Application of multiple linear regression for post-processing of the RegCM4 model outputs in forecasting precipitation

The seasonal forecasts of precipitation play a major role in agricultural and water resource management and also in monitoring extreme events such as drought and flood. The Earth Systems Physics (ESP) group of the Abdus Salam International Centre for Theoretical Physics (ICTP) maintains and distributes a stateof- the-science regional climate model called the ICTP Regional Climate Model (RegCM), which has been successfully used in different regions of the world for a diverse range of climate-related studies. This study was performed with two aims: (1) to evaluate the performance of the RegCM4 dynamic model in forecasting monthly, seasonal and annual precipitation in four selected stations in the northwest of Iran, i.e. Tabriz, Ardabil, Khouy and Urumia; and (2) to examine the accuracy of a stepwise regression technique for post processing of the outputs of the model for a 30-year period from 1982 to 2011. In order to run the RegCM4, the required observed weather data of the study stations were collected from the Iran Meteorological Organization (IRIMO) archive, while the rest of the data were collected from the ICTP database including three sets of the weather data: NCEP/NCAR Reanalysis Product version 1 (NNRP1) with a 6-hour time step and a horizontal resolution of 2.5°×2.5° on the reanalysis data from the National Center of Environmental Prediction (NCEP) of the United States, Sea Surface Temperature (SST) of the Optimum Interpolation Sea Surface Temperature (OISST) type, retrieved from the National Oceanic and Atmospheric Administration (NOAA) database and surface data (SURFACE), which were consisted of three topographic features: Global Topographic (GTOPO), vegetation or land use Global Land Cover Characterization (GLCC), and soil type Global Zobler (GLZB) data, with a horizontal resolution of 30×30 seconds from the United States Geological Survey, for the period 1982–2011. To determine a suitable rainfall scheme, the normal year of 2009 was chosen for running the model using different schemes. Accordingly, the Kuo scheme with a minimum bias compared to the observed precipitation amounts in the entire 36 synoptic stations of the region was selected as the best scheme. The time step was set to 100 seconds, with a spatial resolution of 30×30 km2, and the number of grid points were 152 in longitude (iy) and 168 in latitude (ix) during the study period. The geographical area center was placed at 30.5° N and 50° E. Nine significant variables (excluding, total precipitation; tpr) having the highest correlation with precipitation were determined as q2m, t2m, ps, v1000, v500, u1000, u500, omega1000, and omega500. For post-processing of the outputs of the model, the multiple linear regressions (MLR) approach was used. Except for the warm months, the output of the RegCM4 showed a wet bias, and overestimation. Applying the multivariate linear regression equation (and sometimes two-variables) to the output of the model led to a better agreement between the observed and simulated values of precipitation, such that in 75% of the cases, the bias and relative error decreased for the monthly, seasonal and annual forecasts. At all stations, except for Urumia, performing the post processing improved the accuracy of the RegCM4 output at all time scales. Further scrutiny is recommended for explaining the variations among the stations.