A novel method for estimating daily evapotranspiration based on one, two, and three-day meteorological records using the long short-term memory model combined with feedforward neural networks.

In agricultural and hydrological applications, daily reference evapotranspiration (ET) is a key parameter. Based on one, two, and three-day antecedent meteorological parameters including average air temperature (T), sunshine hours (S), average relative humidity (RH), average wind speed (W), and reference evapotranspiration (ET), this study develops a new model framework for estimating daily ET magnitudes. The Long Short-Term Memory (LSTM) model was combined with the Feedforward Neural Network (FFNN) model in order to develop the hybrid approach, which was used to estimate daily ET values at two stations, Babolsar and Bandar Anzali, Iran, between 1990 and 2022. Additionally, 70% of the data in the present study were used for training purposes and 30% for testing purposes. Based on the statistical metrics, the proposed LSTM-FFNN model was benchmarked against the standalone LSTM model. Results from the comparison at Babolsar station indicated that the LSTM-FFNN model performed better than all scenarios with an error of approximately 1 millimeter per day in scenarios 12, 15, and 16. According to the standalone LSTM model, scenarios 5, 8, 11, 15, and 16 had a 1.57 mm/day error rate, and likewise, for the Bandar Anzali station, the combined LSTM-FFNN model performed the best in scenarios 14 and 15 with the lowest error rates (RMSE=0.96 mm/day), while the standalone LSTM model performed better in scenarios 15 and 16. It was found, however, that the proposed LSTM-FFNN model outperformed the standalone LSTM model at both sites. These results suggest that the proposed LSTM-FFNN model is capable of accurately estimating daily ET magnitudes, which may assist in the optimal management of water resources within agriculture. [ABSTRACT FROM AUTHOR]