A novel machine learning-based framework for integrating scheduling and control in continuous multiproduct chemical process operations
Abstract
The integration of scheduling and control in chemical process management offers significant efficiencies and economic benefits. However, practical implementation remains challenging due to the complex mathematical formulation. Therefore, pursuing formulations that are both accurate and computationally tractable has consistently remained a central, rapidly evolving research topic.This work presents a machine learning (ML)-based framework for integrated scheduling and control, where deep neural networks (DNNs) are leveraged to address the complexity arising from both the process models and the control layer. The DNNs are trained offline to approximate two key components: (i) the nonlinear open-loop process dynamics (i.e., process model) and (ii) the closed-loop scheduling-adaptive control policy, which establishes a mapping from higher-level scheduling decisions to lower-level optimal control actions. The two DNNs are then embedded into a classical scheduling formulation with an economic objective, resulting in an integrated mixed-integer nonlinear program (MINLP), which accounts for both process dynamics and control logic, while eliminating the need for repeated online evaluation of the complex process model or solving the nonlinear model predictive control (NMPC) problem. The integrated MINLP is solved using a Genetic Algorithm, which handles both discrete and continuous decisions.Application to a multi-product continuous stirred-tank reactor (CSTR) demonstrates the potential of the proposed framework to achieve tractable scheduling–control integration, generating optimal schedules that satisfy varying demand targets and rush-order scenarios. Compared with a baseline case, where product transition times are represented only by statistical estimates, the proposed approach yields an economic gain of 3.65x105 $ in the scheduling objective.
Related
- Efficient Detailed Scheduling of Multiproduct Pipelines using Action-Decoupled Deep Reinforcement Learning
- Integrating Data Center Flexibility into Power System Operations: A Framework for Temporal Redispatch
- Separately continuous maps into regular spaces and continuity of operations in separable semitopological groups
- Integrating YOLOv8-based disaster detection with deep reinforcement learning for closed-loop UAV emergency response scheduling: A simulation study
- Development of approximate scheduling-adaptive controllers for multi-products continuous chemical processes using deep learning techniques and model predictive control
BibTeX
@article{qassime2026novel,
title = {{A novel machine learning-based framework for integrating scheduling and control in continuous multiproduct chemical process operations}},
author = {Mehdi Abou El Qassime and Ahmed Shokry and Antonio Espuña and Eric Moulines},
journal = {Computers \& Chemical Engineering},
year = {2026},
month = sep,
doi = {10.1016/j.compchemeng.2026.109628},
url = {https://hal.science/hal-05755672v1},
note = {CC BY},
}
RIS
TY - JOUR
TI - A novel machine learning-based framework for integrating scheduling and control in continuous multiproduct chemical process operations
AU - Mehdi Abou El Qassime
AU - Ahmed Shokry
AU - Antonio Espuña
AU - Eric Moulines
PY - 2026
DA - 2026/09/19
JO - Computers & Chemical Engineering
DO - 10.1016/j.compchemeng.2026.109628
UR - https://hal.science/hal-05755672v1
AB - The integration of scheduling and control in chemical process management offers significant efficiencies and economic benefits. However, practical implementation remains challenging due to the complex mathematical formulation. Therefore, pursuing formulations that are both accurate and computationally tractable has consistently remained a central, rapidly evolving research topic.This work presents a machine learning (ML)-based framework for integrated scheduling and control, where deep neural networks (DNNs) are leveraged to address the complexity arising from both the process models and the control layer. The DNNs are trained offline to approximate two key components: (i) the nonlinear open-loop process dynamics (i.e., process model) and (ii) the closed-loop scheduling-adaptive control policy, which establishes a mapping from higher-level scheduling decisions to lower-level optimal control actions. The two DNNs are then embedded into a classical scheduling formulation with an economic objective, resulting in an integrated mixed-integer nonlinear program (MINLP), which accounts for both process dynamics and control logic, while eliminating the need for repeated online evaluation of the complex process model or solving the nonlinear model predictive control (NMPC) problem. The integrated MINLP is solved using a Genetic Algorithm, which handles both discrete and continuous decisions.Application to a multi-product continuous stirred-tank reactor (CSTR) demonstrates the potential of the proposed framework to achieve tractable scheduling–control integration, generating optimal schedules that satisfy varying demand targets and rush-order scenarios. Compared with a baseline case, where product transition times are represented only by statistical estimates, the proposed approach yields an economic gain of 3.65x105 $ in the scheduling objective.
SN - 0098-1354
SN - 1873-4375
ER -