LAUSR

Abstract This paper describes the design of an advanced control algorithm for the cooling system of a large business and commercial centre. This complex system comprises phenomena that are difficult to model with physical principles, such as the demand of the users, heat transport phenomena in a large and complex pipe network, and the behaviour of cooling elements installed by third parties. Motivated by these features, a learning-based model predictive control (MPC) approach is proposed in this paper. The data-driven procedure requires only high-level prior information, making it easy to implement and to replicate on similar systems. Specifically, to derive a dynamic model of the plant, a comparison among AutoRegeressive eXogenous (ARX), Output Error (OE), Echo State Networks (ESN) and Long Short Term Memory (LSTM) neural networks has been performed. The latter have been eventually selected in view of their higher predictive performance on a validation dataset. Then, an output feedback MPC scheme has been designed to cope with the nonlinearity of the model and the presence of boolean control inputs, corresponding to the on/off switching of the cooling units. The resulting MPC algorithm has been tested on a grey-box model of the system, showing significant potential improvements with respect to the baseline controller currently employed.