Learning based algorithms for temperature control and fouling prediction in heat-exchangers

My thesis broadly explores different data-driven algorithmic frameworks for solving two class of problems pertaining to heat-exchangers: temperature control and fouling resistance modeling and prediction. Designing robust and accurate temperature controllers for heat exchangers is very challenging primarily because of complexities associated with the synthesis of dynamics of industrial heat exchangers. It is practically infeasible to synthesize an accurate dynamical model of heat-exchanger's flow and heat transfer physics because of unmodeled system dynamics and random noise. The use of non-model based control approaches such as PI control is limited by the challenges faced in the accurate tuning of its parameters. Further, an optimally tuned PI controller will not be universally applicable across the entire operating range of heat-exchanger's outlet temperatures due to changing controller set-points caused by large-scale stochastic fluctuations in the heat exchanger's flow and temperature inputs. We propose a couple of data-driven temperature control solutions for heat-exchangers based on $n$-step advance deep neural networks and a hybrid control approach based on steady state neural network with a proportional controller. We compared the performance of these two control approaches with PI control and open loop (no control) scenarios on a simulated water-air heat exchanger system. The heat-exchanger control objective was to accurately track the user-defined set-point temperature signal at the air-outlet duct of the heating coil system. We assess the response characteristics of the developed control approaches for two set-point temperature signals- a constant temperature and a high frequency (periodic) square pulse. We obtained over 70% reduction in absolute set-point temperature tracking error compared to a PI controller, over a 50-minute duration for both the developed control approaches. Further, we explained the working of the neural network + proportional control module in great detail along with the rationale behind our design choices for ensuring the controller's stability and robustness. The complexities involved in fouling prediction in heat-exchangers have deterred traditional model based approaches and other empirical methods. Many existing data-driven prediction approaches are typically application-driven or heat-exchanger specific. Hence, these algorithms are limited either by the scale of data they can handle or by the geometrical and flow configurations of the heat-exchanger for which they were conceived. Here, we develop a generalized and scalable statistical model for accurate prediction of fouling resistance using commonly measured parameters of industrial heat-exchangers. This prediction model is based on deep learning which is a family of computational data-driven methods that allows a scalable algorithmic architecture to learn non-linear functional relationships between a set of target and predictor variables from large number of training samples. The efficacy of this modeling approach is demonstrated for predicting fouling in a simulation of a flue-gas and water cross flow heat exchanger designed for waste-heat recovery. The results demonstrate that the average coefficient of determination, which quantifies the accuracy of predictions is over 99% in predicting flue-gas side, water side and overall fouling resistances. This prediction framework is also evaluated under varying levels of input noise and it was demonstrated that averaging predictions over an ensemble of multiple neural networks achieves better accuracy and robustness to noise. We find that the proposed deep-learning fouling prediction framework learns to follow heat-exchanger flow and heat transfer physics, which we confirm using locally interpretable model agnostic explanations around randomly selected operating points.LimitedAuthor requested closed access (OA after 2yrs) in Vireo ETD syste