Data-driven multi-model control for a waste heat recovery system

Abstract

We consider the problem of supervised learning of a multi-model based controller for non-linear systems. Selected multiple linear controllers are used for different operating points and combined with a local weighting scheme, whose weights are predicted by a deep neural network trained online. The network uses process and model outputs to drive the controller towards a suitable mixture of operating points.The proposed approach, which combines machine learning and classical control of linear processes, consists in the design of a controller for a waste heat recovery system (WHRS) mounted on a Heavy-Duty (HD) truck engine to decrease fuel consumption and meet the future pollutant emissions standard.The proposed control scheme, which can be applied to any nonlinear system with an existing linear controller bank, is successfully evaluated on an Organic Rankine Cycle (ORC) process simulator and compared to a standard linear controller and to several strong multi-model baselines without learning.

Deep KKL: Data-driven Output Prediction for Non-Linear Systems

“…The idea of the proof is to compare y pθ obtained from h θ with the prediction y p defined in (7) . This concludes the proof.…”

“…These procedures learn the dynamics directly from a set of observations of the system. In its most modern form, Deep Learning, high-capacity deep neural networks are trained from massive amounts of data, with impact on many applications in control theory by complementing classical methods [6], [7] or even replacing them, for instance through Deep Reinforcement Learning [8]. Depending on the concrete application and the amount of available data, recent work tends to demonstrate that neural networks may benefit from hybridization with more classical modeling techniques.…”

“…(40)With Grönwall inequality, it yields,Z(z, p) − Z θ (z, p) ∀(z, p). (41) This implies with y p defined in(7) : |y p (t + p) − y pθ (t + p)| ≤ δ e + p) − y pθ (t + p)| ≤ |y(t + p) − y p (t + p)| + |y p (t + p) − y pθ (t + p)|, (42) and employing Proposition 1 it finally implies |y(t + p) − y pθ (t + p)| ≤ kL 2 e −λt+L1p |z y 0 | − 1 + 1 , (43)where k is obtained from P and L 1 = A + L 2 b and L…”

Experimental Investigation and CFD Analysis of Pressure Drop in an ORC Boiler for a WHRS Implementation

“…In fact, this last one represents up to another 40% of the total fuel energy at maximum power point [ 1 , 2 , 3 , 4 , 5 ]. The Organic Rankine Cycle (ORC) technology is an effective method to recover energy from low-temperature waste heat that has been studied in depth in the last decades and has been implemented in a wide variety of fields in the industry, such as marine engines [ 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 ], light duty vehicles [ 19 , 20 , 21 , 22 , 23 , 24 ], heavy-duty vehicles [ 22 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 ], biodiesel engines [ 35 ], heat and power generation [ 36 , 37 , 38 , 39 ], geothermal energy [ 40 , 41 , 42 ], solar energy [ 43 , 44 , 45 , 46 ], or biogas plants [ 47 , 48 ]. Energy saving, efficiency improvement, and emissions reduction of WHRS based on an ORC vary depending on the application, the energy output, or the exergy effici...…”

Towards improving the estimation performance of a given nonlinear observer: a multi-observer approach

“…Because the gains are different, each mode exhibits different properties in terms of speed of convergence and robustness to measurement noise. We run all modes in parallel and we evaluate their estimation performance in terms of a quadratic cost using monitoring variables, inspired by supervisory control and estimation techniques, see e.g., [18], [26]- [29]. Based on these running costs (i.e., monitoring variables), we design a switching rule that selects, at any time instant, the mode which is providing the best performance.…”