Control of Waste Heat Recovery Systems using Nonlinear Locally Optimal and Machine Learning Methods

Abstract

Waste heat recovery systems are designed to capture thermal energy from mechanical systems that would normally be transferred to their surroundings. Due to the stochastic nature of waste heat sources, control systems implemented to maintain process setpoints often have issues working with the apparent nonlinear, time-varying system. This work proposes using a Trans-critical Organic Rankine Cycle (TORC), where an organic working fluid is evaporated above its critical point, as a waste heat recovery system. The TORC system in this work is modelled as a 13-dimensional dynamic model with additive gaussian noise. An Extended Kalman Filter (EKF) is implemented to construct a full state estimate given a subset of noisy measurements which can be obtained with conventional sensors. Two different control systems are then implemented on this system. The first, a Cerebellar Model Articulation Control (CMAC), involves a proportional control output and a Neural Network learned output which satisfy the Lyapunov stability criterion. The second, an Iterative Linear Quadratic Regulator (ILQR), uses linearized points along a trajectory with a quadratic cost-function minimizing algorithm to choose control outputs. It was found that both the CMAC and ILQR can reliably track process setpoints and exhibit significantly less drift than linear control methods such as Proportional-Integral-Derivative Control.