Resumo:
Real-Time Optimization (RTO) is a widely used tool in the industry to determine optimal operating points, maximizing the economic efficiency of continuous processes. However, optimization convergence failures remain a significant challenge, as the complex interaction between operational variables can prevent the calculation of the best operating condition. To address this limitation, this study proposes the use of machine learning techniques to diagnose the causes leading to RTO non-convergence, utilizing real data from a petroleum distillation unit. The initial statistical analysis enabled the characterization of process variable behavior, while unsupervised clustering methods, evaluated using the Reval library, allowed for the identification of patterns and segmentation of scenarios associated with optimization convergence and non-convergence. Additionally, the application of supervised classification models made it possible to assess the impact of process variables on RTO stability, highlighting those with the greatest influence on optimization failure. Explainable AI techniques, such as SHAP (Shapley Additive Explanations), were employed to interpret the results and provide greater transparency in analyzing critical factors. The results demonstrate that machine learning can be an effective tool for diagnosing RTO failures, enabling the anticipation of adverse conditions and allowing for more precise operational adjustments. Thus, this study contributes to improving the reliability system of real-time optimization and reinforces the importance of integrating data science techniques into industrial process control and optimization systems.
