High Fidelity Approximation of Slow Simulators Using Machine Learning for Real-time Simulation/Optimization

摘要

Simulation and optimization of industrial processes is cost effective and profit productive. Often, high fidelity models require extensive resources to code and require long execution times. In this work, we examine using machine learning techniques to replace simulation models with high fidelity approximations. We test linear genetic programming, linear regression, and machine learning paradigms. The results show that high fidelity approximations (R~2 of 0.99) are possible that execute in a fraction of the time required by the original simulator. These solutions are coded into web services so that a plant manager can input standard information into a user friendly web page, but produce results in a few milliseconds as opposed to hours. This advantage allows for real-time dynamic planning and optimization on the plant floor.