Prediction function control method for oxygen content of coking heating furnace
Technical Field
The invention belongs to the technical field of automation, and relates to a prediction function control method for the oxygen content of a coking heating furnace.
Background
Coking plays an important role in improving the economy of the petrochemical industry. In the control system of the coke oven, the control of the oxygen content is a very important issue, and it directly affects the pressure in the chamber, the temperature of the radiation, and the like. Because the coke oven, the distillation tower and the coke tower are an integrated process which interacts with each other. The volume of gaseous oil in the distillation column is closely related to the circulating temperature of the oil flowing into the coke oven, however the temperature of the oil has a direct effect on the oxygen content. The temperature of the oil remaining in the coke oven and the coking rate in the coke drum also affect the amount of fuel required to be fed into the oven, these factors also being closely related to the oxygen content. Process disturbances, time delays, and nonlinearities affect the oxygen content in the coke oven, causing the switches in the coke drum to cause many periodic oscillations in the process. It is also difficult for a typical PID controller to control the oxygen content to a set value due to non-linearity, complex dynamics and variable disturbances in the process. Although model predictive control (DMC) has been widely used in industrial processes, the complexity of coke ovens limits the applicability and efficiency of DMC control. Although multi-model strategies and non-linear control can be used, the multi-model strategy requires a lot of operations and extensive experiments, which results in that the control performance is greatly dependent on the form of the operations, and for non-linear control, it is a great problem facing the industry to have a sufficiently effective non-linear model and an optimal method of the corresponding non-linear model.
When very complex non-linearities occur in an industrial process, data from the industrial process becomes very important. The ideal controller can be designed by building a dynamic model of the process from a large amount of available data in the industry, and then obtaining the structure of the model. Research in the related art has emerged, but a number of challenges have arisen to be resolved. For example, there are many model parameters to identify and the model disturbance through negative feedback tends to be an infinite predictor.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a control method for controlling the oxygen content in a coking heating furnace by combining a data structure model with a prediction function control design. The control method has the advantages that the controller is designed by utilizing the black box principle in the industry, and the linear control method can be effectively utilized. Compared with the traditional PID method, the control method improves the anti-interference performance in the process in practical application.
The method comprises the steps of firstly modeling the industrial process through a large amount of data acquired in the industrial process, regarding a model as a linear part and a nonlinear part, then respectively modeling the nonlinear part by utilizing a Back Propagation Neural Network (BPNN), and modeling the linear part by utilizing a traditional linear two-point method. And performing rolling optimization and feedback correction on the established model by using a Predictive Function Control (PFC) method, thereby determining a control variable output at the next moment.
The technical scheme of the invention is to establish a model through data acquisition, and continuously perform rolling optimization to determine the controlled quantity by using a prediction mechanism controlled by a prediction function. The method can obviously improve the dynamic performance and the stability of the system.
The method comprises the following steps:
1. modeling of controlled objects
1-1. Since the model of the controlled object is composed of two parts, linear and nonlinear, it can be described as the following form:
y(k)=y L (k)+y NL (k) (1)
wherein y is L (k) Is the output of the linear model at step response, y NL (k) Is through y L (k) And y (k) electrically i A non-linear model of deviation determination between (i =1,2, \8230;, N), y (k) calculation of luminance i And (i =1,2, \8230;, N) is the output of the controlled object y (k) in the actual process at the time point i, and N is the number of output sampling points.
1-2. The linear part of the model can be obtained by a step response, y (t) is the actual output of the model, y (∞) is the steady-state output of the model, U 0 Is the amplification of the input signal. The output y (t) can be usedExpressed in terms of form, the gain of the model may be expressed as
According to the characteristics of the model process, the linear part of the whole model can be described as a first-order model plus a hysteresis link or a second-order model hysteresis link. The linear model was chosen to be of the form:
thus y * (t) can be described in the following manner.
Wherein take y * (t 1 )=0.39,y * (t 2 )=0.63,t 2 >t 1 >, τ, the delay time τ and the response time T can be obtained as follows:
1-3. The non-linear part of the model can be obtained by the following steps:
is provided withWherein y is Li (k) (i =1,2, \8230;, N) is y L (k) The value at the corresponding time point. Y of non-linear part in model NL (k) By BPNNThe model is as follows:
wherein w 2 (i,j)(j=1,2),w 3 (i) Is the connection weight of the neural network weight link layer, I is the number of output nodes, and g (x) activation function can be selected as g (x) = 1/(1 + e) -x ) The delay link in the model is equivalent to d = tau/T s . By sampling time T s The model of the controlled object is discretized and then expressed in the following form:
y(k)=α 1 y(k-1)+β 0 u(k-d-1)+y NL (k) (6)
wherein alpha is 1 ,β 0 Are the corresponding coefficients in the equation.
2. Predictive control function (PFC) controller design:
2-1. Output prediction
Setting:
to predict the outcome of future processes, g (θ (k)) andrespectively at the central point theta 0 Andthe linearization process yields the following equation:
wherein φ (k) = [ u (k-d-1), y m (k-1)] T ε andis a non-linear function.
Further from the above expression:
can find out from the above formulaIs closely related to the future prediction part,is a higher order function and is therefore negligible. The following form can thus be obtained:
whereinIs a constant term, the following equation can be obtained:
y(k)=a 1 y(k-1)+b 0 u(k-d-1)+C (13)
wherein
2-2. Difference operator Δ =1-z from both ends of equation (13) above -1 Transforming the output of the model after differentiation into y m (k) The results are in the form:
y m (k)=A 1 y m (k-1)+A 2 y m (k-2)+B 1,0 Δu(k-d-1) (15)
wherein A is 1 =1+a 1 ,A 2 =-a 1 ,B 1,0 =b 0 。
2-3. The PFC controller is designed using the model of equation (15) above and can be divided into three parts. The first part being y past (k + d + p) which is determined by the past inputs and outputs, the second part being G p U p Is determined by the present and future input quantities, and the third part is the prediction error, which is determined by the feedback errorWhereinAt real time k, the spatial choice of the prediction function is N y :
Wherein U is p =(Δu(k),Δu(k+1),…,Δu(k+p-1)) T
In the strategy of predictive control, the controlled variable is closely related to the characteristics of the system process and the initial set point.
