Using weather radar to improve the prediction accuracy of LSTM neural networks for anomaly detection of water level measurements in UDS

P. Aarestrup*, P. S. Mikkelsen

*Corresponding author for this work

Research output: Contribution to conferenceConference abstract for conferenceResearchpeer-review

32 Downloads (Pure)

Abstract

Large amounts of data are currently collected in urban drainage systems in the process of digitalizing the water sector. The data is to be used in decision-making when moving towards a more data-driven urban water management (Eggimann et al., 2017). The large amount of data also poses a threat if wrong, leading to potentially wrong decisions, hence anomalies in the data need to be identified and the process automated for it to be scalable. In order to have a well-functioning anomaly detection, models with a high prediction accuracy are needed to avoid wrong classifications of the data.

Feed-Forward Neural Networks (FFNN) have been used before to predict time series in urban drainage systems (Savic et al., 2013; Karimi et al., 2019). Recent studies have shown that Long Short-Term Memory (LSTM) networks, which is a sub-type of Recurrent Neural Networks (RNNs) can learn time-dependencies through gating mechanisms and be used to predict water levels and flows using rain gauges as a predictor (Palmitessa et al., 2021). However, the models are still prone to the spatial distribution of precipitation when using rain gauges or cannot be used if rain gauges are not present. To solve this, we here proposed to use weather radars as a predictor. This is tested and compared to using rain gauges when predicting water levels in a sewer overflow chamber in Bellinge, Denmark using data from a 9 month period in 2019.
Original languageEnglish
Publication date2022
Number of pages3
Publication statusPublished - 2022
Event12th Urban Drainage Modeling conference - Hybrid event, Costa Mesa, United States
Duration: 10 Jan 202212 Jan 2022

Conference

Conference12th Urban Drainage Modeling conference
LocationHybrid event
Country/TerritoryUnited States
CityCosta Mesa
Period10/01/202212/01/2022

Fingerprint

Dive into the research topics of 'Using weather radar to improve the prediction accuracy of LSTM neural networks for anomaly detection of water level measurements in UDS'. Together they form a unique fingerprint.

Cite this