River Flooding Forecasting and Anomaly Detection Based on Deep Learning

Abstract

Pluvial floods are rare and dangerous disasters that have a small duration but a destructive impact in most countries. In recent years, the deep learning model has played a significant role in operational flood management areas such as flood forecasting and flood warnings. This paper employed a deep learning-based model to predict the water level flood phenomenon of a river in Taiwan. We combine the advantages of the CNN model and the GRU model and connect the output of the CNN model to the input of the GRU model, called Conv-GRU neural network, and our experiments showed that the Conv-GRU neural network could extract complex features of the river water level. We compared the predictions of several neural network architectures commonly used today. The experimental results indicated that the Conv-GRU model outperformed the other state-of-the-art approaches. We used the Conv-GRU model for anomaly/fault detection in a time series using open data. The efficacy of this approach was demonstrated on 27 river water level station datasets. Data from Typhoon Soudelor in 2015 were investigated by our model using the anomaly detection method. The experimental results showed our proposed method could detect abnormal water levels effectively.

Keywords

• Floods
• river water level
• forecasting
• deep learning
• anomaly detection