Endüstriyel fraksiyonlandırma kolonunun simulasyon hatalarını veri-bazlı modelleme teknikleri ile tahmin edilmesi

Safe and efficient operation of refinery units remains the prime concern for refineries. The daily growth of fuel consumption and the necessary of proving it show that operating of a refinery unit in economical way is vital case. Unfortunately, the downstream refining industry faces many technical and operation difficulties due to the introduction or increase of environmental restrictions that lead to tighter product specification that require investing in new refinery technologies, improved process operations, as well as adaptation to the changes in demands. For most of the refinery units it is usually possible to improve the operating conditions with small touches and these tiny operational improvements often mean a huge economic benefit and is worth of effort. Operational improvements depend on accurate estimates of process behaviours, which can be achieved both by using process simulators, which embody first-principle models of the processes, and/or by data-driven models derived with operation data. Due to the inevitable deviation of theoretical models from the actual behavior of the process, many process engineers prefer data-driven. On the other hand; compared to data-driven models, it is possible to extract much more information about the process by means of a process simulator. In this thesis, we propose and test a novel approach that corrects process simulator outputs for product qualities of an industrial hydrocracker fractionator column located in Turkish Petroleum Refinery Co. with data-driven models. The data-driven models use Support Vector Regression, Artificial Neural Networks, Random Forest, and linear regression model. With such an approach, process engineers can both correct some crucial actual outputs (product qualities) of process simulator with an acceptable error and at the same time they can extract more information related to the process, which is not possible to obtain with only data-driven models. We also introduce a method for feed characterization with operational data, which is a critical input of the process simulators. This enables using the process simulator for any time. Finally, we show six different approaches for simulation error modelling and evaluate their results by considering the measurement uncertainties.