Data-Driven Fouling Detection in Refinery Preheat Train Heat Exchangers Using Neural Networks and Gradient Boosting

Area of Science:

• Chemical Engineering
• Process Monitoring
• Artificial Intelligence

Background:

• Heat exchanger fouling in refinery crude distillation units (CDUs) significantly impacts energy efficiency and operational reliability.
• Early detection of fouling is critical for preventing performance degradation and economic losses.

Purpose of the Study:

• To develop and compare virtual sensing approaches for real-time fouling detection in CDU preheat trains.
• To evaluate the performance of data-driven (LSTM, XGB) and semi-empirical (ɛ-NTU) models in predicting heat exchanger performance and identifying fouling.

Main Methods:

• Utilized Long Short-Term Memory (LSTM) neural networks and Extreme Gradient Boosting (XGB) for data-driven modeling.
• Employed the ɛ-NTU method as a semi-empirical baseline for comparison.
• Trained models on clean operational data to establish baseline performance and detect deviations indicating fouling.

Main Results:

• LSTM models demonstrated high accuracy in capturing dynamic operational trends for fouling detection.
• XGB models offered a computationally efficient alternative with limitations in extrapolating to novel conditions.
• Both LSTM and XGB models showed superior fouling detection sensitivity compared to the ɛ-NTU method.
• Inefficiencies from a single fouled exchanger were linked to substantial CO2 emissions and economic losses.

Conclusions:

• AI-enabled virtual sensors provide a powerful tool for real-time fouling monitoring in industrial heat exchangers.
• Implementing these tools can enhance predictive maintenance, boost energy efficiency, and reduce environmental impact in refineries.