Optimal real-time control of water distribution systems undergoing cyber-physical attacks

In recent years, researchers have developed intrusion detection tools to improve the security of Water Systems in the presence of novel and sophisticated cyber-attacks. Nevertheless, such tools lack prescriptive analytics conceived to determine the best course—that is, how to control a system undergoing an attack. Here, we fill in this gap by presenting a numerical modelling framework based on the coupling of DHALSIM with a Deep Q-Learning agent.  The former is a simulation environment providing a high-fidelity representation of both hydraulic processes and ICT devices, while the latter is an optimal control algorithm tasked with the problem of curbing the impact of cyber-physical attacks by re-operating key hydraulic devices, such as pumps and valves. Their integration is made feasible thanks to a new feature implemented in DHALSIM—the stepwise simulation—allowed by the addition of a new simulator wrapper, namely Epynet. The evaluation phase is run considering the system in normal operating conditions and undergoing cyber-attacks. Our results show a good behaviour in terms of Demand Satisfaction Ratio. In particular, the control agent manages to satisfy the customers demand, without overflowing the tanks in the system.  To our knowledge, these results are the first in the unexplored area of attack recovery in water distribution systems.