Control
Modeling - For Control
The Aspen Control Platform (ACP) provides three different
controller formulations. The ACP provides the industry-leading
Aspen DMCplus® formulation based on the FIR model form, as well as
linear MIMO State Space and non-linear MISO State Space models.
(Click to
enlarge)
Switching between formulations is accomplished with a simple
mouse click. Older versions of Aspen Nonlinear Controller and Aspen
DMCplus controllers are easily imported with automatic model,
script, and tuning conversion.
The modeling tools are implemented in a modern (.NET)
drag-and-drop environment for model building, simulation,
configuration and deployment. The entire control application can be
built, simulated, and deployed within aspenONE Advanced Process
Control, providing a true competitive advantage with a seamless
interchange between the three industry-leading control
formulations.
| Formulation |
Prediction
|
Filter |
Optimizer |
Controller |
| FIR (Aspen DMCplus) |
FIR models. Current CV predictions are kept in memory and
current moves are added to it. |
Bias is added to each output. First order time constant can be
used to filter the update. Rotation factor used for ramps. |
Upper and lower ranked limits for CVs (linear or quadratic
type). Ranked ETs can be specified on inputs and outputs. Costs on
inputs and outputs. Algorithm is a revised simplex method for LPs.
Interior point for QPs. Interior point can also be used for
LPs. |
Number of moves from 8 to 64 is specified. Matrix
inversion is used to create an analytical solution to the
unconstrained problem. "Clipping" of the inputs is used to enforce
constraints. Option to use QP solution. Also, option added to treat
CV con-straints more rigorously. Problem solution time is cubic in
number of MV and number of moves. |
| State Space MISO |
MISO models. State of the linear part of each model
is maintained and updated each cycle. |
Bias is added to each output. Noise ratio can be
used to filter. Internal states are not updated (except through
prediction). Rotation factor used for ramps. |
Upper and lower ranked limits for CVs (linear or
quadratic type). Ranked targets can be specified on inputs and
outputs (target can be an ET or upper or lower limit or minimum
move). Costs on inputs and outputs. Algorithm is an SQP. Each QP
iteration uses the interior point method. Trust regions are used to
expand the search region. This usually takes 11 QP iterations. |
Number of moves for each MV is specified.
Algorithm is an SQP. Each QP sub-problem is solved with an active
set method. A line search is used to expand the search region. This
usually takes 2 to 6 QP iterations. Problem solution time is cubic
in number of MVs and number of moves. Dependence on number of
states is not known but probably small (and models usually low
order). |
| State Space MIMO |
MIMO models. State of the model is maintained and updated each
cycle. |
Internal states are updated using a Kalman filter. Typical
noise and disturbance used for tuning. Input and "internal"
disturbances can be specified. |
Upper and lower ranked limits for CVs (only quadratic). Ranked
ETs can be specified on inputs and outputs. Costs on inputs and
outputs. Algorithm is an interior point method. |
Number of moves is specified (no limit). Algorithm
is an interior point method - extensively customized to take
advantage of the problem structure. Problem solution time is cubic
in number of MVs, quadratic in number of states, and linear in the
number of moves. |