Wherein λ j Is a weight coefficient, f j (i) Is the value of the basis function at the sampling instant i, M being the basis function
Where μ is the smoothing coefficient, y s Is a set value
2-4, optimizing the performance indexes as follows:
thereby obtaining a control increment to calculate the control quantity at the time k:
and 2-5, applying the obtained control quantity to the controlled object, and repeating the steps 2-2 to 2-4 until the next moment to continuously solve the control quantity u (k + 1) of the controlled object in a circulating mode.
The invention has the following beneficial effects:
the invention decomposes a complex model in industry into a simple step response model and a nonlinear part, and combines the prediction control and the prediction function control of a nonlinear system by utilizing a neural network, thereby reducing the complexity of the system structure and lightening the burden of operation. The local linear representation of the nonlinear excitation function is utilized to convert the multi-step nonlinear prediction into a series of simple and intuitive multi-step linear prediction forms. Compared with the traditional PID control, the dynamic performance and stability of the system can be more effectively improved by utilizing the prediction function control, and the method is applied to the setting of an industrial controller.
Detailed Description
Taking the oxygen content control process in the coking heating furnace as an example:
1. model for establishing coking heating furnace
1-1. Since the oxygen content in a coke oven can be described as two-part, linear and non-linear, the following expression can be obtained:
y(k)=y L (k)+y NL (k) (1)
wherein y is L (k) Is the output of the model in step response, y NL (k) Is through y L (k) And y (k) electrically i A non-linear module for determining a deviation between (i =1,2, \8230;, N), y (k) & lty & gt i And (i =1,2, \8230;, N) is the output of the controlled object output y (k) in the actual process at the time point i, and N is the number of output sampling points.
1-2. The linear part of the model can be obtained by a step response, y (t) is the actual output of the model, y (∞) is the steady-state output of the model, U 0 Is the amplification of the input signal. The output y (t) can be usedExpressed in terms of form, the gain of the model may be expressed as
According to the characteristics of the model process, the linear part of the whole model can be described as a first-order model plus a hysteresis link or a second-order model hysteresis link. We chose the linear model to be of the form:
thus y * (t) can be described in the following manner.
Wherein get y * (t 1 )=0.39,y * (t 2 )=0.63,t 2 >t 1 &τ, the delay time τ and the response time T can be obtained as follows:
1-3. The non-linear part of the model can be obtained by the following steps:
is provided withWherein y is Li (k) (i =1,2, \8230;, N) is y L (k) The value at the corresponding time point. Y of non-linear part in model NL (k) Can be implemented by Back Propagation Neural Networks (BPNN)The value of (a) is minimum, and the model is as follows:
wherein w 2 (i,j)(j=1,2),w 3 (i) Is the connection weight of the neural network weight link layer, I is the number of output nodes, and g (x) activation function can be selected as g (x) = 1/(1 + e) -x ) The delay link in the model is equivalent to d = tau/T s . By sampling time T s Discretizing the model of the controlled object and expressing the model into the following form:
y(k)=α 1 y(k-1)+β 0 u(k-d-1)+y NL (k) (6)
wherein alpha is 1 ,β 0 Are the corresponding coefficients in the equation.
2. The specific method for designing the Prediction Function Control (PFC) controller comprises the following steps:
2-1, output prediction:
setting:
to predict the outcome of future processes, g (θ (k)) andrespectively at the central point theta 0 Andthe linearization process yields the following equation:
wherein φ (k) = [ u (k-d-1), y m (k-1)] T ε andis a non-linear function.
Further derived from the above expression:
can find from the above formulaIs closely related to the future prediction part,is a higher order function and is therefore negligible. The following forms can thus be obtained:
whereinIs a constant term, the following equation can be obtained:
y(k)=a 1 y(k-1)+b 0 u(k-d-1)+C (13)
wherein
2-2. Difference operator Δ =1-z from both ends of equation (13) above -1 Transforming the output of the model after differentiation into y m (k) The results are in the form of:
y m (k)=A 1 y m (k-1)+A 2 y m (k-2)+B 1,0 Δu(k-d-1) (15)
wherein A is 1 =1+a 1 ,A 2 =-a 1 ,B 1,0 =b 0 。
2-3, designing PFC controller by using the model of the above formula (15), we can divide it into three parts. The first part being y past (k + d + p) which is determined by the past inputs and outputs of the coke oven, and the second component is G p U p Is determined by the present and future inputs to the coke oven, and the third component is the prediction error, which is determined by the feedback error WhereinAt real time k, the spatial choice of the prediction function is N y :
Wherein U is p =(Δu(k),Δu(k+1),…,Δu(k+p-1)) T
In predictive control of the oxygen content of a coke oven, the amount of control is closely related to the characteristics of the system process and the initial set point.
Wherein λ j Is a weight coefficient, f j (i) Is the basis function at the sampling instanti value, M being a basis function
Where μ is the smoothing coefficient, y s Is a set value
2-4. Optimizing the performance index as follows:
so as to obtain the control increment of the oxygen content, and calculate the oxygen control quantity at the k moment:
and 2-5, applying the obtained control quantity to the controlled object, and repeating the steps 2-2 to 2-4 until the next moment to continuously solve the control quantity u (k + 1) of the controlled object in a circulating mode.