딥러닝 알고리즘의 성능을 향상하기 위한 최적 모니터링 주기 결정: 남강 유역의 사례를 중심으로

Abstract

The intention of this study is to address the effect of temporal resolution of input data on the prediction accuracy of deep learning algorithms. A popular watershed model Soil & Water Assessment Tool (SWAT) was used to generate time series of discharge and 9 water quality variables over different time intervals (i.e., 1, 2, 3, and 7 days) for three years (i.e., 2015-2017). We then assessed the performance of two deep learning algorithms, recurrent neural network (RNN) and gated recurrent unit (GRU), on those prepared data sets. Our finding showed that the prediction accuracy of the RNN algorithm increased progressively with the time resolution of input data from weekly data to daily data. The RNN also achieved its superior performance on the daily data, regardless of dependent variables. The only exception was the variable carbonaceous biochemical oxygen demand. When a comparison was made between those algorithms, the RNN algorithm was found to have slightly better accuracy than the GRU algorithm. Collectively, all these results revealed that the temporal resolution of input data was strongly responsible for changes in the performance of deep learning, specifically for those used in time series prediction.