Introduction
In the chemical industry, efficiency is paramount. Plants producing materials like Low-Density Polyethylene (LDPE) operate under tight margins where minor improvements in energy consumption or waste reduction translate into millions of dollars in savings. Optimizing these processes requires solving complex Nonlinear Programming (NLP) problems. However, two major hurdles persist: the massive scale of models governed by Differential-Algebraic Equations (DAEs) and the "control gap," where optimization takes too long to be useful for real-time plant adjustments.
Methodology
This research presents a framework to bridge the control gap using three core technical pillars:
- Simultaneous Collocation: Instead of solving physics equations and optimization iteratively, this method discretizes the time horizon into finite elements. It treats the plant state at every time step as a decision variable, transforming a dynamic problem into a large-scale algebraic puzzle that can enforce safety constraints at every point.
- IPOPT (Interior Point Optimizer): A large-scale software package designed to solve these puzzles using barrier methods and parallel linear algebra, capable of handling millions of variables.
- Advanced Step NMPC (as-NMPC): This is the key to real-time speed. The controller solves a "background" optimization based on a predicted state. When actual measurements arrive, it uses NLP Sensitivity—a mathematical update—to instantly adjust the pre-calculated solution rather than re-solving from scratch.
Case Study: LDPE Production
The methods were evaluated using a high-fidelity simulation of a commercial-scale LDPE process, which involves complex recycles and a tubular reactor up to 2 km in length.
The model's accuracy was validated by estimating over 100 kinetic parameters across various operating scenarios. As shown in the reactor temperature profiles, the model closely matches actual plant data for different product grades.
Key Results
- Massive Speedup: In grade transition tests, the computation time for a control move was reduced from 351 CPU seconds to just 1.04 seconds using the as-NMPC approach.
- Waste Reduction: The optimization strategies reduced grade transition times from 5 hours to 2.8 hours, preventing the creation of 23 tons of off-spec waste per transition.
- Scalability: Parallel implementation allowed the system to handle up to 32 data sets for parameter estimation simultaneously, overcoming the memory limitations of standard serial solvers.
Conclusion and Future Work
While the as-NMPC method provides the speed necessary for real-time operations, its effectiveness depends on the accuracy of background predictions. Future research will focus on explicitly handling model uncertainty and measurement noise, as well as better integrating high-level business planning with low-level process control.
Summary updated