Soft sensor modelling for optimization of distribution control system in oil refineries by applying hybrid data mining techniques

A data-driven soft sensor is a sensor that uses data from available online sensors (such as temperature, pressure, and flow rate) to forecast quality attributes that cannot be monitored naturally or can only be measured at a high cost, infrequently, or with long delays. Oil refineries use control systems, which are connected to PLCs or distributed control systems (DCS). The DCS system is the unit responsible for attaining and providing such data as daily reports for the process, to construct soft sensors utilizing past data from the laboratory observations/measurements and processes data. To determine the quality of crude oil, one can consider the prolonged in-depth laboratory-based tests or the rather expensive approach of online analysers. Implementing light naphtha product quality criterion measurement is surrounded with several essential concerns such as missing data, detecting outliers, selecting input variables and training, validating and maintaining the soft sensor which must be addressed and delt with beforehand. Hence, obtaining heavy-duty soft sensors for oil refineries remained a challenge which in return makes it difficult to improve the end product while simultaneously increasing production. The adaptive neuro-fuzzy inference system, a hybrid soft computing technology combining a fuzzy logic system (FLS) and a neural network (NN), was used to develop a virtual sensor adaptive neural fuzzy inference system (ANFIS) in this research. Rough set theory (RST) and its discretization approach were used to minimize the fuzzy rule sets and redact characteristics of the decision table attributes. It was then used to create the soft sensor modelling for ANFIS, while using the discretization method helped in converting continuous data into a comprehensible data mining format that can be used for data mining. This research is aimed at monitoring and controlling light naphtha production by examining the American petroleum institute gravity (API gravity) and Reid vapour pressure (RVP) variables in real time. It further aims at breaking the privacy barriers between the oil industries as well as soft sensor modelling for data source interpretation to predict API gravity and RVP in real time for crude oil unit's top splitter in the refinery. By comparing the prediction models to other machine learning techniques, with regard to the proposed prediction model, the root means square error improved to be 0.019 for RVP prediction model and 0.4137 for API prediction model and the determination of correlation yield satisfactory results with value of 0.96 for RVP prediction model and 0.99 for API prediction model , the model has been proven to be accurate, and the simulated soft sensor model has been employed as feedback for the cascade PID controller. Whether for operator information, cascaded to base-layer process controller, or multivariable controllers, the proposed virtual sensor model can competently replace the actual online analyser. The objectives of this research were realized by the optimization of the controller of the splitter in the crude distillation unit of the AlDoura Oil Refinery's crude distillation unit. The ability to translate the expert's knowledge into the created model using the gaussian membership function, as demonstrated by the ANFIS model, results in excellent generalisation ability. It has been determined that the Al Doura oil refinery's real-time process was explored, and the data received from these two sources was used to expand the information provided by the data collected. Feedback measurement values from a cascade controller positioned at the top of the splitter in a rectifying section's crude distillation unit (CDU) are used to determine each response variable. A steady-state control system was achieved through the incorporation of an embedded virtual sensor into the suggested adaptive soft sensor paradigm. For the oil refinery's quality control, a cascade ANFIS controller and a soft sensor model were used in the predictive control system's implementation to keep the distillate product's purity within the stipulated range. Overshoots and undershoots are eliminated in the proposed ANFIS-based cascade control compared to the conventional proportional-integral-derivative (PID)-based cascade control. The rise time and settling time are also significantly improved by 26.65 percent and 84.63 percent. Results from other machine learning techniques are also compared to those from the prediction and control models.

Text ALI HUSSEIN HUMOD AL JLIBAWI - IR.pdf Download (1MB)

Actions (login required)

View Item