Modeling approaches for data-driven model predictive control of acid gases in waste-to-energy plants

The economic and environmental sustainability of waste-to-energy (WtE) plants can be improved through advanced control techniques such as model predictive control (MPC), which enables stricter regulation by incorporating constraints, handling multiple objectives, and projecting system behavior forward in time. To be effective, MPC requires accurate process models. Due to the complexity and variability of WtE processes, data-driven modeling offers a more practical and flexible alternative to first-principles approaches, making it well-suited for use in data-driven MPC (DDMPC). This study tested several approaches for modeling HCl abatement in WtE plants using routine monitoring data, aiming to identify models suitable for DDMPC. A key part of the process was careful data preprocessing to ensure continuity and reliable results. Various model structures were explored, starting with linear models, then enhancing them with nonlinear transformations (e.g., squared and cubic), and ultimately using fully nonlinear models to capture highly nonlinear behaviors. Among the tested structures, the nonlinear ARX (AutoRegressive with eXogenous input) model with Neural Networks as the nonlinear mapping function for the model output yielded the best overall performance. However, the study showed that a simpler model, defined by fewer parameters, can still effectively capture the HCl removal dynamics in the investigated plant. The linear ARMAX (AutoRegressive Moving Average with eXogenous input) model, with its simpler structure, is sufficient for this purpose and is preferred for DDMPC development due to its lower computational cost and suitability for real-time optimization.