Deep Learning-based Surrogate Models for Water Distribution Systems

Simulation-based optimization is widely used for designing and managing water distribution networks. The process involves the use of accurate computational models, such as EPANET, which represent the physical processes taking place in the water network and reproduce the control logic governing its operations. Unfortunately, running such models requires expensive computations, which, in turn, may hinder the application of simulation-based optimization to large and complex problems. This issue can be overcome by resorting to surrogate models, that is, simplified data-driven models that accurately mimic the behaviours of physical-based models at a fraction of the computational costs. In this work, we explore the potential of Deep Learning Neural Networks (DLNN) for building surrogate models for water distribution systems. Different DLNN architectures, including feed-forward and recurrent neural networks, are trained and validated on datasets generated through EPANET simulations. The DLNN models are then used in lieu of the original EPANET model to speed-up the evaluation of the objective function employed in a simulation-based optimization problem. The effectiveness of the proposed technique is assessed on a realistic case-study involving cyber-attacks on a water network. In particular, the DLNN surrogate models are employed by an evolutionary optimization algorithm that schedules the operations of hydraulic actuators in order to best respond to the attacks and facilitate the recovery process.