Anomaly Detection of Automatic Rain Gauge Measurement Using Artificial Neural Network Long Short Term Memory Method

Abstract

Purpose: The purpose of this research is to accurately detect anomalies in the results of automatic rain gauge measurements using the Long Short Term Memory (LSTM) method, so that measurement errors can be immediately identified and the equipment can be repaired immediately.  

Design/methodology/approach: Detection of anomalies from rain gauge measurements is carried out using quality control (QC) methods based on range and step check, spatial check and error check which produce anomaly labels which are totaled to become Total Anomaly QC. Total Anomaly QC is transformed via one-hot encoding and then the results of the Total QC data transformation are used to build an anomaly detection classification model using the LSTM algorithm.

Findings/result: The model performance was tested with a confusion matrix. LSTM is able to classify data anomalies in the western, eastern and coastal clusters quite well. The accuracy value of these clusters is more than 0.9, so that >90% of the anomalies are classified correctly. The results of this research can improve BMKG's ability to detect rainfall measurement anomalies from automatic rain gauges and assist in maintaining the validity of rainfall data so that equipment maintenance is carried out on time.

Originality/value/state of the art: This research uses different methods and parameters from previous research. The results obtained are quite satisfactory as shown by an accuracy above 0.9.