Use this link to cite:
http://hdl.handle.net/2183/31379 Control de un laboratorio de control de temperatura mediante redes neuronales recurrentes
Loading...
Identifiers
Publication date
Authors
Blanco Fernández, Cristian
Sierra García, Jesús Enrique
Santos, Matilde
Advisors
Other responsabilities
Journal Title
Bibliographic citation
Blanco Fernández, C., Sierra-García, J.E., Santos, M. (2022) Control de un laboratorio de control de temperatura mediante redes neuronales recurrentes. XLIII Jornadas de Automática: libro de actas, pp.193-200 https://doi.org/10.17979/spudc.9788497498418.0193
Type of academic work
Academic degree
Abstract
[Resumen] El control predictivo (MPC – Model Predictive Control) de procesos es una estrategia extendida, que se basa en la resolución de un problema de optimización en tiempo real, lo que puede ser computacionalmente muy costoso en función de la naturaleza del problema en cuestión. Para superar esta limitación, se ha investigado la posibilidad de utilizar redes neuronales entrenadas para sustituir a este tipo de controladores. La idea subyacente es que para problemas que muestran un comportamiento predecible, se puede entrenar una red a partir de un controlador optimizado para que pueda sustituirlo. De esta forma los costes computacionales se trasladan al entrenamiento de la red, permitiendo el control en tiempo real sin necesidad de realizar operaciones computacionalmente complejas. Este trabajo explora esta idea a partir de un laboratorio de control de temperatura. Se entrenan dos tipos de redes neuronales recurrentes, y se compara su funcionamiento con el de un controlador tradicional.
[Abstract] Model Predictive Control (MPC) is an extended control strategy based on the resolution of an optimization problem in real time, which can be a computationally expensive process depending on the nature of the problem. To overcome this limitation, the use of neural networks already trained as a substitute for this type of controllers has been investigated. The underlying concept is that, for a sufficiently predictable system, a neural network can be trained, using data from an optimized controller, which can replace the MPC. With this approach the higher computational cost lies in the training of the network instead of the online operation of the system. This idea is explored using a temperature control lab. Two types of recurring neural networks are trained, and the performance and computational cost are compared with a conventional controller.
[Abstract] Model Predictive Control (MPC) is an extended control strategy based on the resolution of an optimization problem in real time, which can be a computationally expensive process depending on the nature of the problem. To overcome this limitation, the use of neural networks already trained as a substitute for this type of controllers has been investigated. The underlying concept is that, for a sufficiently predictable system, a neural network can be trained, using data from an optimized controller, which can replace the MPC. With this approach the higher computational cost lies in the training of the network instead of the online operation of the system. This idea is explored using a temperature control lab. Two types of recurring neural networks are trained, and the performance and computational cost are compared with a conventional controller.
Description
Editor version
Rights
Atribución-NoComercial-CompartirIgual 4.0 Internacional (CC BY-NC-SA 4.0)
https://creativecommons.org/licenses/by-nc-sa/4.0/deed.es


