Investigating methods in data preparation for stochastic rainfall modeling: A case study for Kermanshah synoptic station rainfall data, Iran

Given the climate changes, achieving rainfall forecast is of high importance and facing such challenges affected markedly in vast areas of societies. Accordingly, numerous nonlinear and linear methods have been developed. Most hydrological phenomena like rainfall are consisted of both linear and nonlinear parts. Modeling such phenomenon with stochastic methods like seasonal auto regressive moving average model (SARIMA), which are linear, demands data preparation prior to modeling. In this study, by investigating different forms of data preparation methods, variations in stochastic modeling results are scrutinized. The pre-processing methods used are categorized in two parts, normalization and stationarzition of data. The rainfall series is initially normalized by 4 transforms, namely: Manly(Mn), John-Draper (JD), Yeo-Johnson (YJ) and Scaling (Sc). The series, then, are stationarized by differencing, standardization (Std) and spectral analysis (Sf). After achieving preferred results by numerous tests, the preprocessed data are then modeled by stochastic SARIMA model. With regards to error and model sufficiency indices and graphs results, the acceptable results, but not the best, was obtained by the Sc-Diff combination, with SARIMA (0,0,1) (3,0,3)12 model and coefficient of determination, 0.355, variance accounted for, 0.353, root mean square error, 0.313, scatter index, 1.030, mean absolute error, 21.355), corrected Akaike Information Criterion, 1227.03. The results revealed that concerning the severe fluctuations in data, a supplementary method, like hybridization with artificial intelligence (AI) methods, is needed to achieve preferable results.