MATLAB Codes and Python Notebooks for developing steady-state and dynamic Energy Constrained Neural Network (ECNN) models using noisy transient data
Author: Angan Mukherjee (am0339@mix.wvu.edu)
Last Page Update: Aug 06, 2024
We are very welcome to your contribution. Please feel free to reach out to us with your feedback and suggestions on how to improve the current models.
This public repository contains MATLAB codes and Python notebooks for developing and simulating different steady-state and dynamic ECNNs for nonlinear chemical systems using noisy transient data. The corresponding publication for this work is:
Mukherjee, A. & Bhattacharyya, D. "Development of Steady-State and Dynamic Mass and Energy Constrained Neural Networks for Distributed Chemical Systems Using Noisy Transient Data". Ind. Eng. Chem. Res. (2024)
Available online at: https://pubs.acs.org/doi/10.1021/acs.iecr.4c01429
These codes will be updated in subsequent versions to enhance robustness of the proposed algorithms and user friendliness.
Note that our previous work focused on the development of Mass Constrained Neural Networks (MCNNs) for modeling nonlinear dynamic chemical processes. The respective codes can be accessed publicly on GitHub at: https://github.com/mukherjeeangan26/Steady_Dynamic_MCNN.
The MCNN and ECNN can be combined together for modeling systems involving simultaneous mass and heat transfer, thus leading to the development of Mass-Energy Constrained Neural Network (MECNN) models.
Test case steady-state and dynamic data have been provided in Excel spreadsheets with respect to the operation of a superheater system for developing a lumped parameter model. Note that the codes uploaded in this repository are generic and can be applied to create energy (or mass-energy) constrained models for any steady-state and dynamic systems (data). The training and validation / simulation datasets need to be loaded at the beginning while running the inverse problem / forward problem codes, indexed as '_MainFile.m'. The rows of the input and output data matrices refer to the observation indices and time steps for steady-state and dynamic modeling respectively while the corresponding columns signify the different input / output variables.
For the models developed in this work, MATLAB's default optimization solver for constrained optimization problems (i.e., 'fmincon') seldom led to convergence issues, especially during dynamic optimization. Therefore, it is desired to implement Interior Point with Filter line search algorithm (IPOPT) for estimation of optimal parameters for the equality constrained network models. Therefore, the IPOPT solver has been implemented through the OPTimization Interface (OPTI) Toolbox for both steady-state and dynamic ECNN. The open-sourced OPTI Toolbox can be accessed at the following link: https://github.com/jonathancurrie/OPTI
The user requries to install OPTI in the MATLAB window by running the opti_Install.m and following recommended specifications.
The equality constrained optimization problems have been solved using the Python-based open-source optimization modeling language, Pyomo v6.7.1 through the IDAES framework. Note that using the IDAES framework is not mandatory for running the codes uploaded in this repository. One may simply use the interior point optimization solver (IPOPT) in PYOMO to solve the constrained optimization problems proposed in this work.
More details about Pyomo can be found at: https://github.com/Pyomo/pyomo
More details about IDAES can be found at: https://github.com/IDAES/idaes-pse
One may follow the following installation instructions for downloading Python / Pyomo / IDAES:
- Install Anaconda: https://www.anaconda.com/download
- Run 'Anaconda Prompt'
- Create a new environment by: conda create -n my-new-env
- Activate the new environment by: conda activate my-new-env
- Install IDAES by: pip install idaes-pse
- Get IDAES extensions by: idaes get-extensions
- If a user wants to dodge using IDAES framework, the easiest way to install a version of Ipopt is: conda install -c conda-forge ipopt. More information can be found in PYOMO documentation (available online at: https://pyomo.readthedocs.io/en/stable/contributed_packages/parmest/index.html).
However, for latest updates on IDAES and installation commands, please look up the IDAES GitHub page (https://github.com/IDAES/idaes-pse).
Other libraries may be installed before running the Python notebooks, as necessary.
In addition, to access the training / validation codes for the unconstrained all-nonlinear static-dynamic neural networks used in the dynamic MCNNs, please refer to the following link: https://github.com/mukherjeeangan26/Hybrid_StaticDynamic_NN_Models. Additional details about the unconstrained all-nonlinear network architectures and training algorithms have been discussed in the following publication:
Mukherjee, A. & Bhattacharyya, D. "Hybrid Series/Parallel All-Nonlinear Dynamic-Static Neural Networks: Development, Training, and Application to Chemical Processes". Ind. Eng. Chem. Res. 62, 3221–3237 (2023). DOI: https://pubs.acs.org/doi/full/10.1021/acs.iecr.2c03339