Design of an advanced control system and a dual-horizon optimizer for a refinery thermal cracking furnace through predictive fouling modeling

In industrial processes, such as oil refining, after addressing the fundamental requirements of ensuring the essential functionality and safety of the control system through the basic control layer, the next step involves process optimization. Advanced process control (APC) applications enable the stabilization of operations and the achievement of measurable economic goals. Notably, APC applications often rely on model predictive control (MPC). The success of MPC is rooted in its ability to provide a systematic and standardized approach to multivariable process control, incorporating constraints and utilizing models derived from experimental tests. Additionally, it can employ non-linear models, such as those derived from system physics, to predict future process behavior. This Thesis addresses the design of an APC system with a dual-horizon optimizer for a refinery thermal cracking process through predictive fouling modeling. The APC models developed are divided into: • short-term models, developed using AspenTech's DMC3™ software, with the aim of providing short-term predictive control; • a long-term model based on a time series model, developed in MATLAB, for predicting furnace fouling. Both models are equipped with an optimizer to achieve short-term objectives, including maximizing yields while adhering to safety, environmental, and process constraints, as well as long-term objectives to ensure fouling control to reach the end of the run at the set conditions. The project was carried out at API Refinery of Ancona S.p.A., located in Falconara Marittima (AN), Italy, in collaboration with Aspen Technology. The results obtained from the predictive fouling model simulations highlight the possibility of providing a reliable prediction, suggesting a long-term optimization that can yield an increase of over 1% in production, with a substantial economic return. This allows the project to proceed to the commissioning phase for the online implementation of the APC by 2024, along with further model development in collaboration with AspenTech through the utilization of deep learning neural networks.