Assessing the Performance of LGBM on Non-stationary Molecular Sieve CO2 Dehydration Data: A Data Pruning Technique

The effectiveness of Light Gradient Boosting Machine (LGBM) remains questionable in its application to time series data, particularly in forecasting CO2 dehydration systems using molecular sieves. Therefore, this research seeks to improve accuracy in predicting natural gas adsorption processes by analyzing non-stationary data with one-week training window configurations. It was noted that different features obtained different RMSE and MAE values for different windows which varied from 0.09 to 25.54 s of computational time for one-week windows. Moreover, the one-week training period provides an optimal balance system performance and generalization along with operational practicality which makes it well-suited for dynamic systems such as molecular sieve setups. Proper preprocessing combined with LGBM provide a strong basis for gas processing monitoring and optimization using a data-driven approach. Reliability and transparency of predictions can be improved through research on adaptive windowing techniques and hybrid forecasting models.