Adaptive model predictive process control using neural networks

• US 5,659,667 A
• Filed: 01/17/1995
• Issued: 08/19/1997
• Est. Priority Date: 01/17/1995
• Status: Expired due to Fees

• Current Assignee: Regents of the University of California (University of California)
• Original Assignee: Regents of the University of California (University of California)
• Inventors: Baum, Christopher C., Jones, Roger D., Buescher, Kevin L.
• Primary Examiner(s): Downs, Robert W.
• Assistant Examiner(s): CHUNG, JI YONG DAVID
• Application Number:
  US08/373,736
• Time in Patent Office: 945 Days
• Field of Search: 364/148, 364/149, 364/178, 364/173, 364/151, 395/22, 395/23, 382/236, 382/238
• US Class Current: 706/23
• CPC Class Codes: G05B 13/027 using neural networks only
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

A control system for controlling the output of at least one plant process output parameter is implemented by adaptive model predictive control using a neural network. An improved method and apparatus provides for sampling plant output and control input at a first sampling rate to provide control inputs at the fast rate. The MPC system is, however, provided with a network state vector that is constructed at a second, slower rate so that the input control values used by the MPC system are averaged over a gapped time period. Another improvement is a provision for on-line training that may include difference training, curvature training, and basis center adjustment to maintain the weights and basis centers of the neural in an updated state that can follow changes in the plant operation apart from initial off-line training data.

Claims

1. An improved method for controlling at least one process output parameter of a plant with a control value generated by adaptive model predictive control (MPC) using a neural network, the process improvement comprising:

(a) repetitively sampling at times t(k) a process output parameter and associated control value at time intervals k having a first duration;

(b) sequentially storing said process output parameters and associated control values sampled at each of said time intervals k over rg and sg of said time intervals k, respectively, where g is an integer greater than one and defines a gapping time interval g at a second interval duration greater than said first interval duration, and r and s are arbitrary integers greater than one and determined by the size of a register for storing said process output parameters and associated control values;

(c) forming from stored ones of said process output parameters and associated control values a gapped network state vector comprising a sequence of process output parameters selected at times t(k), (y(k-g+1),y(k-2g+1), . . . ,(y(k-rg+1)), and averaged control values, (u(k-g+1),u(k-2g+1), . . . ,u(k-sg+1)), where (u(k-ig+1)=(u(k-ig+1)+u(k-ig+2)+. . . +u(k-ig+g))/g;

(d) applying said gapped network state vector to a controller for outputting an updated control value to apply to said plant at time t(k+1) after time t(k); and

(e) repeating steps (a) through (e) at subsequent time intervals of said first time duration to maintain said process output parameter at a selected value.

2. An improved plant control method according to claim 1, further including the steps of:

(f) training said neural net on historical control values and associated ones of said process output parameters to obtain a first set of learned weights and basis centers for operation of said neural net;

(g) continually inputting said output process and said network state vector at each one of said first time intervals to an on-line training processor for outputting updated weights and basis centers to said neural net for use at a next one of said first time intervals.

3. An improved plant control method according to claim 1 further including the steps of:

inputting said gapped network state vector to said neural net for outputting a predicted process output parameter at time t(k+1);

determining an error between said predicted process output parameter at time t(k) and said process output parameter at time t(k); and

adding said error to said predicted process output parameter to form a first corrected predicted process output parameter at time t(k+1).

4. An improved plant control method according to claim 2 further including the steps of:

inputting said gapped network state vector to said neural for outputting a predicted process output parameter at time t(k+1);

determining an error between said predicted process output parameter at time t(k) and said process output parameter at time t(k); and

adding said error to said predicted process output parameter to form a first corrected predicted process output parameter at time t(k+1).

5. An improved method for controlling at least one process output parameter of a plant with a control value generated by adaptive model predictive control (MPC) using a neural network, the process improvement comprising:

(a) repetitively sampling at times t(k) a process output parameter y_k and associated control value u_k at time intervals k having a first duration;

(b) sequentially storing said process output parameters and associated control values sampled at each of said time intervals k;

(c) forming from stored ones of said process output parameters y_j and associatsd control values u_j a network state vector comprising a sequence of process output parameters and control values;

(d) applying said network state vector and said output process parameter at time t(k) to an on-line training processor;

(e) outputting updated values of weights and basis center locations to said neural net for use in predicting a future output process parameter including the steps of

retaining a selected set of associated output process parameters and control values from said network state vector at time t(k-1) to form a net input vector; and

processing said net input vector and said basis center by determining the distance between said net input vector and said basis center and moving said basis center toward said net input vector if said distance exceeds a first selected value or adding a new basis center if said distance exceeds a second selected value; and

(f) repeating steps (a) through (e) at subsequent time intervals t(k+1) of said first time duration to update said weights and basis center locations after each said first time interval duration.

6. The improved plant control method of claim 5, wherein the step of outputting updated values of weights and basis center locations further includes the steps of:

retaining a selected set of associated output process parameters and control values from said network state vector at time t(k-1) to form a net input vector;

processing said net input vector and said output process parameter at time t(k) to output a first updated value of said weight; and

further processing said net input vector and said output process parameter at time t(k) with difference training to output a second updated value of said weight from said first updated value of said weight.

7. The improved plant control method of claim 6, further including the step of yet further processing said net input vector and said output process parameter at time t(k) with curvature training to output a third updated value of said weight from said second updated value of said weight.

