Reference evapotranspiration estimation using long short‐term memory network and wavelet‐coupled long short‐term memory network
Xiaoxu Long, Jiandong Wang, Shihong Gong, Guangyong Li, Hui Ju
Abstract
Abstract Evapotranspiration (ET) is a vital component of the hydrological cycle, and accurate estimation of reference evapotranspiration (ET 0 ) is of great importance in agriculture water resources planning and management. In this study, long short‐term memory network (LSTM), artificial neural network (ANN), extreme learning machine (ELM), and their wavelet‐coupled models were used for daily ET 0 estimation. For comparison purposes, this study also investigated the ET 0 estimation capability of three typical empirical models, that is, the Hargreaves–Samani (HS) equation, Penman (PM) equation, and Priestley–Taylor (PT) equation. Daily meteorological data were obtained from two weather stations, Beijing and Baoding, located in the northern part of the North China Plain. Results demonstrated that single machine learning (ML) models, for example, ANN, ELM, LSTM, are steadier and generally have better overall performance than wavelet‐coupled ML models. Besides, the LSTM model performed best among the single ML models and was far superior to the HS, PM, and PT models. The LSTM model helped the root mean square error (RMSE) and mean absolute error (MAE) of the ANN and ELM models decrease by 5%–69% and 8%–72%, respectively. Moreover, the LSTM model helped the RMSE and MAE of the empirical models decrease by 6%–83% and 5%–84%, respectively.