Artificial Neural Network for Monthly Rainfall Forecasting Using Teleconnection Patterns (Case Study: Central Plateau Basin of Iran)

Rainfall is final result of complex global atmospheric phenomena and long-term prediction of rainfall remains a challenge for many years. An accurate long-term rainfall prediction is necessary for water resources management, food production and evaluation flood risks. Several large scale climate phenomena affect the occurrence of rainfall around the world; of these large scale climate modes El Nino southern Oscillation (ENSO) and Multivariate ENSO Index (MEI) are well known. Many studies have tried to establish the relationship between these climate modes for daily, monthly and seasonal rainfall occurrence around the world but the majority of these studies did not consider the effect of lagged climate modes on future monthly rainfall predictions.
This study focuses on investigating the use of combined lagged teleconnection patterns as potential predictors of monthly rainfall. Direct Multi Step Neural Network (DMSNN) approach was used for this purpose. Four regions (east, center and west) of Central Plateau Basin of Iran were chosen as case studies, each having many rainfall stations. Hence, precipitation data in a common statistical period of 1981-2014 in 20 synoptic stations in the study area were selected and that the data during 1981-2004 were considered to develop the model and the data during 2004-2014 were used for validation the model in order to predict the next 6 months in monthly time scale. Based on the cross correlation function (CCF) results, MEI (Multivariate ENSO Index) and SOI (Southern Oscillation Index) had strong impact on precipitation of the region.
Direct Multi Step Neural Network (DMSNN) modelling was also conducted for the 20 stations of Central Plateau Basin of Iran using the combined lagged MEI and SOI. Multilayer Perceptron (MLP) architecture was chosen for this purpose due to its wide use in hydrologic modeling. To determine the best combination of learning algorithms, hidden transfer and output functions of the optimum model, the Levenberg–Marquardt and backpropagation algorithms were utilized to train the network, tangent sigmoid equations used as the activation functions and the linear equations used as the output function.
The values R2 (Correlation Coefficient), RMSE (Root Mean Square Error), and MAE (Mean Absolute Error) parameters were used to explore the efficiency of the model.
ANN models generally showed lower errors and are more reliable for prediction purposes. After calibrating and validating the models they were tested on out-of-sample sets. ANN was able to perform out of sample test with correlation coefficient of of 0.81 for the South, and 0.4 for West of Central Plateau Basin of Iran. Although the effect of SOI and MEI in the west is quite weak, however with the use of combined lagged SOI–MEI sets Direct Multi Step Neural Network (DMSNN) modeling, long term rainfall forecast can be achieved. Thus, the results showed that the predicted data preserved the basic statistical properties of the observed series.
The results of this research showed that teleconnection indices are suitable inputs for intelligent models for rainfall prediction. Computing the best structure of artificial neural network models showed that DMSNN can predict rainfall most accurately.
Accurate long term rainfall forecasting can contribute significant positive impacts in water resources management. Central Plateau Basin of Iran climate is greatly fluctuating; at times it goes through severe drought years, then suddenly it experiences wet periods and dry. During drought periods, water supply and irrigation sectors are affected severely; proper prediction of such drought period helps water managers and users to have well-planned, coordinated allocation of resources. Also, prediction of the wet years helps flood management authorities to have well-planned flood disaster management. In addition to predicting rainy month in advance, the developed ANN models are also capable of predicting the intensity of seasonal rainfall.