Author(s): Enric Bonet; Maria Teresa Yubero; Lluis Sanmiquel; Marc Bascompta
Linked Author(s):
Keywords: No Keywords
Abstract: Historically, one of the most common causes of dam failure has been overtopping, primarily in earthfill dams. Such events accounted for approximately 34% of the events registered in the United States, according to the Association of State Dam Safety Officials according to ICOLD (1995). Other causes which have also contributed to dam failures throughout history, notably water infiltration through the dam body in masonry dams, leading to erosion of the mortar that binds the rocks forming the dam body according to Lane (1935) and Yao & Liu (2021). Accurately quantifying the flow rate from cracks in the mortar is, therefore, an important parameter to consider during dam maintenance and operation. In this article, a tool is developed using Artificial Intelligence, specifically artificial neural networks, for predicting water leakages in a masonry dam. The tool learned from historical data collected from the Santa Fe del Montseny Dam (Barcelona) over the past 10 years (with gaps in the dataset during this period), taking into account factors such as temperature, precipitation, reservoir levels, water levels in observation wells within the dam, as well as water leakage measurements. The leakage flow prediction tool was developed in a MATLAB environment. An artificial neural network was used, with different options such as a hold-out sample and k-fold validation being implemented. In this study, different layer sizes, different number of neurons, and different k-fold values were considered in an attempt to minimize the leakage prediction error of the tool. The results indicate that the tool can predict infiltration flow with an error close to 12% (using a hold-out sample, or test dataset), making it a valuable tool for decision-making pertaining to the dam’s security. Indeed, the capacity to accurately predict leakage flows is a useful tool for dam monitoring and identify deviations from historical patterns that should be further investigated and/or acted upon.
Year: 2024