Advanced control techniques for CHP-DH systems: A critical comparison of Model Predictive Control and Reinforcement Learning

District heating (DH) network is a key infrastructure to decarbonize the heating sector through the centralized production of heat distributed to final users. The implementation of advanced control techniques is increasingly common in the field of energy optimization since they can provide a more efficient way of minimizing energy demand by appropriate scheduling of the control variables. The aim of this work is to present the application of two control strategies, i.e., Model Predictive Control (MPC) and Reinforcement Learning (RL), to a system based on a DH network supplied by a Combined Heat and Power plant (CHP-DH plant). The analyzed case study is a real CHP-DH plant operating in the small Italian town of Osimo (central Italy). The DH network currently connects more than 1200 users, generating peak heat demand of about 9.7 MWth. The heat generator is composed of a natural gas fueled internal combustion engine coupled with natural gas boilers. The work provides a comparison between the current control strategy (deduced from measured data) and the performance of the CHP-DH plant controlled with an MPC and an RL control. The results showed the effec-tiveness of the two controls in satisfying the thermal demand of the users, while minimizing the thermal losses towards the ground. Both MPC and RL allow to implement control strategies different from the current control in terms of supply temperature and flow rate circulating in the network. Referring to the winter months, in which the current operation of the system tends to prefer high supply temperatures, the advanced controls made it possible to reduce the thermal heat supply by reducing the thermal losses of about 3.9 % with the MPC and 6.54 % with the RL, corresponding to emission avoidances up to 23.3 tCO2 and 12.6 tCO2, respectively. The paper, as well as showing the application of the controls, contains a critical discussion of all the positive aspects and weaknesses found in the application of the MPC and the RL control to the case study.