Forecasting evapotranspiration in different climates using ensembles of recurrent neural networks

Accurate ahead evapotranspiration forecasting is crucial for irrigation planning, for wetlands, agricultural and forest habitats preservation, and for water resource management. Deep learning algorithms can be used to develop effective forecasting models of ahead evapotranspiration. In this study, three Recurrent Neural Network-based models were built for the prediction of short term ahead actual evapotranspiration. Two variants of each model were developed changing the employed algorithm, selecting between long short-term memory (LSTM) and nonlinear autoregressive network with exogenous inputs (NARX), while the modeling was performed in the context of an ensemble approach. The prediction models were trained and tested using data from two sites with different climates: Cypress Swamp, southern Florida, and Kobeh Valley, central Nevada. With reference to the subtropical climatic conditions of South Florida, LSTM models proved to be more accurate than NARX models, while some exogenous variables such as sensible heat flux and relative humidity did not affect the results significantly. An increase of the forecast horizon from 1 to 7 days resulted in a slight reduction in the accuracy of both the LSTM- and NARX based models. Considering instead the semi-arid climate of Central Nevada, NARX models generally provided more accurate results, which were only slightly affected by relative humidity, sensible heat flux, and forecast horizon. On the other hand, LSTM models performance decayed if sensible heat flux and relative humidity were neglected, and if the forecast horizon was increased from 1 to 7 days. Deep learning-based models can provide very accurate predictions of actual evapotranspiration, but the performance of the models can be significantly affected by local climatic conditions.