Attention-based neural network fusion for fouling prediction in Ethylene-Vinyl Acetate heat exchangers

Fouling, a phenomenon in which materials originating from the process fluid settle onto heat-exchange surfaces, significantly reduces thermal efficiency, increases energy consumption, and raises maintenance costs, particularly in high-pressure tubular reactors used for Ethylene-Vinyl Acetate (EVA) polymerization. Accurate, real-time prediction of fouling factor is therefore essential for optimizing operational efficiency, maintaining product quality, and preventing unplanned downtime. This paper proposes a novel attention-based neural network that integrates parallel Bidirectional Gated Recurrent Unit branches with a Multi-Head Attention mechanism to enhance temporal feature extraction and focus on the most informative time steps. In addition to the neural architecture, the framework incorporates Mutual Information-based feature selection stage to retain highly relevant process variables, derived from temperature, pressure, and flow rate measurements collected through sensors across the reactor system. The model was trained on six years of industrial EVA reactor data from Versalis. Experimental results demonstrate that the proposed model consistently outperforms baseline architectures, achieving the lowest test MSE ( 3 . 48 × 1 0 − 3 ), RMSE ( 4 . 17 × 1 0 − 2 ), and highest R 2 (0.82) on normalized data. These improvements highlight the model ability to capture complex temporal dependencies and generalize under varying operational conditions. The proposed approach offers a scalable and effective solution for predictive fouling monitoring in polymerization heat exchangers, with potential applicability across other energy-intensive chemical manufacturing processes.