JP2643506B2 - Predictive controllers for industrial processes - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a predictive control device used for an industrial process such as a plant.

[Prior Art] In general, predictive control is known as an effective control method for a plant having a large delay time or a dead time element.

Conventional prediction control is performed using a controller including a predicted value generation unit 1 and a control signal generation unit 2 as shown in FIG. 9, for example. This controller has a plant model expressing the dynamic characteristics of the plant to be controlled in a predicted value generation unit 1, predicts a future value of a plant output calculated from this model and a past plant input U, and calculates a predicted value ΔY
The control output CV is calculated by the control signal generator 2 based on

In such a prediction control method, as shown in FIG.
The predicted value ΔY starts from the current measured value of the plant output. Therefore, the current measured value Y of the plant output is input from the plant output observation unit 3 to the control signal generation unit 2 in FIG. And if there is a discrepancy between the predicted plant output value and the actual measured value,
The measured values are used entirely and the predicted values up to that point are corrected in a biased manner.

However, in an actual plant, the measurement is usually accompanied by noise, and the predicted value is fluctuated by this noise as shown in Fig. 11, resulting in unstable controller output and unnecessary wasteful movement. Will occur.

 Therefore, the following countermeasures have been considered.

Make the controller output CV blunt.

In FIG. 9, a noise filter 4 is inserted between the plant output observation unit 3 and the control signal generation unit 2.

[Problem to be Solved by the Invention] However, as a countermeasure, if a smoothing filter is inserted to blunt the controller output, the original control function is lost, and the intended control specification cannot be satisfied. Also, lowering the feedback gain G of the controller results in prolonged settling to the set value.

On the other hand, according to the countermeasures, the noise is a first-order lag filter,
It can be removed by a moving average filter, an exponential smoothing filter, or the like.
When the sample cycle of the measurement system of the plant output is extremely short as compared with the control cycle of the controller, the filter is effective, and unnecessary movement of the controller output as shown in FIG. 11 is suppressed.

Therefore, a case is considered in which the sampling of the plant output measurement system and the control cycle of the controller are synchronized. In this case, if the true value of the plant output is a constant value, noise can be substantially removed by applying a moving average filter.

However, in an operating plant, any of the output values of the above filters causes a phase delay with respect to the true plant output. This is because the filter uses past plant output values (measured values) in the smoothing process. As described above, since the measured value has a phase delay, deterioration of control performance by the phase delay cannot be avoided as a result.

The problems of the conventional predictive control described above are summarized as follows.

Conventional predictive control does not consider noise. Therefore, when there is an influence of noise, it is necessary to give up precise control. Further, if a general noise filter is added, the original advantage of the prediction control is lost. The cause is
This is because the use of the above filter causes a phase delay on the filter output side.

An object of the present invention is to provide a predictive control device that reduces the influence of measurement noise and does not cause a phase delay in an industrial process such as a plant.

Means for Solving the Problems The present invention calculates a predicted output value of a controlled object from a model expressing dynamic characteristics of a controlled object process and an input to the process, and outputs a control signal based on the predicted value. Do
A predictive control device for an industrial process, comprising, as a state observer of a controlled object, a Kalman filter represented by a model including a disturbance element, and inputting measured values of input and output of the controlled object and disturbance to the Kalman filter. Thus, the calculation of the predicted value and the noise removal are performed simultaneously.

[Operation] In the Kalman filter, a tentative estimated output value obtained from the control target model is compared with a measured output value, and a difference between the tentative estimated output value and the measured output value is set as a measurement error. Here, since the measurement error includes an error of both the measurement input value and the measurement output value, both of them are targeted.

In operation, the measured value of the control target input is input to the Kalman filter, the control target model is advanced by one sample, and a provisional estimated state quantity is obtained without phase delay.

Next, the control target output value predicted this time (temporary estimated output value) is compared with the actually measured output value, and the temporary estimated state amount is optimally corrected based on the difference, and the optimal estimated state amount ( Control target output). As a result, measurement noise is reduced.

Here, the optimal correction means using a Kalman filter gain which is an optimal correction gain. The Kalman filter gain is uniquely obtained as an optimal gain for the controlled object model given the error variance included in the measured value of the input / output of the controlled object.

According to the present invention, the Kalman filter is represented by a model including a disturbance element. This is from the viewpoint that it is indispensable to handle a process-specific disturbance in order to control a process to be controlled according to the present invention with high accuracy. On the other hand, in the conventional Kalman filter, all inputs are treated as noise, except for signals whose contents are known, such as manipulated variables such as plant inputs, which are also processed as white noise. If used as it is, the disturbance of the process cannot be handled, and the control is not performed with high accuracy. Therefore, in the present invention, as described above, the Kalman filter includes a disturbance element, that is, is configured with a model that can process a disturbance as an input, and by inputting the measured values of the input and output of the control target and the disturbance to the model, Predictive control of the process is executed with high accuracy.

Embodiment FIG. 1 shows a configuration of a prediction control device according to the present invention.
This device is divided into a PV (predicted plant output) calculation unit 11 and a control calculation unit 21.

