Evaluation of data-driven models for post-combustion CO2 capture: A comparative analysis of accuracy, robustness and feasibility

Abstract

This study examines the predictive performance, preprocessing impact, computational feasibility, and robustness of data-driven models in simulating absorber behaviour in carbon capture systems under real-world conditions. Five algorithms — Dynamic Mode Decomposition (simplified and full), Autoregressive Integrated Moving Average, Random Forest, Support Vector Regression, and Long Short-Term Memory networks — were evaluated across three preprocessing scenarios: Robust Principal Component Analysis with and without interpolation, and unprocessed data. Results show that preprocessing generally improves accuracy, with RPCA-based approaches outperforming untreated datasets across most horizons, although its impact on robustness under noise remains limited. Robustness analysis was conducted on the three best-performing models — DMD, ARIMA, and LSTM — revealing distinct behaviours. Dynamic Mode Decomposition was the most computationally efficient, providing near-instantaneous training and prediction, and maintained acceptable performance under noise. ARIMA exhibited strong robustness and predictive capacity, with minimal performance degradation across noise levels. In contrast, Long Short-Term Memory networks, while effective for long-term forecasting, displayed high computational costs and significant sensitivity to stochastic training effects. These limitations resulted in inconsistent performance across noise levels, even under low perturbations. The study highlights trade-offs between accuracy, feasibility, and robustness, stressing the importance of aligning model choice with deployment constraints. While black-box methods offer strong predictions, their sensitivity to randomness and computational demands hinder practical use. Robust and reproducible approaches like DMD balance efficiency and reliability, making them well-suited for industrial carbon capture applications.