Deep learning-based embedded mixed-integer model predictive control

Abstract

We suggest that using deep learning networks to learn model predictive controllers is a powerful alternative to online optimization, especially when the underlying problems are complex, as in the case of mixed-integer quadratic programs. The use of deep learning has two important advantages compared to classical shallow networks regarding its embedded implementation. A better function approximation can be achieved with the same number of neurons and less weights are necessary due to the use of more, but smaller, layers. This reduces significantly the memory footprint of the necessary code for its embedded implementation. As with shallow networks, deep neural networks are extremely easy to implement and to deploy on embedded platforms. The potential of the approach is illustrated with simulation results of an energy management system in a smart building, including the implementation of the proposed controller using a low-cost microcontroller.