Desenvolvimento de estratégias para a utilização das técnicas da engenharia 4.0 e tratamento de sinais na construção de controle preditivo de processos em linguagem Python

Abstract

Currently, with the expansion of technology, industrial processes have access to increasingly efficient equipment, allowing for the simultaneous monitoring of multiple aspects. This marked the beginning of a new revolution, Industry 4.0, based on the interconnectivity of data within industrial plants. Consequently, the control field has also adapted, creating what is known as advanced control. Model Predictive Control (MPC) is one of the new techniques developed, based on the use of a dynamic system model to predict its future behavior. In this context, this strategy enables the use of technologies from the new industrial era, particularly regarding artificial intelligence models and machine learning, as internal models for system prediction. Currently, in terms of machine learning applications, deep learning is a subclass based on artificial neural networks and is being widely explored and reported in the literature. Among its models, it is worth highlighting the long short-term memory (LSTM) networks and gated recurrent unit (GRU) networks, as well as the potential for using these techniques coupled with advanced data pre-processing strategies like variational mode decomposition (VMD), referred to as hybrid models. Thus, the methodology of this work aims to develop, implement, and compare, in Python, strategies of classical and advanced control (MPC) for the catalytic reaction of ethylene in a CSTR reactor. Regarding classical control, the proportional-integral-derivative (PID) controller achieved the best performance for servo-type disturbances, i.e., changes in the setpoint of the controlled variable. Additionally, when evaluating its application in a noisy system, it was observed that it did not perform adequately. In terms of advanced control, the controllers used in this work were divided based on the internal model employed. The first group used only deep learning models, LSTM and GRU, and achieved satisfactory performance. When evaluated under noisy conditions, both controllers lost performance but presented better or comparable results relative to the PID application. The second group used hybrid models, combining VMD for noise handling with predictions using LSTM or GRU. However, this group did not achieve the expected performance in any scenario, especially when exposed to noise. At the end of the study, the MPC-LSTM and MPC-GRU controllers achieved the best metrics for noise-free data, while the MPC-LSTM and MPC-VMD-LSTM controllers performed best under noisy conditions. However, the results indicate the need for further refinement of parameter tuning techniques, particularly for hybrid model MPC controllers. In conclusion, the use of MPC compared to PID is promising for more complex applications or those requiring greater controller flexibility. In contrast, for simpler applications, the PID has the advantage of being well-established and widely adapted in the industry, in addition to having significantly shorter execution times than MPC.