Abstract Precipitation is one of the most important part of weather and climatology. prepare an accurate estimation of daily precipitation data is difficult. First requirement of analyzing in meteorology is preparing the statistical data for a long-term. The analysis of incomplete data series is biased. There is a serious need for the estimation of the lost data. Although there are broad ranges of methods to complete the lost data, percentage of statistical gaps is known as one of the most important limitation of their utilization. Under-studied area of Mashhad-Chennaran plain, with area of 9909 square kilometers, which is part of the western part of the Qaraqom drainage basin located in the northeast of Iran. After the data adequacy test, 30 rain gauge stations out of 34 rain gauge stations in Mashhad-Chennaran plain were used in this research. The aim of this thesis is reconstruct the lost data of rain gauge stations with high accuracy to present a completed statistical series (at least 30 years). The present study compares two methods for restoring and completing lost rainfall data values, which are included: Multiple calculations by chain equations through matching predictive mean (PMM) and random forest (RF) machine learning algorithm available in R programming software. In order to increase accuracy, reduce errors and improve the data interpretation with the mentioned methods a cluster analysis was employed on the stations in this project by Ward's hierarchical method. This algorithm was selected as the best algorithm among the existing algorithms in this method based on the Cofentic correlation coefficient criterion (0.73). The 8 clusters were selected by the optimal case and the average width of the silhouette (0.26). Moreover, the 30 rain gauge stations were selected based on the average and standard deviation of seasonal rainfall depth, geographic location (longitude and latitude) and Altitude above sea level were categorized into the mentioned clusters. The duration of the statistical period in each cluster was determined based on the common base year of the stations in that cluster. In the next step, a complete series of daily data with more time limitation frame (from 2015 to 2017) at 6 levels (5%, 10%, 15%, 20%, 30% and 40%) were lost by R software, artificially and completely randomly, to assess the models more accurate during the reconstruct and validation process, and to compare the models of the 8 rain gauge stations that had a complete series of daily rainfall data were used in the mentioned period. Then the incomplete series was ccompleted by PMM and RF methods and eva‎luated by dr and R2 performance criteria as well as RMSE, NRMSE, GSD, CRM, MBE, MAE and RLE error criteria on a daily and monthly scale. Hence, the optimal values of the function parameters were calculated based on the number of complete series in the PMM method, which was equal to 5 (m=5) and the 50 time repetitions (maxit = 50). Also, these were used as the basis for reconstruct with the PMM method. In the RF method, the number of calculation trees 500 (n_est=500) and the number of repetitions 100 (max_iter=100) were chosen as the optimal parameters of the model. Furthermore, based on the performance and error criteria of the models, the RF method showed better results in most of the lost levels, also the reconstructed data were closer to their real values. In the end, restoring process was employed on the main data (long-term missing data) of the rain gauge stations in the desired clusters by PMM and RF methods. While the model eva‎luation results showed that the RF method has a better performance than the PMM method in most of the clusters, when the stations have a severe statistical gap or have many zero data in their statistical series PMM showed better results, because RF method suffers performance degradation and reconstructs data more than its true value, which also seems to be related to the length of the statistical period.

سعيد اسلاميان

ابوالفضل مساعدي

عليرضا گوهري , رضا مدرس