(687h) A Model Predictive Control Framework Utilizing a Hybrid Subspace Based Model
AIChE Annual Meeting
2020
2020 Virtual AIChE Annual Meeting
Computing and Systems Technology Division
Optimization-based Estimation and Control
Wednesday, November 18, 2020 - 9:45am to 10:00am
Subspace based models have demonstrated an excellent ability to model complex and highly
non-linear batch processes. Methods have been developed to accommodate batches of different
lengths in the model building stage. The usage of these models in model predictive control
(MPC) schemes been applied in different application areas [1-2] in the recent past. Even more
recently, a parallel hybrid modeling scheme [3] using a first principles model and a data based
residual model was proposed. The residual model was developed using subspace based technique
and built with error between the process measurement data and the first principles output of
historical batches. The hybrid model was effective in capturing the dynamics of the process and
provided better predictions than subspace (built with process data only) and the first principles
model. In this work, a model predictive scheme utilizing the hybrid model is proposed. The key
consideration in the present formulation is to retain the linearity of the model utilized within the
MPC. However, presence of the non-linear first principles model in the hybrid model renders the
MPC optimization problem into a non-convex optimization problem. In this proposed approach,
a subspace model built with output data of first principles model is first developed and then
appended with the residual model to have the same parallel hybrid model structure. This linear
hybrid model is now used as the predictive tool inside the MPC. A batch crystallization process
is considered as the motivating example. Fines or crystals due to nucleation are produced during
the crystallization process which are undesirable in downstream processing industries. The aim
of the MPC scheme is to reduce the volume of fines at batch termination while maintaining a
desired quality and following certain physical constraints. The MPC embedding the hybrid
model is implemented and the results are compared with purely data driven subspace model
predictive controlle. The hybrid MPC is shown to achieve superior results compared to the other
strategies.
non-linear batch processes. Methods have been developed to accommodate batches of different
lengths in the model building stage. The usage of these models in model predictive control
(MPC) schemes been applied in different application areas [1-2] in the recent past. Even more
recently, a parallel hybrid modeling scheme [3] using a first principles model and a data based
residual model was proposed. The residual model was developed using subspace based technique
and built with error between the process measurement data and the first principles output of
historical batches. The hybrid model was effective in capturing the dynamics of the process and
provided better predictions than subspace (built with process data only) and the first principles
model. In this work, a model predictive scheme utilizing the hybrid model is proposed. The key
consideration in the present formulation is to retain the linearity of the model utilized within the
MPC. However, presence of the non-linear first principles model in the hybrid model renders the
MPC optimization problem into a non-convex optimization problem. In this proposed approach,
a subspace model built with output data of first principles model is first developed and then
appended with the residual model to have the same parallel hybrid model structure. This linear
hybrid model is now used as the predictive tool inside the MPC. A batch crystallization process
is considered as the motivating example. Fines or crystals due to nucleation are produced during
the crystallization process which are undesirable in downstream processing industries. The aim
of the MPC scheme is to reduce the volume of fines at batch termination while maintaining a
desired quality and following certain physical constraints. The MPC embedding the hybrid
model is implemented and the results are compared with purely data driven subspace model
predictive controlle. The hybrid MPC is shown to achieve superior results compared to the other
strategies.
[1] Subspace Identification-Based Modeling and Control of Batch Particulate Processes Abhinav Garg
and Prashant Mhaskar, Industrial & Engineering Chemistry Research 2017 56 (26), 7491-7502, DOI:
10.1021/acs.iecr.7b00682
[2] Model predictive control of uni-axial rotational molding process, Abhinav Garg, Felipe P.C. Gomes,
Prashant Mhaskar, Michael R. Thompson, Computers & Chemical Engineering, Volume 121, 2019,
Pages 306-316, ISSN 0098-1354
[3] Hybrid Modeling Approach Integrating First-Principles Models with Subspace Identification
Debanjan Ghosh, Emma Hermonat, Prashant Mhaskar, Spencer Snowling, and Rajeev Goel, Industrial &
Engineering Chemistry Research 2019 58 (30), 13533-13543