A two-stage predictive modeling and data augmentation approach for optimizing co-pyrolysis products from oily sludge and solid wastes.

Co-pyrolysis of oily sludge (OS) and solid wastes supports resource recovery and safe residue disposal. However, complex interactions between feedstock composition and operating parameters hinder process prediction and design. In this study, a two-stage predictive modeling and data augmentation approach was developed using literature-derived data covering oilfield, storage-and-transportation, and refinery oily sludge. In the first stage, a multi-task gradient boosting regression (GBR) model predicted oil, gas, and char yields with high accuracy (test R2 > 0.96). For oil yield prediction, the five highest-ranked predictors were OS ash and fixed carbon contents (OS_Ash and OS_FC), solid waste oxygen and ash contents (SW_O and SW_Ash), and the mass fraction of OS (S_ratio). In the second stage, these five variables, together with temperature, were used to predict aromatic hydrocarbon yield. A Wasserstein generative adversarial network with gradient penalty (WGAN-GP) augmentation improved aromatic hydrocarbon prediction R2 by 3.74-5.58%. Constrained particle swarm optimization (PSO) identified a model-predicted candidate operating condition at an OS mass fraction of 56 wt% and a temperature of 546 °C, with predicted oil and aromatic hydrocarbon yields of 83.90 wt% and 14.29 wt%, respectively. The two-stage approach supports data-driven prediction and optimization of oily sludge co-pyrolysis.