Industrial applications of hybrid modelling techniques

Abstract

In the present study, the application of hybrid modelling techniques is applied to industrial applications. Many of the studies currently known to the literature for the fields under examination are either purely model-based, theory-based or lab/pilot scale empirical studies. In this work, we present a hybrid approach whereby empirical data is used to form statistical models for relationships where no clear fundamental relationship can be described mathematically. Equally, first-principles models are employed where no suitable data can be gathered empirically. Finally, the process understanding, heuristics and recollections of plant operators, engineers and maintenance personnel can be integrated formally into the decision-making process of process design/optimisation. The first half of this work is concerned with process development of a proprietary modular Gas-to-Liquids process, briefly comprised of a packed bed plate-fin ’mini-channel’ Fischer-Tropsch reactor. Currently, little can be predicted about the flow or temperature performance of a complex reactor geometry in the design phase. Data-driven models provide a simplistic approximation with no added understanding. At commercially relevant scales, the parameters of interest are both costly and hazardous to iterate through empirical trial and improvement. By integrating offline analysis, online data and a novel temperature sensing scheme, we increase the spaciotemporal resolution of data while adding process understanding. The second theme of this work is related to flue gas filtration in large-scale Biomass and Energy-from-Waste Power Generation plants. Flue gas filtration is overlooked as an opportunity for process improvement. We argue that a filtration system designed on the basis of lowest CAPEX, and operated at the lowest maintenance cost will not provide the lowest total cost of ownership. By integrating industrial historic data, maintenance records, commercial data and multivariate modelling methods, we produce a set of recommendations for improved operation. Commercially available solutions are benchmarked in predictive hybrid models on a ROI basis