Long-Term Seasonal Rainfall Forecasting using Regression Analysis and Artificial Neural Network together with Larg-Scale Circulation Indices

Abstract

Long-term rainfall prediction several months in advance plays an important role in water management, especially for countries dependent on agriculture. The objective of this study is to forecast the long-term rainfall of eight rain gauge stations in the Phetchaburi River Basin, Thailand, 12–18 months in advance using linear regression (simple linear regression (SLR)/multiple linear regression (MLR)) and non-linear relations (polynomial regression (PR)/artificial neural network (ANN)). Seven atmospheric circulation indices: ONI, DMI, MEI V. 2, NINO4, NINO3.4, NINO3, and NINO1+2, together with historical rainfall data, were used as predictors in the models. To avoid bias in empirical equation construction, one-year cross-validation was also applied together with a one-month moving window average approach from January to July of the preceding year (12–18 months lead time) to seek suitable periods of predictors for predicting rainfall. The results reveal that the surface temperature indices of the Indian Ocean (DMI) and Pacific Ocean (NINO) are the most essential for forecasting rainfall. MEIV2 and ONI were positively correlated with local rainfall only when non-linear regression was used. Non-linearity models showed better forecasting skills compared to linear regression. The suitability of periods varied according to the statistical models and selected predictors.