Machine learning based optimization of thermal cracking furnace in a visbreaker unit

Machine learning (ML) is a branch of artificial intelligence that leverages advanced algorithms to autonomously learn from large datasets, recognize patterns, and make accurate predictions with minimal human intervention. In the oil and gas industry, where refining operations are highly complex, ML-based approaches offer significant advantages over conventional mechanistic models, which often struggle to capture the dynamic nature of refinery processes. One critical unit in a refinery is the Visbreaker, which plays a key role in reducing the production of residual oil during crude oil distillation while increasing the yield of valuable middle distillates, such as naphtha and fuel oil. It operates by thermally breaking down large hydrocarbon molecules in residual oils through high-temperature heating in a furnace, generating lighter hydrocarbons like LPG and gasoline. However, managing the Visbreaker presents challenges, particularly coking of furnace tubes when processing heavy residual feeds. This coke buildup can lead to frequent shutdowns for maintenance, disrupting operations and reducing efficiency. This thesis focuses on ML-based optimization models for refinery processes, particularly the Visbreaker, utilizing real-time sensor data for enhanced decision-making. Machine learning algorithms, including Decision Trees, Random Forests, and Artificial Neural Networks (ANNs), were employed to predict critical parameters such as furnace coil temperatures. The robustness of these models was validated using 500 days of historical data. Additionally, ML models were developed to estimate the remaining operational time before a shutdown. Building upon these predictive models, an AI-driven optimization framework, using an ANN-based genetic algorithm (ANN-GA), was developed to recommend optimal operating conditions. By integrating ML models with real-time data, this study enables proactive decision-making and optimization. Operators can anticipate operational bottlenecks like coking and adjust parameters in advance, ensuring the Visbreaker operates efficiently while minimizing risks. ML-based approaches thus provide a robust solution for managing refinery complexities, enhancing productivity, and improving sustainability.