Deep Neural Network architecture as predictive optimal controller for {HVAC+Solar cell + battery} disturbance afflicted system
Abstract — Model predictive control is an advanced method to control the dynamics of a system while satisfying a certain set of constraints. The method itself is gaining more and more popularity in all sorts of industries ranging from chemical plants and oil refineries where they have been used since the 80s, to power plants, and more generally power systems. Deep neural networks are neural networks with more than one layer. They are considered to be universal function approximators, which would make them perfectly suited for data driven control applications. Although they have been around since since the 70’s, lack of data has made it impossible in practice to showcase their real abilities until the last two decades where massive automation and connectivity have made it possible to generate a massive amount of data. In this paper we show how it is possible to simulate an mpc optimized control offline on an HVAC system connected to a grid, a battery and a solar panel (the system is introduced in the introduction) and train a deep neural network (6 layers, 50 units each layer) in order to achieve MPC without explicit control computation or instruction, but rather exclusively the data generated by the offline simulation. Index Terms — Model predictive control, Optimal control, Deep Neural Networks, Optimization, Deep Learning, HVAC, Power Electronics, Power Systems, Power Control, fuzzy logic.