Real-Time Optimization (RTO)

In previous chapters we have emphasized control system performance for load

and set-point changes.
Now we will be concerned with how the set points are specified.

In real-time optimization, a computer is used to optimize set points for control loops.

Implementation of RTO/Supervisory Control:

1) Set point calculations are performed on a digital computer (e.g., steady-state

optimization).
2) Set points could be transmitted to either:
(a) analog controllers (before 1990)
(b) digital controllers (after 1990)
3) Case (a) is referred to as "analog supervisory control"; case (b) is "direct digital
control'.
Selection of Processes for RTO

Sources of information for the analysis:

1) Profit and loss statements for the plant

Sales, prices
Manufacturing costs etc.
2) Operating records
Material and energy balances
Unit efficiencies, production rate etc.

Categories of Interest:

1) Sales limited by production

Increases in throughput desirable
Incentives for improved operating conditions and schedules.
2) Sales limited by market
Seek improvements in efficiency.
Example: Reduction in manufacturing costs (utilities, feedstocks)
3) Large throughput units
Small savings in production costs per unit are greatly magnified.

4) Process Variability
Excursions in process variables => offspec products and a need for
larger storage capacities.
 Reduction in variability allows set points to be moved closer to a
limiting constraint, e.g. product quality.

5) High Raw Material or Energy Consumption

e.g., minimize energy consumption by optimal allocation of fuel supplies
and steam.

6) Losses of Valuable Components in waste Streams

Detect via chemical analysis plant exit streams (e.g., air and water). Adjust
air/fuel ratio in furnace to minimize hydrocarbon consumption.

Common Types of Optimization Problems

1. Operating Conditions
Tower reflux ratio
Reactor temperature
2. Allocation
Fuel use
Feedstock selection
3. Scheduling
Cleaning (e.g., heat exchangers)
Maintenance
Batch processes

Formulation and Solution of Optimization Problems (see book by Edgar et al.,

Optimization of Chemical Processes)

In order to perform on-line optimization, a series of steps are required:

Step 1: Determine the process variables of interest.

Step 2: Definition of the Objective Function
 Difficult part of the problem!
Example: Minimize the amount of offspec product from a distillation
column while avoiding a flooding condition.
- Relate off-spec product to costs of utilities and feedstocks.
Treat flooding condition as a constraint on vapor and
liquid flow rates.
Specific objective functions will depend on plant configuration and supply
and demand.

(See TABLE 1 - next slide)

TABLE 1: Objective Functions for Distillation Columns

Step 3. Development of Process Models

Process model is used in on-line optimization. Consists of:
- Equality constraints which relate principal process variables.
- Inequality constraints (e.g., physical limits on pumps, compressors,
metallurgical limits)
Process model is usually a steady-state model.
Some batch optimization methods use the process as the model. (e.g.,
design of experiments)

Step 4. Simplification of the Process Model

Reduce the size of the optimization problem as much as
possible without losing the essence of the problem.
- Ignore process variables which have a negligible effect on the objective
function.

Step 5. Computation of the Optimum

Choose a computational technique to determine optimum.
Virtually all optimization techniques are iterative in nature linear
programming, nonlinear programming.
Good estimate of the optimum accelerates convergence.
- Experience on which inequality constraints might be active is also useful.

Step 6. Sensitivity Studies

 Can be used to determine which variables and parameters are most
important in determining the optimum.

One Dimensional Search Techniques

Selection of a method involves a trade-off between the number of objective function
evaluations (computer time) and complexity.
(1) "Brute Force" Approach
Small grid spacing (x) and evaluate f(x) at each grid point can get close to the
optimum but very inefficient.
(2) Equal Interval Search
Example: Compare objective function values at 3 equally spaced
points.
Suppose we are interested in the region, a x b . Thus initially,
the "region of uncertainty, L" is L0 = b - a .
Consider two cases:

For case (1), maximum lies in (x2, b).

For case (2), maximum lies in (x1, x3).
 Polynomial fitting technique

Fit quadratic polynomial to three data points. Find analytical optimum,

then calculate f(x*). Discard worst value of f, then repeat the process.

Multivariable Optimization
Computational efficiency is important.

"Brute force" techniques are not practical for problems with more than 3
or 4 variables to be optimized.

Typical Approach: Reduce the multivariable optimization problem to a

series of one dimensional problems:
(1) From a given starting point, specify a search direction.
(2) Find the optimum along the search direction, i.e., a one-dimensional search.
(3) Determine a new search direction.
(4) Repeat steps (2) and (3) until the optimum is located

Two general categories for MV optimization techniques:

(1) Methods requiring derivatives of the objective function.
(2) Methods that do not require derivatives.

Nonderivative methods are attractive for real-time optimization

applications where a process measurement is used in place of the objective
function.
Two nonderivative methods which have been used in industrial
applications:
(1) Evolutionary Operation (EVOP)
(2) Simplex or Pattern Search Method.

Evolutionary Operation (EVOP)

Procedure:
(1) Choose a base point.
(2) Evaluate objective function at base point and a set of regularly spaced points around
the base point.
For 2 variables, form a square around the base point.
For 3 variables, form a cube.
(3) Move the process to the point that has the largest value of the objective function.
(4) Use this point as the new base point.
(5) Repeat steps (2) and (4) until no further improvement occurs.
 Disadvantage of EVOP:
- It is slow due to the large number of steadystate points
which must be evaluated at each iteration.

Sequential Simplex Method [8]

Uses a regular geometric figure (a simplex) as a basis.
- Two variable problem: use an equilateral triangle.
- Three variable problem: use a regular tetrahedron.
Objective function is evaluated at the vertices of the simplex.
General direction of search is projected away form the worst vertex.
Thus search direction can change and only one new operating point must be evaluated at e
Example: The two variable problem, max f(x1, x2)

Assume

f1 min f ( f1 , f 2 , f 3 )

Points 2,3, and 4 form a new equilateral triangle.

The pattern search approach usually results in a zigzag pattern as it
moves to the optimum (see figure 20.6-next slide).

Constrained Optimization
Optimization problems commonly involve equality and inequality
constraints.
Nonlinear Programming (NLP) Problems:
a) Involve nonlinear objective function (and possible nonlinear constraints).
b) Efficient off-line optimization methods are available (e.g. conjugate gradient, variable
metric).
c) On-line use? May be limited by computer time and storage requirements.
Quadratic Programming (OP) Problems:
a) Quadratic objective function plus linear equality and inequality constraints.
b) Computationally efficient methods are available.
 Linear Programming (LP) Problems
Both objective function and constraints are linear.
Solutions are highly structured and can be rapidly obtained.

Linear Programming (LP)

Has gained widespread industrial acceptance for on-line optimization, blending etc.

Linear constraints can arise due to:

1. Production limitation e.g. equipment limitations, storage limits, market constraints
2. Raw material limitation
3. Safety restrictions, e.g. allowable operating ranges for temperature and pressures
4. Physical property specifications e.g. product quality constraints when a blend pro

n
P y i Pi
i 1

5. Material and Energy Balances

- Tend to yield equality constraints.
- Constraints can change frequently, e.g. daily or hourly.
 Effect of Inequality Constraints
- Consider the linear and quadratic objective functions on the next page.
- Note that for the LP problem, the optimum must lie on one or more constraints.

General Statement of the LP Problem:

n
max f ci xi
i 1

subject to: xi 0 i 1, 2,K , n

n

a
j 1
ij x j bi i 1, 2,K , n

Solution of LP Problems
- Simplex Method
- Examine only constraint boundaries
- Very efficient, even for large problems