Prediction of Seasonal Southwest Monsoon Rainfall over India using Artificial Intelligence

Abstract

Indian Southwest Monsoon Rainfall (ISMR) plays a very important role in Indian economy – water management, agricultural operations, ground water recharge etc. The all India average rainfall during June to September, monitored by ground rain gauges distributed across the country since the 1870s, has been nearly steady around 90 cm. The interannual variation (IAV) of ISMR, though only 10%, has proved to be a very stable and important parameter to monsoon prediction with dynamical methods. Over the decades, its tele connection with many antecedent global parameters have proved to be potential predictors of ISMR well in advance of the season.

Various methods, both statistical (eg. Auto Regression, ARIMA etc based on teleconnections) and dynamical (based on physical processes) methods have been employed over the last few decades. Although the statistical models are good time series predictors, being based on stationary data, the results have not proved to be much reliable, particularly during poor monsoon years.

In our approach, we seek to compare and contrast the performance of previous machine learning and statistical approaches and propose the best possible prediction model for seasonal forecast of ISMR. The data used in our technique comprises six critical global oceanic parameters, which affect the southwest monsoon over the Indian landmass, along with the seasonal rainfall time series data for the period 1958 to 2020. We aim to reduce the error margin of the predictions prevalent in the previous techniques to less than 4% of Long Period Average (LPA).