New Generation of Multivariable Predictive Controllers based on Computational Intelligence for the control of industrial processes with complex dynamics (neoPREDiCOM).

Agencia Estatal de Investigación. Ministerio de Ciencia e Innovación. (PID2020-120087GB-C21). 2021-2024

NeoPREDiCOM is a project coordinated between the Universitat Politècnica de València (UPV) and the University of the Basque Country (UPV/EHU). Specifically with the Intelligent Control Research Group (GICI) (https://www.ehu.eus/es/web/gici/).

Main goals:

Design, develop and validate a novel multivariable and multi-objective predictive control architecture based on computational intelligence techniques, capable of controlling in real time industrial processes of complex dynamics, with learning and robust model generation capabilities.

Specific goals:

There are two types of objectives: specific scientific objectives (where their achievement generates more transversal and general milestones) and specific technological objectives (whose achievement implies the application of scientific results to practical or real cases, generating specific milestones for the specific applications proposed).

Specific scientific goals:

[OE-C1] Design and develop new multi-objective optimization algorithms and/or variants of the existing ones to characterize the optimal and sub-optimal non-dominated solutions in their neighborhood. This algorithm will be part of the controller structure and will provide it with robustness when handling sub-optimal solutions (Subproject 1. CPOH – IP1).

[OE-C2] Study and develop single-variable models using ensemble neural networks (Ensemble Model) (Subproject 2. GICI – IP2).

[OE-C3] Design and develop new algorithms and procedures for automated decision making based on linguistic preferences (Subproject 1. CPOH-IP1).

[OE-C4] Develop a new iMO-NMPC proposal for the single-variable case (SISO) incorporating novel contributions in the optimizer and decision manager (Subproject 2. GICI – IP1).

[OE-C5] Design, develop and analyze MIMO models based on neural networks of the processes to be worked with during the project. These models and the resulting analyses will be integrated into the controller structure (Subproject 2. GICI – IP1).

[OE-C6] Adapt and implement the iMO-NMPC control strategy for multivariable systems (MIMO). (Subproject 2. GICI – IP2)

[OE-C7] To design and develop new reinforcement learning strategies based on models capable of simultaneously managing different objectives linked to different rewards that provide improvements over classical methods based on a single reward index (Subproject 1. CPOH – IP1).

[OE-C8] Adapt the multi-objective reinforcement learning strategy for use in the context of the proposed predictive control architecture. The model-based learning strategy will be modified in terms of reward management (cost indices) and model of the environment (process) to find the optimal policy (controller). (Subprojects 1 and 2. CPOH-IP1 and GICI-IP1-IP2).

Specific technological goals:

[OT-1] Validate the proposed control architecture on HiL schemes, where the suitability of both the strategy and the hardware in terms of computational cost can be verified (Subprojects 1 and 2. CPOH-IP1 and GICI-IP1-IP2).

[OT-2] Validate the proposed control architecture on different application cases: management and control of a fuel cell, control of an aluminum rolling mill process (Subprojects 1 and 2. CPOH-IP1 and GICI-IP1-IP2).