Industrial, Large-Scale Model Predictive Control With Structured Neural Networks
Tech ID: 31733 / UC Case 2020-085-0
In process industries, model predictive control (MPC) is the industry standard of advanced production operations. All of these systems solve an optimal control problem which regulates their output, and the problem is solved at intervals determined by the many controllers. Solving the optimal control problem can be a slow calculation and there are many instances where a model can be inaccurate. Increasing the speed of this calculation and the accuracy of the control model would significantly enhance the efficiency of controlled environments.
Researchers at the University of California, Santa Barbara have developed the resources to train a neural network that optimizes large-scale control processes at a higher speed. Large, integrated production facilities could be evaluated as a whole rather than piece by piece because of the comprehensive decision-making accuracy and rapid execution of a trained neural network. Solving the optimal control problem, the calculation that slows down current MPC solutions, is executed offline with this technology. The dynamic system state and steady-state variables that are calculated offline are then delivered to the highly accurate neural network to execute final and quickest step when the system is online. The online implementation of this technology will be largely independent of a plant’s size and/or complexity and thus retains its benefits in myriad applications.
- Highly accurate control problem solutions
- Unprecedented MPC speed
- Multi-process control evaluation
- Scalable implementation
- Production facilities
- Chemical plants
- Pharmaceutical controls