Soft Sensor Design for a Petrochemical Process
Soft Sensor. Oil. NARMAX/NARX models. Systems Identification. Structure Selection. Parameter Estimation. Data Selection.
Oil is an element used as a raw material in several production chains in the most diverse sectors of the economy. In this way, it is possible to say that most sectors are dependent on oil in some way. Furthermore, its extraction process is characterized by the geographic region where the reserves are located, which can be offshore and onshore, i.e., located in deep waters or inland. Offshore oil extraction is a very complex process and, therefore, several equipment and instruments are necessary to carry out and control the production in wells located on the ocean floor. The arrangement used in this process is generally composed of a Stationary Production Units (SPUs), which stores the production and is connected to the other elements on the seabed through pipes (e.g., risers and flowlines); manifolds, which centralize oil production from different wells; Wet Christmas Trees (WCT), which utilizes a set of valves, control the production of oil in the wells; and, finally, between the oil reserve and the WCT, there is the production column of the well. Inside this production column, more precisely at the bottom, is located the Permanent Downhole Gauge (PDG) sensor, which is used to measure the pressure and temperature of the well. This sensor is subjected to extreme operating conditions and, as a consequence, has a very short service life. The exchange or maintenance of this sensor is rarely done because it is difficult to access and requires that production be stopped, which results in great economic losses. Therefore, to overcome the production problem without data from the PDG sensor, the use of soft sensors appears as an alternative. Soft sensors are mathematical models capable of estimating a process variable using other variables as input. In this project, it is proposed to use the systems identification methodology (i.e., i. Dynamic tests, data collection, and pre-processing; ii. Choice of the mathematical representation of the model; iii. Selection of structures for the model; iv. parameters estimation; and v. Model validation.) to model a soft sensor to estimate the output of a PDG sensor, which is used in an offshore oil extraction process. To this end, optimization techniques and computational intelligence algorithms will be implemented and developed to select the best data and database variables, as well as select structures and estimate parameters for polynomial NARX/NARMAX models. Preliminary results demonstrate that multi-objective algorithms, such as NSGA-II, are effective in selecting regressors for polynomial NARX/NARMAX models. The models resulting from this approach are parsimonious and present good behavior in both static and dynamic regimes. Therefore, multi-objective algorithms prove to be very useful and promising for the system identification process and, thus, they will be further investigated and applied, not only for structures selection, but also in the other stages of the project.