Physics-informed dynamic hybrid modeling for real-time renewable CO₂ tracking in refinery co-processing.

Area of Science:

• Chemical Engineering
• Environmental Science
• Data Science

Background:

• Refinery co-processing is crucial for integrating renewable bio-feedstocks.
• Accurate attribution of CO2 emissions from these feedstocks is challenging due to process complexity.
• Existing methods fail to address dynamic feedstock variability and nonlinearities.

Purpose of the Study:

• To develop an adaptive modeling framework for real-time CO2 emission attribution in refinery co-processing.
• To improve the accuracy and interpretability of renewable CO2 emission estimations.
• To provide a practical solution for transparent refinery decarbonization.

Main Methods:

• A hybrid adaptive modeling framework combining constrained Recursive Least Squares (RLS) and sparse Generalized Additive Models (GAM).
• Integration of conformal prediction for robust uncertainty quantification.
• Application of non-negativity constraints for physical interpretability.

Main Results:

• The proposed framework achieved high predictive accuracy (RMSE: 345.4, R2: 0.950) on an industrial dataset (>86,000 samples).
• Outperformed 15 baseline methods in estimating CO2 emissions.
• Demonstrated real-time, interpretable, and uncertainty-aware tracking of renewable feedstock contributions.

Conclusions:

• The hybrid adaptive model offers a superior solution for CO2 emission attribution in refinery co-processing.
• This approach facilitates transparent and data-driven decarbonization strategies.
• Enables accurate monitoring of renewable bio-feedstock impact on refinery emissions.