Dynamic modelling of phase‐change solvent carbon capture process based on AT ‐ BiGRU ‐ MHA

While it is crucial to address the pronounced nonlinearity, substantial time delays, and complex dynamics of post‐combustion carbon capture (PCC) systems using phase‐change solvents, most existing studies rely on idealized simulation data and thus have difficulty representing real industrial noise and long‐range temporal dependencies. To overcome these limitations, this study develops a high‐fidelity dynamic prediction model. Using operational data from the China Huaneng Group phase‐change CO 2 capture platform, we propose an improved AT‐BiGRU‐MHA neural network. The model utilizes a bidirectional gated recurrent unit (BiGRU) to extract temporal features and integrates an adaptive multi‐head attention mechanism—with positional encoding and a learnable temperature coefficient—to emphasize critical historical states. A global–local feature fusion strategy is introduced to enhance responsiveness to recent trends. Experimental results demonstrate high accuracy for 2 key variables. For the CO 2 capture rate, the model achieves an RMSE of 0.0158, a MAPE of 1.0460%, and an of 0.9753. For the stripper bottom temperature, it yields an RMSE of 1.9796, a MAPE of 0.3359%, and an of 0.9846, consistently outperforming baseline models. Attention‐weight visualization enhances the transparency of the model and offers insights into the underlying process dynamics, showing that different heads capture distinct long‐range and short‐term dynamics. This model offers a continuously differentiable foundation for nonlinear model predictive control (NMPC), facilitating CO 2 capture rate tracking and energy management under variable operating conditions.