Forecasting monthly rainfall using artificial neural network

Abstract

Rainfall is one of the most difficult climatic variables to be forecasted. Due to complex relationship with other weather parameters, precipitation often comes out to be erratic and irregular in many cases. Being primary source for survival and agricultural production in major parts of the world, rainfall draws attention of the research workers to model and forecast the amount of precipitation precisely. In Indian context, prediction of rainfall is mainly important for the monsoon months (June to September) as more than 75% of the rainfall, in the country, is received during this season. It is always beneficial to study the rainfall in regional scale, such as block or district, for higher precision level that results effective crop and hydrological planning of the region. Artificial Neural Network (ANN) model, in many situations, is capable in explaining the behaviour of complex system, such as rainfall, to a substantial extent. In the present study, modelling of monthly rainfall (during monsoon) time series for Birbhum and North 24 Parganas districts of West Bengal state have been done employing promising Neural Network Autoregression (NNAR) technique. Lagged values of the time series were used as inputs to a multilayer feed-forward neural network with one hidden layer. The weights assigned to each node were learned and a nonlinear (sigmoid) activation function was applied in the hidden layer. NNAR was compared with traditional Autoregressive Integrated Moving Average (ARIMA) method for both in-sample and out-of sample forecast accuracy using different diagnostic measures. Predictive power of NNAR over ARIMA model, for the data under consideration, was established and residuals, obtained from the former, were found to be non-autocorrelated. Finally, forecasting of monthly rainfall during monsoon season was done from 2019 till 2020 for the two districts.