Anomaly Detection of Streamflow Time Series Using LSTM Autoencoder
Streamflow data obtained from the stream-gauge stations usually comprises of an ample volume of outliers. Anomaly detection is a requisite step in streamflow monitoring and analysis, especially in the context of water resources management, planning and flood risk studies. This study suggests a hybrid deep-learning anomaly detection method that combines an autoencoder and a long-short-term memory (LSTM) network. Multiple LSTM cells that collaborate with one another to understand the long-term dependencies of the data in a time series sequence make up the LSTM network. Based on the reconstruction error of the autoencoder's decoding phase, anomaly identification is accomplished. The applicability of the proposed method is demonstrated by considering the streamflow data (from 1985 to 2015) of Thumpamon streamgauge station of Greater Pamba River basin, Kerala. The hybrid framework exhibits promising results after computing the accuracy, precision, recall and the F1-Scores values as 99.51%, 100%, 89.89% and 94.73% respectively.
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.