An iTransformer-based Approach for BTX Yield Forecasting in the Coke Oven Gas Purification Process

Abstract

The complexity of the BTX recovery process makes it challenging to accurately predict its yield. The nonlinear and intricate relationships between process variables (e.g., temperature, pressure, flow rate) and the BTX yield are not adequately captured by traditional modeling approaches. To overcome this limitation, this study applies the iTransformer model to predict the BTX yield. The iTransformer is a state-of-the-art machine learning architecture capable of effectively learning long-term dependencies and inter-variable interactions in multivariate time series data, making it well-suited for modeling the dynamic characteristics of the COG purification process. We propose a novel architecture combining GAT-Linformer with iTransformer as the backbone. This approach aims to maximize prediction accuracy by deeply learning complex latent patterns and interac- tions among variables within the data, independent of physical simulations. Experimental results using real-time process data demonstrate the superior performance of the proposed model compared to existing baselines.