K. Nithyanandam, S. Bhavsar, C. Kulkarni
Impact Innovations LLC,
United States
Keywords: carbon capture, emissions, CESAR-1, time-series forecasting, causal inference
Summary:
Carbon capture technologies are essential for reducing carbon emissions and other pollutants from industrial and electric power generation facilities. Solvent-based systems, particularly those using amines, represent one of the most promising solutions for capturing carbon dioxide (CO2) from flue gas before it escapes into the atmosphere. However, these amine solvents are volatile and can degrade over time, leading to the emission of unwanted byproducts, especially when exposed to high temperatures and oxidative environment. These challenges are further amplified as fossil fuel plants adjust to the growing integration of variable renewable energy sources, which require them to operate under dynamic and cyclic conditions. As a result, managing emissions from carbon capture systems has become increasingly complex, making it more difficult to comply with environmental regulations. Currently, there are no robust modeling frameworks available to predict and manage emissions from carbon capture plants, especially under variable operating conditions. Traditional steady-state models are limited in scope and cannot capture the transient behavior or complex interactions that occur within these systems. Moreover, most process models are designed primarily for performance optimization and typically do not account for emissions, in part due to limited knowledge of the underlying chemical pathways and reaction kinetics. To address this gap, this project developed a suite of advanced machine learning (ML) tools, augmented by physics and causal reasoning, to forecast and control emissions in real time under dynamic operating conditions. In this project, we built a stand-alone module designed to integrate within a broader modular software framework incorporating: (1) an automated code testing and version control pipeline to ensure code quality and reliability; (2) machine learning models trained on operational data to forecast emission profile of amines and key degradation byproducts such as ammonia and aldehydes; (3) physics-informed feature engineering to improve model transferability across different systems; and (4) graph-based causal analysis to identify and evaluate the sensitivity of key process variables in reducing emissions. The final forecasting model tested on a nine-day dataset with five-minute sampling intervals achieved high accuracy, with less than 10 percent average prediction error for single horizon forecasts and less than 15 percent average prediction error for multi-horizon forecasts up to 25 minutes. This work lays the foundation for a real-time digital twin that will enable proactive emissions control, optimize solvent use, and minimize both regulatory and operational risks associated with unexpected emission events. More broadly, this approach could benefit other emissions-intensive technologies by offering a scalable, data-driven alternative to costly and time-consuming first-principles modeling.