Explicit Machine Learning-based Model Predictive Control of Nonlinear Processes via Multi-parametric Programming
- Let us consider a second-order irreversible exothermic reaction A >> B:
-
The first-principles equations (i.e., energy and material balance equations) for this system are given as follows:
-
Where,
- 𝐶𝐴: Concentration of reactant A (kmol/m3)
- 𝑇: Temperature of the reactor (K)
- 𝐶𝐴0: Concentration of reactant A in the feed
- 𝑄 : Heat input rate (kJ/h)
- F: feed volumetric flow rate (m3/h)
- 𝑇0: Inlet temperature (K)
The full description of the process parameters can be found at here
-
The State and Manipulated variables for this system are:
- States variables: 𝐱=[𝐶A−𝐶As, 𝑇−𝑇s]
- Control / Manipulated variables: 𝐮=[𝐶A0−𝐶A0s, 𝑄−𝑄s]
- Step 1: Perform extensive open-loop simulations to obtain sufficient data for model training.
Please refer to the files under the "Open-loop Simulation" folder
- Step 2: Build a ML model to capture the nonlinear system dynamics of the CSTR.
Please refer to the files under the "Train_Model" folder
- Step 3: Approximate the nonlinear behavior of the trained ML model using piecewise linear affine functions.
Please refer to the files under the "Space_discretization" folder
- Step 4: Generate the solution maps for the discretized state-space via multi-parametric programming.
Please refer to the file "GetExplicitML-MPC_Sols.py"
- Step 5: Carry out closed-loop simulations to test the effectiveness of the Explicit ML-MPC.
Please refer to the file "ExplicitML-MPC.py"
You can find our paper here
@article{wang2024explicit, title={Explicit machine learning-based model predictive control of nonlinear processes via multi-parametric programming}, author={Wang, Wenlong and Wang, Yujia and Tian, Yuhe and Wu, Zhe}, journal={Computers & Chemical Engineering}, pages={108689}, year={2024}, publisher={Elsevier} }