Rainfall Forecasts of Yemen Based on Statistical and Probabilistic Approaches

Since reliable rainfall data are not available for the Republic of Yemen, this study analyzed hypothetically generated monthly rainfall data from January 1981 to December 2018, using POWER, 2019 information. Therefore, for each geomorphological basin, 3 representative hypothetical station locations are proposed with the help of Theissen polygon approach. After determining the representative annual average rainfall datasets of four basins (Red Sea, Arab Sea, Gulf Aden and Rub 'Al Khali), five independent parametric and non-parametric data quality tests for each basin; Homogeneity, Consistency, Normality, Trend and Stationarity Unit-root are applied. To predict rainfall data for the next 3 years, 27 different ARIMA models were proposed for each basin and tested through Akaike Information Criteria (AIC). Among them, the best 3 representative models for each basin were selected and determined by suggested weighted average (1, 2 and 3). For this reason, the annual average dataset of each basin from 1981 to 2008 (28) was used to train these models, and the remaining annual average dataset from 2009 to 2018 (10) was used to test these trained datasets. Then, among the three selected models of each basin, the most suitable model was selected and used to predict the annual rainfall data for 3 consecutive years (2019, 2020 and 2021). These models are ARIMA (0,1,1) for the Red Sea Basin, ARIMA (0,1,1) for the Arab Sea Basin, ARIMA (2,1,2) for the Gulf Aden Basin, and ARIMA (0,1,2) for the Rub 'Al Khali Basin. As a part of this study, three different frequency distributions among the commonly used ones; Normal, log-normal and Pearson Type III (Gamma) were selected and the most representative frequency distribution function for each basin was determined by selecting that distribution having the closest p value to 1.00. Also, based on Moving Averages with different time windows (2 to 9), the annual rainfall trend of each basin was determined. It has been determined that all basins showed a decreasing trend in the range of 1 - 2 mm/year. Similarly, annual average rainfall data sets for each basin were systematically analyzed for wetness or dryness periods, and all basins were interpreted to be under the influence of the dry spell.