -5 , 1^10 -5 , is the Jacobian information of the controlled object, which is obtained by identification using formula (27).
Citation Information
Patent Citations
Crowd-sourcing system and method for controlling harmful emissions of pesticide waste liquid incinerator to reach standard
CN103472729A
Combustion furnace capable of inhibiting generation of unburned objects during garbage combustion process
CN105698185A
Improved neural network boiler combustion system modeling method based on object combustion mechanism
CN107977539A
Effluent ammonia nitrogen concentration prediction method based on a self-organizing recurrent fuzzy neural network with multivariate time sequence analysis
CN111354423A
Intelligent identification method of sludge bulking based on type-2 fuzzy neural network
US20200024168A1
