Prediction of hydrocarbon adsorption–desorption dynamics in activated carbon columns using long short-term memory networks for intelligent removal systems

Hydrocarbon (HC) emissions from the petrochemical industry pose environmental and health risks, as HCs are precursors for the generation of fine particulate matter and have carcinogenic effects. Meeting increasingly stringent environmental regulations requires accurate prediction of adsorption–desorption dynamics and effective optimization of industrial emission control systems. This study develops and evaluates long short-term memory (LSTM) models to predict HC adsorption–desorption behavior in activated carbon columns. Univariate and multivariate LSTM models were compared to assess how HC concentration and flow rate influence system dynamics. The models were trained and validated with experimental data collected at varying inlet concentrations and tested on an independent dataset to evaluate generalizability. The multivariate LSTM model achieved superior performance for system-wide dynamics, reaching an R² of 0.9336 for HC concentration during desorption. Both models predicted flow rate with high accuracy; however, the univariate model more effectively predicted HC concentration during adsorption (R² = 0.7451), underscoring its suitability for single-parameter prediction. Confusion matrix analysis further demonstrated the robust cycle identification capabilities of both approaches (accuracy >99 %), with lower misclassification rates for the multivariate model. Implementation of these models in a proposed parallel adsorption–desorption system confirmed their practical value for real-time process optimization. The LSTM framework enables dynamic adjustment of column transitions and cycle timing, while facilitating early detection of performance degradation. This approach to process modeling and control proposed in the present study thus contributes to more intelligent, efficient, and sustainable industrial emission management practices based on adaptable model selection tailored to specific applications and computational constraints.