Anomaly Detection in Reservoir Water Level Data Using the LSTM Model Based on Deep Learning

Abstract

Weather and hydrological phenomena have been changing due to climate change as evidenced by localized torrential rainfall and precipitation falling by more than 30% compared to the annual average. From 2013 to 2017 the ninety-nine reservoirs reached a water storage rate of 0%, making a secure stable water supply for agriculture uncertain. There is an increased need for information regarding agricultural water management to respond to the changes in the agricultural environment and climate. Therefore, automatic water level measurement facilities have been introduced to determine the real-time reservoir storage capacity and agricultural water supply. According to the Ministry of Agriculture, Food and Rural Affairs' guidelines for the installation and operation of water level measurement instruments, automatic water level facilities are currently installed at 1,734 reservoirs and 1,880 irrigation canals, with water level data generated at 10-minute intervals. The official recognition of hydrological water level data for agricultural reservoirs increased from six in 2016 to forty-nine in 2019. Anomaly detection algorithm methods for data regarding the agricultural reservoir level as well as quality control measures based on agricultural reservoir characteristics are required to minimize data quality degradation and generate reliable hydrological data over time. Though it was practically impossible to analyze the correlation between the water level or run-off and influential factors such as weather and terrain, recently a non-linear hydrological analysis has been possible using models such as Artificial Neural Networks (ANNs). This study aims to present an anomaly detection algorithm for reservoir level data using deep learning based LSTM (Long Short-Term Memory) models, in combination with other neural networks for managing quantitative information of agricultural water supply.