Abstract
Water distribution networks (WDNs) are critical infrastructures that directly
impact urban development and citizens¿ quality of life. Due to digitalization technologies, modern networks have evolved towards cyber-physical systems, allowing real-time
management and monitoring of network components. However, the increasing volume
of data from monitoring poses significant challenges to accurately estimate the hydraulic
status of the system, mainly when anomalous events or unreliable readings occur. This
paper presents a novel methodology for state estimation (SE) in WDNs by integrating convolutional graph networks (GCNs) with long short-term memory (LSTM) networks. The
methodology is validated on two WDNs of different scales and complexities, evaluating
the SE of the sensors. The capability of the GCN-LSTM model was assessed during the
last two months of the time series by simulating failures to analyze its impact on sensor
readings and estimation accuracy. The smaller network showed higher sensitivity of the
sensors to detect failures, while the larger one evidenced more challenges in SE due to the
sensor dispersion. Overall, the model achieved low prediction errors and high coefficient of
determination values between the actual and simulated values, showing good performance.
Likewise, the simulated failures showed that replacing the missing data with the hourly
mean of the last week significantly improved the accuracy of the predictions, guaranteeing
a robust SE in the event of sensor failures. This methodology provides a reliable tool for
addressing various network configurations¿ operational challenges.
