Neural-network based on-line adaptation of model predictive controller for dynamic systems with uncertain behavior

Model-based controllers have positioned themselves in industrial applications, working mainly on top of a layer of PID controllers. Their implementation takes an important amount of time because of the required PID tuning and the model characterization/identification. This paper presents a strategy to perform on-line adaptation for the dynamic matrix coefficients in a DMC controller. Based on the observed PH (Prediction Horizon) elements of the response and controller signal vectors, and based on a non-residing control horizon controller design, the direct control problem is reformulated using the full-effect dynamic matrix (PHxPH) as an unknown. Data is collected and used in two directions: training a RAWN Network (Random Allocation Weight Neural Network), to describe recently observed process behavior, and to solve a least-squares problem for a set of linear equations where the unknowns are the characteristic response coefficients. The paper presents the effect of both approaches, illustrating the adaptation algorithms operation in a highly non-linear process where the controller is designed in a low-gain region. Then the process operating condition is shifted so that it moves to a high-gain region to observe controller response.