Evaluating the impact of different decomposition methods on the accuracy of reference evapotranspiration forecasts in humid regions

Abstract

The forecasting of reference crop evapotranspiration (ET0) plays a crucial role in irrigation scheduling. In this study, we employed deep learning (DL) models, including convolutional neural network (CNN), long short-term memory (LSTM), bidirectional LSTM (Bi-LSTM), CNN-LSTM, and CNN-Bi-LSTM, to forecast ET0 at three different stations (Nanchang, Xinjian, and Dongxiang) in Jiangxi Province, China, for various forecast lead times (1, 3, 5, 7, 10, and 15 days). The experimental results demonstrate that LSTM and Bi-LSTM achieved the highest predictive accuracy across the three stations, with corresponding average values of coefficient of determination (R²), mean absolute error (MAE), root mean square error (RMSE), Nash–Sutcliffe efficiency (NSE), and combined accuracy (CA) of 0.9646, 0.2456 mm day−1, 0.2931 mm day−1, 0.9646, and 0.2043, respectively. Compared with the baseline model, the hybrid models incorporating complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), EEMD, and CEEMD exhibited inferior predictive performance. Additionally, their significantly longer runtime resulted in operational costs that were 8–10 times higher. In contrast, DL-based LSTM and Bi-LSTM are excellent algorithms for forecasting ET0 across various time intervals.