8. The improved plant control method of claim 6, wherein the step of further processing said net input vector and said output process parameter includes the steps of:

retaining values of said net input vector at time t(k-2) and said output process parameter at time t(k-1);

determining a net input vector difference value between said net input vector at time t(k-1) and said net input vector at time t(k-2);

determining an output process parameter difference value between said output process parameter at time t(k-1) and said output process parameter at time t(k-2); and

processing said net input vector difference and said output process parameter difference with said first updated value of said weight to output a second updated value of said weight.

9. The improved plant control method of claim 8, wherein the step of yet further processing said net input vector and said output process parameter further includes the steps of:

retaining values of said net input vector at time t(k-3) and said output process parameter at time t(k-2);

determining a net input vector difference value between said net input vector at time t(k-2) and said net input vector at time t(k-3);

determining a second net input vector difference value between said net input vector difference at time t(k-2) and said net input vector difference at time t(k-3);

determining a output process parameter difference value between said output process parameter at time t(k-2) and said output process parameter at time t(k-3);

determining a second output process parameter difference value between said output process parameter difference at time t(k-2) and said output process parameter difference at time t(k-3); and

processing said second net input vector difference and said second output process parameter difference with said first updated value of said weight to output a third updated value of said weight.

10. The improved plant control method of claim 5, wherein the step of outputting updated values of weights and basis center locations further includes the steps of:

retaining a selected set of associated output process parameters and control values from said network state vector at time t(k-1) to form a net input vector;

processing said net input vector and said basis center to adjust said basis center if a predicted output process parameter differs from said output process parameter at time t(k) by a predetermined value.

11. An improved control system for controlling at least one process output parameter of a plant by generating a control value using adaptive model predictive control (MPC) with a neural network, the control system comprising:

a register for repetitively sampling and storing process output parameters and associated control values at times t(k) with first time intervals k having a first duration, said register sequentially storing said process output parameters and associated control values sampled at each of said first time intervals over rg and sg, respectively, of said first time intervals, where g is an integral multiple of k greater than one and defines a gapping interval g at a second time duration greater than said first time duration, and r and s are arbitrary integers greater than one and determined by the size of said register for storing said process output parameters and associated control values;

an electronic processor for forming from stored ones of said process output parameters and associated control values a gapped network state vector comprising a sequence of process output parameters selected at g time intervals, (y(k-g+1),y(k-2g+1), . . . ,(y(k-rg+1)), and averaged control values, (u(k-g+1),u(k-2g+1), . . . ,u(k-sg+1)), where (u(k-ig+1)=(u(k-ig+1)+u(k-ig+2)+. . . +u(k-ig+g))/g;

a controller for receiving said gapped network state vector and outputting an updated control value to apply to said plant at time t(k+1) after time t(k).

12. An improved control system according to claim 11, further including an on-line training processor connected to receive said gapped network state vector and said process output parameter at time t(k) and outputting updated values for internal weights and basis center locations to said neural net.

13. An improved control system according to claim 11, further including:

a delay register for receiving a predicted value for said output process parameter at time t(k+1) and outputting a predicted value of said output process parameter at time t(k);

a subtraction circuit for outputting a difference between said output process parameter at time t(k) and said output process parameter predicted at time t(k); and

an addition circuit for adding said difference to said predicted output process parameter at time t(k+1) and outputting a corrected predicted output process parameter at time t(k+1).

14. An improved control system for controlling at least one process output parameter of a plant by generating a control value using adaptive model predictive control (MPC) with a neural network, the control system comprising:

a register for repetitively sampling and storing process output parameters and associated control values at times t(k) with first time intervals k having a first duration, said register sequentially storing said process output parameters and associated control values sampled at each of said first time intervals over rg and sg, respectively, of said first time intervals, where g is an integral multiple of k greater than one and defines a gapping interval g at a second time duration greater than said first time duration, and r and s are arbitrary integers greater than one and determined by the size of said register for storing said process output parameters and associated control values;

an electronic processor for forming from stored ones of said process output parameters and associated control values a network state vector comprising a sequence of process output parameters and control values; and

an on-line training processor connected to receive said network state vector and said process output parameter at time t(k), wherein said on-line training processor further includes;

electronic circuitry for adjusting basis centers used by said neural net including;

an error processor for determining an error value between said output process parameter at time t(k) and a predicted output process parameter at time t(k) using a current location for said basis function;

a basis center location processor for determining a distance between said net input vector and said basis center; and

logic circuitry for adjusting said basis center if said error exceeds a predetermined error value or if said basis center location exceeds a predetermined distance from said net input vector;

electronic circuitry for forming a net input vector from said network state vector and outputting said net input vector at time t(k-1); and

a weight training processor receiving said output process parameter at time t(k) and said net input vector from time t(k-1) and outputting a first updated value of processing weights for said neural net.

15. The improved control system of claim 14, wherein said weight training processor further includes difference training circuitry for receiving a value of said output process parameter at times t(k) and t(k-1), said net input vector at times t(k-1) and t(k-2), and said first updated value of said weight and outputting a second updated value of said weight.

16. The improved control system of claim 14, wherein said weight training processor further includes curvature training circuitry for receiving a value of said output process parameter at times t(k-1) and t(k-2), said net input vector at times t(k-2) and t(k-3), and said second updated value of said weight and outputting a third updated value of said weight.

17. The improved control system of claim 14, wherein said on-line training processor further includes electronic circuitry for adjusting basis centers used by said neural net.