An Efficient Approach of Focused Time Delay Neural Network in Drought Forecasting in Central Iran
By: Abbasali Vali, Fatemeh Roustaei
Key Words: Drought forecasting, Precipitation Index, Central Iran.
J. Bio. Env. Sci. 9(1), 231-244, July 2016.
An exact prediction and modeling of drought is essential for watersheds management. The main contribution of this research is in the design, performance and comparison of drought forecasting models using Focused Time Delay Neural Networks (FTDNN). The network was trained to perform one-step-ahead predictions. Standardized Precipitation Index(SPI) were applied in various time scales including 3, 6, 9, 12, 18, 24 and 48 monthly time series in 14 synoptic stations in Central Iran during 1965–2014. Five categories of back-propagation training algorithms namely resilient back propagation (RP), batch gradient descent (GD and GDX), Quasi-Newton (BFGS), conjugate gradient (CGF, CGP, and CGB) and Levenberg-Marquardt (LM) were used. Then, according to the best algorithm, the number of neurons in the hidden layer was optimized and the best performance was identified. The number of epochs, high Correlation Coefficient (R2), least Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) were considered to evaluate the performance of the FTDNN model at each step. The result showed that the Levenberg-Marquardt (LM) was the best algorithm and node 31 was the most efficient for drought prediction. Finally, the designed network was applied on all of the SPIs time series to determine the best in prediction according to statistical parameters. It was found that better results can be achieved by increasing the duration of the time series. According to the results obtained, FTDNN trained by LM is an efficient tool to model and predict drought events especially in long term time series.