This work deals with the use of a new class of Artificial Neural Networks called Auto-Regressive Neural Networks (ARNN) for the emulation of sewer overflows. Sewer overflows are discharges of untreated human, industrial and commercial wastes and sometimes storm runoff water directly into streams, rivers, or lakes. In the Netherlands, approximately 90 % of all sewer systems are combined systems, which means that the environmental regulations applied to such missions are becoming stricter year after year. In fact, the current legislation requires simulations of flow in sewer systems over a 10-year period, with an extension to 25 years in the near future. This implies that a few hundred storms events have to be simulated for each proposed new sewer design or for the improvement of an existing one.
The common technique consists of simulating sewer overflows by hydrodynamic, computer-based models that spend long computational times. The approach tested in this work is the use of ARNN, a type of Artificial Neural Networks, where the computed output in one or more previous time step(s) is used as input for the current time step. One of the main advantages of an ARNN model is that the information regarding the physics of the system is kept in every time step. Therefore, the model only requires a limited quantity of information about the state of the system for the current time step.
The selected sewer system is located at the city of Maartensdijk, in the centre of the Netherlands. In order to apply the new approach, it is necessary to have target values set for the ARNN training. For this purpose, the SOBEK-Urban package developed at WL/DELFT HYDRAULICS is employed. SOBEK must provide the ARNN emulator with not only discharges but also water depths at one overflow weir in the system. In fact, water depth acts as a link between rainfall depths and sewer discharges. Rainfall depths in this case are the inputs of the ARNN model, meanwhile the outputs generated by the emulator are the discharge from the weir and the water depth upstream of the weir. The water depth is then fed back as input in the next time step forming the autoregressive part of the model.
This research also presents the way to select a few representative rainfall events out of few hundred events that can lead to overflows in a sewer system. Finally, the efficiency of the approach is tested by emulating two overflow points simultaneously.
Back to the list of MSc abstracts