1. Towards Integrated Dynamic Real-Time Optimization and Control of Industrial Processes
Two-level strategy
Vertical decomposition
Optimal process operation
Objectives:
Maximize profit
On-spec production
Feasible operation profiles
Implications
Complex dynamic optimization and control problem
Involves repetitive decision making
D-RTO
MPC
Constraints:
Changing market conditions
Process disturbances
Operational & safety constraints
Vertical decomposition approach
D-RTO
D-RTO trigger
Estimation
D-RTO time-scale
MPC time-scale
MPC
Decomposition based on objectives  economic optimization (D-RTO) & tracking (MPC) subproblems
Different models, derived from a first principle model, different set of constraints at each level
Re-optimization may not be necessary at each sampling time
Plant (model)
(incl. base control)
Sensitivity analysis for D-RTO
Interplay between D-RTO and MPC
D-RTO
MPC
updated
D-RTO trigger: sensitivity analysis
Lagrange function sensitivities w.r.t. all estimated disturbances
Optimal solution sensitivities w.r.t. all estimated disturbances
Economic objective only at D-RTO level and disturbance rejection at MPC level
soft constraints can be moved from MPC to D-RTO
Consistent process models ( , ) at each level
a LTV model along reference trajectories by linearization of D-RTO model at MPC level
Longer time horizon for D-RTO to ensure feasibility
compute
compute
and changed active constraint set
One sensitivity integration of process model at each sampling time using previous D-RTO results (and active constraint set) at is required
Compute change in sensitivities ( ) and Lagrange function ( ) can be then calculated
Solution to QP problem:
using second order information (Hessian of Lagrange function)  optimal sensitivities
using first order information  feasible only sensitivities
updates as
D-RTO trigger for a potential re-optimization and quick feasible updates of
based on disturbance sensitivity analysis of optimal solution
a re-optimization is triggered only if the detected persistent disturbances have high sensitivities and causes active constraint set change
uref, yref
If and are larger than a threshold value Sth and changed active constraint set is predicted, a re-optimization should be done
Else linear updates based on optimal solution sensitivities are sufficient
Application to an industrial process
Problem description
Dynamic optimization results and feasible updates due to change in parameter 1 M
Polymer
Catalyst
Reactor
Solvent
Recycle monomer
Separation
Monomer
Continuous polymerization process with frequent grade changes
Optimal grade change problem
minimize off-spec polymer production and transition time
fresh and recycle monomer feed rate and catalyst feed rate as control variables
two uncertain reaction parameters and open-loop unstable operation
Process model of about 2000 DAEs
Reaction parameters randomly perturbed between their bounds
Re-optimization done only when necessary  steer to desired grade
Feasible updates only possible up to +4% change in parameter 1
The integrated D-RTO and control with embedded sensitivity analysis is well suited for large-scale industrial process operation
A software package with integrated dynamic optimizer, MPC, EKF and OPC data server is developed
Nominal solution
Reoptimized solution
Fesible updates