The PV calculation unit 11 includes a state observer 12 using a Kalman filter, a gain calculation unit 13 for providing a Kalman filter gain K, and a screen display prediction value calculation unit 14 as described later, and its output PV (prediction value). Is sent to CRT15 and displayed. The control operation unit 21 generates a state deviation predicted value deviation ΔY for state feedback from the output (optimum state estimated value) X of the state observer 12, and performs a predetermined operation from the output to generate a control signal. The control signal generator 23 outputs a CV. The control signal generator 23 calculates the deviation based on the deviation ΔY and the preset value SP, and calculates the result and the feedback gain G from the state feedback gain calculator 24.
Is used to calculate the control output CV.

It should be noted that the predictive control device of the present invention has one input
Although shown as an output system, it can be extended to a multi-input system as described later.

According to the present invention, the following Kalman filter is used for the state observer 12 of FIG.

As shown in FIG. 2, the Kalman filter expresses the original operation of the plant by a plant model, compares a provisional estimated plant output value ⁰ obtained therefrom with a measurement plant output value, and finds the difference ( ⁻⁰ ). Think of it as a measurement error. Here, the measurement error includes not only the measurement error w of the measurement plant output value, but also the measurement error v of the measurement plant input value used for the plant model calculation for obtaining the estimated plant output value ⁰ . Therefore,
A noise filter is required to consider both of them.

This Kalman filter is represented by the following equation. ⁰ (k) = A (k−1) + BΔ (k) (1) ⁰ (k) = C ⁰ (k) (2) (k) = ⁰ (k) + K ((k) ⁻⁰ (k))
(3) where Δ (k) = (k) − (k−1) (4) (k) = U (k) + v (k) (5) (k) = Y (k) + w (k) ( 6) In the above equation, k is the number of samples, A is a system parameter, B is an impulse response model, C is an observation vector, K
Is the Kalman filter gain, and U (k) and Y (k) are the true values of the plant input and plant output, respectively.

In FIG. 2, the plant input and the plant output are sampled at predetermined sampling intervals by the switching means 25 and 26, respectively. Further, z ^-1 is a delay element based on z conversion.

The Kalman filter utilizes the fact that the above measurement error generally has an average value of 0 and has whiteness (white noise), and uses the estimation error (the difference between the estimated value and the true value: e
= −Y) to calculate an optimal state estimate in a statistical sense that minimizes the sum of squares.

The respective errors v,
The magnitude of w is represented as each measurement error variance, and the Kalman filter gain K is calculated using the variance values of both as parameters. Here, the influence of the measurement error w of the plant output value is one time only at the time of the measurement, but the measurement error v of the plant input value continues to exert an influence for a certain period after that along with the dynamic characteristics of the plant. That means
The Kalman filter also corrects this future value.

 The operation of the above Kalman filter is as follows.

First, the current value (k) of the plant input is input to the Kalman filter, the plant model is advanced by one sample (thereby eliminating phase lag), and a provisional estimated state quantity ⁰ (k) is obtained according to the equation (1).

Next, the plant output value (temporary estimated plant output value) ⁰ (k) predicted this time according to the equation (2) is compared with the actually measured plant output value (k), and the difference is expressed as (3) The provisional estimated state quantity ⁰ (k) is optimally corrected by the equation to obtain the optimal estimated state quantity (k) (the result is:
Measurement noise is reduced).

Here, the optimal correction means using the Kalman filter gain K which is an optimal correction gain. The Kalman filter gain K is given by the following equation.

K = MC ^T (CMC ^T + W) ⁻¹ (7) M = A (1−KC) MC ^T + BVB ^T (8) where V is the measurement error variance of the plant input, W is the measurement error variance of the plant output, M Is the error variance matrix, and ^T represents the transpose of the matrix.

Therefore, given the error variances V and W included in the measured values of the plant input / output, the Kalman filter gain K is uniquely obtained as the optimal gain for the plant model.

FIG. 3 is a block diagram showing a calculation process for generating an output (estimated plant output value) from the input (measured plant input / output value) to the Kalman filter as described above. That is, the plant input measurement value becomes an input to the impulse response model B of the plant, and generates plant dynamics. Then, by using the plant output measurement value and the output calculation value from the impulse response model B and multiplying by the Kalman filter gain K,
An optimal estimate of the plant output is obtained.

Fig. 4 shows the best estimate of plant output
The calculation process for displaying guidance on the CRT screen as PV and giving guidance to the operator is shown.

FIG. 5 shows a process of calculating a predicted value deviation ΔY for performing state feedback control.

Next, FIG. 6 shows an example of a plant control system incorporating the predictive control device 10 of FIG.

Here, the Kalman filter gain K is obtained from the variation (error variance V, W) of the measured value of the plant input / output as described above, and the optimum feedback gain G is obtained by the following equation.

