A hybirid mechanistic and data-driven (Denoise Autoencoder (DAE) - Long-short term Memory (LSTM)) model for estimating the CO2 concentration profile for the CO2 absorber of the carbon capture plant. The tested dimensionality reduction techniques include: DAE, principal component analysis (PCA) / propoer orthogonal decomposition (POD). The absorber has 6 sampling points while there is only one gas analyzer, this work aims to build a soft sensor to use the other continuous process record to infer the CO2 concentration at these points. The DAE-LSTM approach is then compared to a semisupervised estimation method.
Main documents:
Estimate_CO2_profile.ipynb
semi_supervised_stacked_AE/pilot_plant_ssae.ipynb
Related similar approach:
X. Zhang, et.al, A weighted auto regressive LSTM based approach for chemical processes modeling, Neurocomputing, Volume 367, 2019, Pages 64-74, ISSN 0925-2312, https://doi.org/10.1016/j.neucom.2019.08.006.
The semisupervised stacked autoencoder training strategy is proposed in:
X. Yuan, et al., A novel semi-supervised pre-training strategy for deep networks and its application for quality variable prediction in industrial processes, Chemical Engineering Science, Volume 217, 2020, 115509, ISSN 0009-2509, https://doi.org/10.1016/j.ces.2020.115509.
If you use codes/data in your research, please cite this paper:
@article{ZHUANG2022103747,
title = {A hybrid data-driven and mechanistic model soft sensor for estimating CO2 concentrations for a carbon capture pilot plant},
journal = {Computers in Industry},
volume = {143},
pages = {103747},
year = {2022},
issn = {0166-3615},
doi = {https://doi.org/10.1016/j.compind.2022.103747},
url = {https://www.sciencedirect.com/science/article/pii/S0166361522001440},
author = {Yilin Zhuang and Yixuan Liu and Akhil Ahmed and Zhengang Zhong and Ehecatl A. {del Rio Chanona} and Colin P. Hale and Mehmet Mercangöz},
keywords = {Data-driven modeling, Soft sensor, Long short term memory, Carbon capture pilot plant}
}