RNN, GRU, LSTM 모델을 이용한 낙동강 유역 잠재증발산량 예측 및 비교

Abstract

This study aimed to evaluate the applicability of deep learning-based time-series models for predicting reference evapotranspiration (ET₀), benchmarkedagainst the FAO Penman-Monteith (FAO-PM) method. We developed and compared three representative Recurrent Neural Network (RNN)models—basic RNN, Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM)—using daily meteorological data from 21 observationstations across the Nakdong River Basin in South Korea. The results consistently showed that LSTM and GRU models, which incorporate gatemechanisms, significantly outperformed the basic RNN model in prediction accuracy across all stations. The LSTM model demonstrated the best overallperformance, achieving the lowest average Root Mean Square Error (RMSE) of 0.871 mm/day and the highest coefficient of determination (R²) of 0.767during the test period. The GRU model's performance was nearly equivalent to LSTM’s, making it a computationally efficient alternative. While therelative superiority of the models was consistent, the absolute prediction error varied depending on the distinct climatic characteristics of each station. Accuracy was highest at stations with stable wind conditions, whereas errors increased in coastal areas with strong, variable winds. These findingsdemonstrate that LSTM and GRU are robust and reliable data-driven methodologies for accurately predicting ET₀ across diverse climate environments,highlighting their high potential as effective tools for agricultural water resource management.