G = (R + P ^T QP) ⁻¹ P ^T Q (9) where P is a change prediction matrix of the plant output Y generated by the plant operation amount (input) U, and Q and R are
It is an output weighting factor when calculating an optimal feedback gain in the linear square optimal control law. In practice, Q is a tuning parameter of the operator and can be adjusted manually. R is automatically set inside the system.

Reference numeral 16 in FIG. 6 denotes a step response model calculation unit for giving a step input and checking the response of the control system to the input.

Further, the plant 30 receives a control signal from the prediction control device 10.
Flow controller 31, control valve 3 controlled by CV
2. It includes a transmitter 33, a flow analyzer 34, and the like, and outputs of the transmitter 33 and the analyzer 34 are supplied to the predictive control device 10 as a plant input and a plant output, respectively. Flow controller
The input unit 31 is connected by a switch 35 to one of a CAS (cascade) terminal for inputting a control signal CV and an AUT (automatic) terminal for inputting an LSP (local set point).

Although the above embodiment is a one-input one-output system, according to the present invention, the input is extended to a multi-input system by adding a process disturbance to the input. This can be realized only by changing the plant input section in FIG. 2 to multiple inputs as shown in FIG. That is, the each of the plurality of plants input measured values (1) of the B _m
.DELTA. (K) (m = 1, 2,...) Are calculated, and their sum is calculated to obtain an estimated state quantity ⁰ (k).

In FIG. 7, “plant input 1” is “plant input” in FIG. 6, but “plant input 2” and “plant input 3” are disturbances of the process to be controlled (in this case, plant). The disturbance inputted here is different from noise, and is a change in a physical quantity that is grasped in advance, such as a change in a production amount or a raw material in a plant, a change in pressure or temperature, and the like.

FIG. 8 shows a plant control system of a two-input one-output system as a specific example of FIG. By treating the process disturbance as the second plant input to the predictive control device having two inputs, the accuracy of the predictive control can be improved, and the same effect as that of the feedforward control can be obtained. The step response model calculation unit 16 is also provided with two step inputs.

The embodiment has been described above, but the present invention is not limited to this. For example, an apparatus embodying the present invention can be configured for each block as shown in FIG. 1, but each operation can be executed by one computer. Also, the circuit configuration of each unit and the form of a signal representing a measured value or the like can be arbitrarily determined according to the configuration and function of the target system.

[Effects of the Invention] As described above, in the present invention, the prediction and the noise filtering are simultaneously performed using the Kalman filter represented by the model including the disturbance element in the state observer, so that the following effects are obtained.

Both input and output white noise associated with observation of the input and output of the plant is reduced.

Since noise filtering and prediction are realized at the same time, calculation efficiency is high.

The estimated (predicted) value after noise filtering is a statistically optimum estimated value with respect to the true value, and the significance of the filtering is clear.

When an error occurs in the observed value of the plant input, the error can be corrected in a feedforward manner for the future value.

The calculation of the filter gain is uniquely determined (automatically calculated) by the variance of the noise accompanying the plant input and the plant output, so that it does not take much time.

As a result of the filtering, the vibration of the controller output caused by the noise is reduced.

By inputting the measured values of the input and output of the process to be controlled and the disturbance to a Kalman filter represented by a model including a disturbance element, precise control becomes possible.

In conventional predictive control, it is necessary to accumulate past time-series data on plant input, but this is not necessary in the present invention.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an apparatus for performing predictive control according to the present invention, FIGS. 2 and 3 are diagrams showing a calculation process of a Kalman filter used for a state observer, and FIGS. 4 and 5 are respectively FIG. 3 is a block diagram showing a process of calculating a predicted value for screen and a predicted value ΔY for state feedback from the output of the Kalman filter in FIG. 3, FIG. 6 is a configuration diagram of a plant control system to which the present invention is applied, FIG. Is a diagram showing a case where the plant input unit of FIG. 2 is extended to multiple inputs, FIG. 8 is a configuration diagram of a plant control system using a 2-input 1-output control device, and FIG. 9 is an example of a conventional predictive control method. FIG. 10 and FIG. 11 are graphs showing predicted values according to the conventional method. DESCRIPTION OF SYMBOLS 1 ... Prediction value generation part, 2 ... Control signal generation part, 3 ... Plant output observation part, 4 ... Noise filter, 10 ... Prediction control device, 11 ... PV calculation part, 12 ... State observer 13 Kalman filter gain calculation unit 14 Prediction value calculation unit for screen display 15 CRT 16 Response model calculation unit 21 Control calculation unit 22 Deviation generation unit 23 ... Control signal generator, 24 ... State feedback gain calculator, 25,26 ... Switching means, 30 ... Plant.

Claims (1)

(57) [Claims] 1. A prediction control device for calculating an output predicted value of a control target from a model expressing dynamic characteristics of a control target process and an input to the control target process, and outputting a control signal based on the predicted value. A Kalman filter represented by a model including a disturbance element as the state observer of the controlled object, and calculating the predicted value by inputting a measured value of input and output of the controlled object and disturbance to the Kalman filter. A predictive control device for an industrial process, wherein the predictive control device is configured to simultaneously perform noise reduction and noise removal.