Development of an enhanced bidirectional recurrent neural network combined with time-varying filter-based empirical mode decomposition to forecast weekly reference evapotranspiration

Evapotranspiration is one of agricultural water management's most significant and impactful hydrologic processes. A new multi-decomposition deep learning-based technique is proposed in this study to forecast weekly reference evapotranspiration (ETo) in western coastal regions of Australia (Redcliffe and Gold Coast). The time-varying filter-based empirical mode decomposition (TVF-EMD) technique was used to first break down the original meteorological variables/signals into intrinsic mode decomposition functions (IMFs), which included maximum and minimum temperature, relative humidity, wind speed, and solar radiation. Using a partial autocorrelation function (PACF), the significant lagged values were then calculated from the decomposed sub-sequences (i.e., IMFs). A novel Extra Tree-Boruta feature selection algorithm was used to extract important features from the decomposed IMFs. Four machine learning approaches, including bidirectional recurrent neural network (Bi-RNN), multi-layer perceptron neural network (MLP), random forest (RF), and extreme gradient boosting (XGBoost), were used to forecast weekly evapotranspiration using the TVF-EMD-based decomposed meteorological data. Different statistical metrics were applied to evaluate the model performances. The results showed that the decomposition of the input data by TVF-EMD significantly improved the accuracy compared with the non-decomposed inputs (single models without decomposition). The findings indicate that the TVF-BiRNN model, as presented, achieved the highest level of accuracy in simulating weekly ET0 at both the Red-cliffe and Gold Coast stations (Redcliffe: R=0.9281, RMSE=3.8793 mm/week, MAPE = 9.2010%; Gold Coast: R=0.8717, RMSE=4.1169 mm/week, MAPE = 11.5408%). The novel hybrid modeling technique can potentially improve agricultural water management through its ability to generate more accurate ETo estimates weekly. The proposed methodology exhibits potential applicability to various other environmental and hydrological modeling issues.