Sie sind hier:



Development of a Neural Network Based NonlinearModel Predictive Control with Application to DYN10 Case Study

Lehrstuhl: DYN

Betreuer: Corina Nentwich, Reinaldo Hernandez, Taher Ebrahim,

Beginn ab: 9th October 2017

Maximale Anzahl der Teilnehmer: 6

Beschreibung: Model predictive control (MPC) is an advanced control technique
which solves iteratively an optimization problem for the optimal control input over a prediction horizon based on a dynamic model of the process. Designing a classical controller for Multi-input multi-output systems (MIMO), despite possible in principle, is very difficult
to realize specially if the process operation is imposing constraints on the inputs or on the states due to safety or other operational considerations. On the other hand MPC based algorithms
handle these kinds of processes in a very natural way. The core part of the mpc scheme is the underlying dynamic model of the controlled process, such that the quality of the control depends on the accuracy of the model. Most of the real processes are nonlinear innature and complicated enough such that developing a detailed first principle models is very time costly and very difficult if not impossible. Black box models and specially Neural Networks
(NN) provide an alternative for first principlemodels. Neural Networks can be used to capture the dynamics of nonlinear and complex, multi-variable systems from the empirical data without a deep perior knowledge of the controlled process.

OBJECTIVE: The objective of the project group is to develop a NN model of the DYN10 case study, which is then used within the NNMPC scheme. The work has to start with capturing the empirical data needed for training and verification from the real process, then deciding on a set of different structures and determine the best structure with respect to the underlying process. After determining the suitable structure, the developed model can be integrated into the NNMPC scheme. The NN model can be implemented using the Matlab NN-toolbox and the NMPC scheme can be implemented also in Matlab or in Python. In the final phase and after testing the scheme by extensive simulations and in order to validate the entire system, the controller can be tested on the real plant. Furthermore the results obtained can be compared with the ones obtained from a previously developed NMPC scheme based on the first principles model.

TASK DISTRIBUTION: The task of this project group can be divided into the following tasks.
• Getting familiar with the plant (DYN10) and with the existing control system.
• Capturing the required data for the training and the validation of the neural network.
• Determination of the suitable structure and the development of the NN.
• Establishing a complete NN based model for the plant and validating it on a new setsof data.
• Implementation of of the NNMPC scheme.
• Application of the developed control scheme on the real plant.
• Compare the resulting control quality to the previous documented NMPC.
• Documentation of the challenge.

REQUIREMENTS: We are looking for motivated and autonomously working team players whose interests and skills lie in the area of modeling and process automation. Since this project is heavily based on programming tasks, basic knowledge of programming languages like Python orMatlab is required. In addition to that experience with Labview is also a big pro. Since we expect the most of the implementation to be done in matlab,we appreciate students who are willing to learn and are dedicated and ready to invest serious time in a complex but interesting project group.