Learning based Model Predictive Control of a High-Altitude Simulation Chamber

Abstract

The limitation of conventional methods to explicitly model and monitor physical systems emanates from system complication, uncertainties, and so forth. Artificial intelligence approaches, Artificial Neural Networks in particular, resolve this difficulty by efficiently capturing the pattern of physical systems and exploring key relationships of determinant parameters effectively. The development of an artificial neural network model to catch the interrelation of the input and output variables was successful. Time series data collected using a variety of combinations of control variables were used to train a sequential model which predicted the chamber temperature from the inputs of control variables. From such a black box model, developing a model predictive controller that predicts upcoming events and sets control actions accordingly was developed. Optimization is a major essence of Model Predictive Control as each suggested step by the controller must be optimized to the required control law. Such optimization is better realized in mathematically modeled systems. But, for non-linear and non-convex relationships as in neural networks, this is cumbersome. This difficulty is addressed by the use of input convex neural networks which relate model outputs with the inputs in a convex relationship. Optimization is relatively simpler when convexity is granted. Finally, a 3 layers input convex neural network that represent the system specifications was developed and optimized control steps were generated using COBYLA (Constrained Optimization by Linear Approximation) solver.