This work aims to model a demethanizer column of a CRR plant using Physics-Informed Neural Networks (PINNs).
The plant simulation, from which the datasets were obtained, was developed by Marta Mandis in Aspen HYSYS and employed in the article "Exploring nontraditional LSTM architectures for modeling demethanizer column operations" (DOI: https://doi.org/10.1016/j.compchemeng.2024.108591).
- Folder HYSYS: dataset of the CRR plant simulation in Aspen HYSYS
- dataset_completo and datasetFEED_completo: data extraction and preprocessing for MATLAB
- Function normalizeMinMax: function for scaling the data within the range [-1, 1]
- Function denormalizeMinMax: function for rescaling the data to its original range
- Entire_column: PINN model for a less detailed representation of the entire column
- stage1_p3_norm: PINN model for a detailed representation of Stage 1 (Top tray), where reflux is introduced
- feed2_norm: PINN model for a detailed representation of Stage 2 (Feed tray)
- stripping_norm: PINN model for a detailed representation of Stage 3 (Stripping tray)
- reb_norm: PINN model for a detailed representation of the reboiler, treated as a theoretical stage of the column. It is fed by the liquid descending from Stage 30 and produces both the bottom NGL liquid and the vapor stream entering Stage 30, which subsequently rises through the entire column.
The network was trained by enforcing material balance constraints as the sole physical prior.