Statistical Modeling of Weekly Rainfall Data for Better Planning

Abstract

Prediction in short-term rainfall and other forms of precipitation is one of the most critical factors determining the overall impact of climate change. Thus, in this studyan attempt is made to predict weekly rainfall. The chaotic nature of weekly rainfall in South Asian countries indicates a necessity for methods which go beyond theconventional context. Alternatively, time series approach and artificial neural network (ANN) approach are employed to predict the weekly rainfall which exhibits theblend features of the linear and non-linear phenomenon. The rainfall data from 1990 to 2017 in Colombo city were obtained from the Department of Meteorology, SriLanka (DMSL). The correlation structure of the series confirmed seasonal length of 52 and it was found that there is no cyclic pattern. Seasonality of the series wasremoved by an additive decomposition method and then AR (1)-GARCH (1,1) for deseasonalized series was identified as the best fitted model to describe the pattern ofweekly rainfall. ANN approach facilitates mapping the non-linear relationship between input and output using mathematical model without having proper priorknowledge of the considered phenomenon. The feed foreword Multi-Layer Perceptron (MLP) neural network and three recurrent neural network architectures: SimpleRecurrent Neural Networks (SRNN), Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) were utilized to find the best one. The forecastingperformance of the two approaches were assessed using root mean square error, absolute error, mean absolute error and coefficient of correlation measures. Themodels were trained using data from 1990 to 2014 and were validated using data from 2015 to 2017. Both approaches performed well and provided encouragingforecasting result. However, based on the forecasting degree, GRU model outperformed the time series model to forecast weekly rainfall series in Colombo city. Thedeveloped model can be used for better sort-term planning in rainfall.