Controlador Neurofuzzy com aprendizado on-line: teoria e aplicação na indústria de petróleo

Abstract

This work presents a controller, based on the evolution of the "On-line neurofuzzy controller (ONFC)", which is a simple controller, with low computational cost and adaptive capacity. The development and the analysis of a new learning algorithm, introducing relative aspects of stability, application and tuning methods are presented. This controller, called ONFCDw, can be applied in typical industrial plants and it has showed efficient performance in controlling plants where PID controllers have low performance due to non linearity, frequent changes in the operational point, time varying variable disturbances, etc. Its adaptive nature also gives ONFCDw the advantage of automatically adjusting to temporary changes in processes, avoiding or decreasing the necessity of re-tuning. This algorithm, due to its simplicity, can be implemented in most industrial supervisory systems. In this work it is also presented a successfully implementation of ONFCDw in a temperature control of an area of the main distillation tower in an industrial Delayed Coke unit of Petrobras, being part of a R&D project between UFMG and PETROBRAS. The methodology proposed for implementation included a previous analysis in a conventional dynamic simulator and in a process dynamic simulator, wich is capable to simulate equipments and units typically present in refineries. The promising results motivate the evaluation of these systems in other processes with non linear characteristics and intense external disturbances. Finally, it is presented some proposes for continuity and new possible applications.