A modular technique for monthly rainfall time series prediction

Abstract

Rainfall time series forecasting is a crucial task in water resource planning and management. Conventional time series prediction models and intelligent models have been applied to this task. Attempt to develop better models is an ongoing endeavor. Besides accuracy, the transparency and practicality of the model are the other important issues that need to be considered. To address these issues, this study proposes the use of a modular technique to a monthly rainfall time series prediction model. The proposed model consists of two main layers, namely, a prediction layer and an aggregation layer. In the prediction layer, Mamdani-type fuzzy inference system is used to capture the input-output relationship of the rainfall pattern. In the aggregation layer, Bayesian learning and nonlinear programming are used to capture the uncertainty in the time dimension. Eight monthly rainfall time series collected from the northeast region of Thailand are used to evaluate the proposed model. The experimental results showed that the proposed model could improve the prediction accuracy from the single model. Furthermore, human analysts can interpret such model as it contains set of fuzzy rules.