Forecasting Reference Evapotranspiration Using Time Lagged Recurrent Neural Network

The aim of this study is to employ a Time Lagged Recurrent Neural Network (TLRNN) model for forecasting near future reference evapotranspiration (ETo) values by using climate data taken from meteorological station located in Velestino, a village near the city of Volos, in Thessaly, centre of Greece. TLRNN is Multilayer Perceptron Neural Network (MLP-NN) with locally recurrent connections and short-term memory structures that can learn temporal variations from the dataset. The network topology is using input layer, hidden layer and a single output with the ETo values. The network model was trained using the back propagation through time algorithm. Performance evaluations of the network model done by comparing the Mean Bias Error (MBE), Root Mean Square Error (RMSE), Coefficient of Determination (R2) and Index of Agreement (IA). The evaluation of the results showed that the developed TLRNN model works properly and the forecasting ETo values approximate the FAO-56 PM values. A good proximity of predictions with the experimental data was noticed, achieving coefficients of determination (R2) greater than 75% and root mean square error (RMSE) values less than 1.0 mm/day. The forecasts range up to three days ahead and can be helpful to farmers for irrigation scheduling.