A Digital Twin of a Sewage Water System Using Neural Networks

EasyChair Preprint 10983

Abstract

Under a climate change scenario, and with environmental pressures and regulations in the European Union, the correct management in a quasi-real-time and the forecast under different scenarios of the sewage water system (SWS) play a crucial role; they significantly impact urban floods and water quality treatments.  Under this scenario, we developed a Data-Driven Digital Twin (DT) using different Neuronal Networks (NN) for a small SWS basin in northern Italy. The basin under study consists of a 140 km long sewage network and a total of 22 Doppler sensors that measure every six minutes the water velocity, water pressure (depth) and water temperature and three rain gauges that measure every minute for a total of 1140 days of register. Due to the conditions in which the sensors work, it is typical to have low-quality measures. For this reason, we developed, trained, and included in the DT an NN capable of detecting any anomaly value, assigning a possible cause to the problem (i.e., dirty sensor), and suggesting a correct potential value; this model shows an accuracy >90%. After this quality control, the data passes into the main DT.  This research evaluates two approaches: a convolutional layer NN and a Graph NN. Both models mimic the configuration of the SWS and use the same data to be trained. The model was evaluated under scenarios of missing data (i.e., sensor removal). Both models show a general accuracy of>90%. Finally, this project shows DT's successful development and application due to industry, government and academic collaboration.

Keyphrases: Digital Twin, Graph Neuronal Networks, Neuronal Networks, sewage water

@booklet{EasyChair:10983,
  author    = {Riccardo de Socio and Chelsea Owen and Ege Turkyener and Emre Davut and Sergio Lopez Dubon},
  title     = {A Digital Twin of a Sewage Water System Using Neural Networks},
  howpublished = {EasyChair Preprint 10983},
  year      = {EasyChair, 2023}}