Abstract
Fluid catalytic cracking is a crucial process in the refining industry, capable of converting lower-quality feedstocks into higher-value products. Due to the variability in feedstock properties and fluctuations in product market prices, timely adjustment and optimization of the FCC unit are essential. In this context, data-driven models have garnered increasing attention for their capacity to handle the complex, nonlinear reactions involved in the FCC process. However, on account of the limited operating range of the plants and the black-box nature of data-driven models, relying solely on these models for optimization may lead to contradictory decisions in optimization processes. To address these challenges, we integrate gradient information of product yields with respect to key variables derived from the mechanistic model Petro-SIM, into the training process of data-driven models. To mitigate the high computational demands of the Petro-SIM model, we propose the use of active learning methods for efficient sampling and thereby constructing a surrogate model. The results demonstrate that the active learning approach reduces the required sampling size by 25%. More importantly, the data-driven model trained with gradient information improves the accuracy of trend direction prediction by 34.6%, significantly enhancing its effectiveness in supporting the optimization process. The code will be available at https://github.com/xwl514/fcc-hybrid-loss.