Machine learning assisted optimization of an industrial visbreaker plant

Abstract

This study proposes a data-driven prediction-optimization framework to improve operational efficiency and extend furnace run length in an industrial visbreaker unit subject to coke formation. Using ten years of real refinery operating data, Decision Tree and Artificial Neural Network (ANN) models were developed to predict furnace coil skin temperatures and the remaining operational days before shutdown. The ANN achieved acceptable test-set Mean Absolute Errors for four critical coils and approximately 15 days for remaining-cycle prediction, corresponding to less than 13 % of a typical furnace run length. The trained ANN was embedded into a Genetic Algorithm to optimize seven controllable operating variables under industrial constraints. This framework contributes to predicted run-length extensions of 7.5-12.5 % during early-cycle operation and up to 50 % near end-of-cycle conditions. These improvements translate into delayed decoking requirements, improved thermal stability, and enhanced maintenance planning. The main contribution of this work lies in the integration of long-horizon industrial data, lag-based dynamic feature representation, and ANN-GA optimization for an industrial visbreaker unit. Unlike prior studies based on simulated or short-term datasets, the proposed framework demonstrates industrial feasibility and provides actionable decision support for proactive coking mitigation and operational optimization.