Machine Learning Assisted Solutions of Mixed Integer MPC on Embedded Platforms
Yannik Löhr, Martin Klaučo, Miroslav Fikar, Martin Mönnigmann
Abstract
Many control applications, especially in the field of energy systems, require a simultaneous decision for continuous and binary values of control inputs. In optimal control methods like model predictive control (MPC), this leads to the problem of solving expensive mixed-integer programs online. As this solution in practice has to be calculated with low cost embedded hardware with low energy demand, it is necessary to reduce the computational demand in advance. We present an approach to replacing the mixed-integer program by a simpler quadratic program by means of learning techniques. To be more specific, we design a neural network and a support vector machine to classify the optimal control policies for the binary inputs offline and evaluate this decision in the online step as the basis for the solution of the quadratic program. As a result, we achieve a controller suitable for implementation on embedded hardware. We demonstrate its applicability to a domestic heating system. The results indicate a very high quality of the approximation of the primary optimal controller that solves mixed-integer programs online.