Leveraging Artificial Neural Networks for Rainfall Prediction in Medak District, Central Telangana, India

Abstract

This study explores the application of artificial neural networks (ANNs) for monthly rainfall prediction at the Medak station in Central Telangana, India. A total of 113 years of rainfall data was utilized, with 85 years (January 1901 to December 1985) used for model training and 28 years (January 1986 to December 2014) for testing. Input variable selection was carried out using the Gamma test, autocorrelation function, and cross-correlation function. The models were developed using a multilayer perceptron (MLP) trained with two learning algorithms—Levenberg-Marquardt and Delta-bar-delta—and employed two transfer functions: Sigmoid and Tanh. Model performance was evaluated both visually and through quantitative indices, including Root Mean Square Error (RMSE), Correlation Coefficient (R), Coefficient of Efficiency (CE), Percent Bias (PBIAS), and Integral Square Error (ISE). The results demonstrated that models using rainfall data from adjoining stations as inputs outperformed those using lagged data from the same station. Among all models, the M-8 model showed superior performance with higher R and CE values, and lower RMSE, PBIAS, and ISE values. These results indicate that the M-8 model is a promising tool for reliable monthly rainfall prediction in the Central Telangana region.