Neural Network Based Approximated Model Predictive Control with Preview on Standard Automation Hardware

The concept of Model Predictive Control (MPC) has proven itself for demanding control tasks. One reason for this is the ability to consider manipulated variable and state constraints directly during the design phase. In addition, known future setpoints and disturbance variables can be used to calculate the optimal manipulated variable. This is known as a preview and can significantly improve the control behaviour. One problem with implicit MPC is the need to solve an optimization problem in each sampling step and the resulting computational requirements. A common solution is to formulate an explicit MPC and store the optimal control laws for each region. However, if preview steps are to be integrated, the number of required regions increases drastically. Neural networks as universal function approximators are also suitable for imitating an MPC. A Multi-Layer-Perceptron (MLP) is very suitable for processing high-dimensional data and can therefore also model the preview capability. The MLP is trained entirely using simulatively generated training data. The functionality and drastically reduced computing requirements are demonstrated using an experimental test setup of a seesaw system. Here, the algorithms are executed on standard automation hardware, i.e. on an industrial PC (IPC).