MPC-RNN-Controller

This repository contains code for implementing a Recurrent Neural Network (RNN) Model Predictive Controller (MPC) for a 4-tank coupled system, as part of My Master's degree thesis project.

Project Overview

The RNN MPC Controller project aims to develop a predictive control system using a recurrent neural network to predict control inputs required by a linear MPC for a 4-tank coupled system. The project leverages data extracted from a simulated plant implemented in Simulink to train the RNN.

Project Details

• Formula: The control inputs are predicted using the formula U(k+1) = f(U(k-5), ..., U(k), e(k-5), ..., e(k)), where U(k) represents control inputs and e(k) represents error signals at time k.

• Neural Network Architecture: The RNN architecture consists of a single hidden layer with 10 neurons.

• Training Parameters:

  • Learning Rate: 0.001
  • Loss Function: Mean Square Error (MSE)
  • Optimizer: Stochastic Gradient Descent (SGD)
  • Number of Iterations: 3000

Reference

This project is inspired by the paper "Recurrent Neural Network Based Predictive Control Applied to 4 Coupled-tank System" by Elmer Calle and José Oliden, presented at the 2021 IEEE International Conference on Automation/XXIV Congress of the Chilean Association of Automatic Control (ICA-ACCA). Link to Paper

License

This project is licensed under the MIT License. See the LICENSE file for details.