Traditionally, fouling is managed through design allowances and periodic shutdowns for cleaning, but real-time control systems typically do not compensate for fouling, leading to suboptimal operation. Development of the accurate and robust modelling has been challenging due to the high variability of the processes and the complex and time-varying nature of the fouling phenomena. Various approaches have been proposed and studied to detect and monitor fouling in reboilers. For example, empirical correlations have shown to model fouling resistance or film coefficients, dimensional analysis has been used to predict clean-surface performance, and combined thermal-hydraulic indicators have proven to signal fouling onset [2, 3, 4]. However, these methods focus on monitoring and detection rather than actively mitigating fouling effects, and they often require extensive data or fail to isolate fouling dynamics completely. Previous studies using pure ML methods (e.g., Artificial Neural Networks and Gaussian Process Regression) have improved fouling prediction accuracy over empirical models [5], but often suffered from limited interpretability and robustness under variable conditions [6, 7, 8]. These limitations also include difficulties in generalizing beyond training conditions and dependency on extensive, high-quality data, which is often scarce in industrial contexts. As a result, such purely data-driven models frequently lack robustness, limiting their practical implementation in industrial process environments.
To address these limitations, we propose a hybrid modeling framework that integrates physics-based kinetic models with machine learning to predict and capture the dynamic behavior of fouling under varying operational conditions. In the proposed integrated model, fouling kinetics are characterized using the Ebert–Panchal framework, which predicts deposit accumulation by relating key variables such as fluid temperature, flow rate, foulant concentration, and thermal gradients under the assumption of steady-state, uniform deposition with first-order kinetics. However, while kinetic models like Ebert–Panchal provide a solid theoretical foundation, they often fall short in capturing complex, time-varying behaviors and latent features inherent in real-world operations. These models typically assume steady conditions and uniform deposition, overlooking transient phenomena, process variability, and other nuanced factors that can significantly influence fouling evolution. To overcome these shortcomings, a neural network is trained on historical plant data—including temperature profiles, flow rates, heat transfer coefficients, and reboiler duty variations—to capture the dynamic behavior of fouling. This hybrid approach merges the rigor of physics-based kinetics with the adaptability of machine learning, ensuring that predictions remain grounded in fundamental mechanisms while capturing transient features that traditional models miss. which is subsequently integrated into a model predictive control (MPC) scheme. The MPC leverages real-time predictions of fouling evolution to adjust process variables — actively optimizing reboiler operation (e.g., reboiler duty or reflux ratio) — to prevent excessive fouling while ensuring that separation performance and product quality remain optimal. By continuously balancing energy efficiency with fouling mitigation, the controller maximizes uptime and prolongs reboiler runtime, thereby reducing unplanned shutdowns. This integrated framework bridges the gap between predictive maintenance and real-time process optimization, offering a practical, interpretable solution for industrial fouling control.
In conclusion, our integrated framework combines fouling kinetics from the Ebert–Panchal model with data-driven predictions and model predictive control, enabling real-time optimization of key parameters such as reboiler duty and reflux ratio. This integrated approach mitigates fouling, enhances energy efficiency, and ensures consistent process performance, significantly reducing downtime and operational costs.
References: