Author(s): Fernando Salazar; Joaquin Irazabal; Nathalia Silva-Cancino; Juan Mata
Linked Author(s): Fernando Salazar González
Keywords: No Keywords
Abstract: The development of machine learning techniques has promoted their application in various fields of engineering, including hydraulics and hydrology. Models of this type have been shown to be capable of generating highly accurate and useful predictions in several areas. However, many of the applications are limited to cases in which the available information is abundant and of good quality. Less attention has been paid to situations – common in professional practice – in which the available data has issues such as periods without records, potentially erroneous values or heterogeneous reading frequency. These problems are even more relevant when dealing with relatively large databases, when visual exploration of each of the series is unfeasible. We present the results of the application of autoencoder neural networks for the automatic filtering of potentially anomalous values from a dam monitoring database with several hundred variables.
Year: 2024