Regression Optimization for System-level Production Control

We propose a novel generalized prediction-optimization framework to optimize set point controls for a network of processes in a production plant, wherein a regression model is used to capture the physical representation of each process's behavior and the relationship between its inputs and outputs. We introduce a nonlinear optimization problem to model the optimal set-point problem. For piece-wise linear regressors, we reformulate the problem into a mixed-integer linear program. For highly nonlinear models such as deep neural networks, we propose a decomposition primal-dual algorithm for solving it. Using a real-world use case of oil sands processing, we show the benefit of our approach by the ability to efficiently identify a set of feasible control variables, while giving a high production output.