Long Short-Term Memory Networks to Predict One-Step Ahead Reference Evapotranspiration in a Subtropical Climatic Zone

Precise estimation and forecast of reference evapotranspiration (ET0) stand crucial for developing an efficient irrigation scheduling that helps better utilization of scanty water resources. One of the tools to predict ET0 is to employ machine learning algorithms that predict near future ET0 values based on past values from the ET0 timeseries. The aim of this research is to provide one-step ahead predictions of ET0 with different deep and machine learning methods using calculated past values of ET0. In this context, daily values of ET0 were computed via the FAO-56 Penman-Monteith approach that employ five climatic variables. For predicting one-step ahead ET0, this study evaluates the prediction accuracy and estimation capability of deep learning algorithms, i.e., a Long-Short Term Memory (LSTM) and a bi-directional LSTM (Bi-LSTM) network. The prediction performances of the LSTM and Bi-LSTM were compared with those of a Sequence-to-Sequence Regression LSTM network (SSR-LSTM) and an Adaptive Neuro Fuzzy Inference System (ANFIS). According to the findings, all four methods produced one-step ahead ET0 amounts in satisfactory precision and error levels with Correlation Coefficient, Index of Agreement, and a(20)-index ranging from 0.698 to 0.999, 0.833 to 0.999, and 0.624 to 1.0, respectively. Results further revealed the superiority of Bi-LSTM followed by SSR-LSTM, ANFIS and LSTM models identified by the ranking values computed using Shannon's Entropy. The overall results indicate that the Bi-LSTM model could be successfully employed to predict one-step ahead ET0 values quite precisely.

Automated summary

This study aims to provide one-step ahead predictions of ET0 using different deep and machine learning methods. The results show that the bi-directional LSTM performed the best, followed by the Sequence-to-Sequence Regression LSTM network, ANFIS, and LSTM models in terms of prediction accuracy and estimation capability.