A Methodology for The Use of Long-Range Dependency and Machine Learning Algorithms for Rainfall Forecasting

Abstract

Forecasting rainfall is found to be an important indicator to the undisturbed continuation of social and economic activities. This derives the significance of predicting rainfall, providing better decision making in sectors such as agriculture, power, and tourism. This paper discusses the methodology behind applying long range dependency and machine learning in rainfall forecasting. It was determined that a time series should be established with data from past rainfall figures. This data will be analyzed through the Auto Correlation Factor (ACF) and the Partial Auto Correlation Factor (PACF) to check for seasonality and to identify the lag after which the other lags are insignificant. A time series process is categorized as long memory if the serial dependence or the ACF decays slower than an exponential decay. An exponentially decaying ACF can be categorized as having short memory. The spectral density of the time series will also be measured to gauge how far the peaks are from zero. The times series with peaks closer to zero will be considered as standard long memory, while the further peaks will be categorized as generalized long memory. The Ljung-Box Test and the Brusch-Godfrey Test will be used to assess whether any group of autocorrelations are different from zero. Based on the outputs from this analysis, it is possible to apply the time series into models such as ETS and ARIMA for prediction. The residual values from the models can determine the best fit model. Which can be further verified through machine learning.