Effects of nozzle pressures on multi-output predictions of product gas composition and process efficiency from plastic waste pyrolysis oil gasification using machine learning and transformer models

Machine learning models (MLMs) have been used in gasification systems because they can capture nonlinear relationships without requiring explicit reaction equations. Previous studies focused mainly on predicting product gas composition, while process efficiency metrics and nozzle pressure have received less attention. This study evaluated the effect of nozzle pressure on multi-output prediction of product gas composition (H2, CO, CH4, and CO2) and process efficiency metrics (CCE and CGE) for entrained flow gasification of plastic waste pyrolysis oil. Two input configurations were compared: one including O2/fuel ratio, steam/fuel ratio, temperature, and nozzle pressure, and the other excluding nozzle pressure. Nine models from three families were evaluated: traditional MLMs, multilayer perceptron based neural network models (MLPs), and a Feature Tokenizer Transformer (FTT). The results showed that including nozzle pressure improved predictive performance, increasing test R2 by ∼0.08 for product gas components and ∼0.03 for process efficiency metrics. MLMs demonstrated strong predictive accuracy for both gas composition (test R2 = ∼0.977) and process efficiency metrics (test R2 = ∼0.969). While MLMs showed strong predictive capacity overall, the best-performing model varied by target, suggesting model selection must be tailored to specific output variables. SHAP and attention analyses indicated steam/fuel ratio had the strongest influence on process efficiency prediction, whereas O2/fuel ratio dominated product gas composition. These findings indicate that inclusion of nozzle pressure can improve the multi-prediction of gas composition and process efficiency metrics under the tested entrained flow gasification conditions, thereby broadening the applicability of MLMs in gasification systems.