Abstract:

Accurately estimating evapotranspiration using deep learning models under limited meteorological parameter conditions is of great significance for the efficient utilization and management of limited water resources in arid regions. However， current deep-learning-based hybrid model CNN-BiLSTM-Attention neglect the parameter optimization， making it difficult to meet the requirements of practical application. In this study， a novel hybrid model RIME-CNN-BiLSTM-Attention was proposed to achieve accurate prediction of daily reference evapotranspiration （ET₀） in Linze County under limited meteorological parameter conditions by using the RIME Optimization Algorithm to optimize the hyper-parameter of CNN-BiLSTM-Attention. Compared with CNN-BiLSTM-Attention， the mean absolute percentage error （MAPE） of RIME-CNN-BiLSTM-Attention decreases from 14. 56% to 14. 09%， and the coefficient of determination increases from 0. 8654 to 0. 8930； Furthermore， the numerical results show that the hybrid model RIME-CNN-BiLSTM-Attention outperformed than that of HHO-CNN-BiLSTM-Attention， OOA-CNN-BiLSTM-Attention， and NGO-CNN-BiLSTM-Attention， where CNN-BiLSTM-Attention was separately optimized by the Harris Hawks Optimization （HHO）， Osprey Optimization Algorithm （OOA）， and Northern Goshawk Optimization （NGO）， suggesting that the proposed hybrid model RIME-CNN-BiLSTM-Attention has more robust model performance and higher calculation accuracy， and can achieve accurate estimation of ET₀ in the study area.

Key words: reference evapotranspiration, RIME optimization algorithm, convolutional neural network, bidirectional long short-term memory network, attention mechanism