Comparison of missing data imputation methods applied to daily temperature and precipitation data in Turkey

2023-8-07
Gezgen, Didem
A significant portion of the data under analysis contains missing values, which hinders the generation of meaningful results, particularly when dealing with time-dependent data where the order of observations is crucial. This issue leads to unreliable outcomes in statistical analyses applied in fields such as meteorology and economy. To address this challenge, handling missing values meticulously in time-dependent data is imperative. In this thesis, daily average temperature and total precipitation data, obtained from the General Directorate of Meteorology of Turkey, were utilized. The primary objective was to impute the missing values in these datasets using various methods and subsequently compare their performance. Missing values were intentionally introduced into the temperature and precipitation data. The methods employed for imputation included Simple Arithmetic Average Method (SAA), K-Nearest Neighbor Method (KNN), Random Forest Method (RF), Multiple Imputation by Chained Equation Method (MICE), and Generalized Adversarial Imputation Network (GAIN). The outcomes were assessed based on the Root Mean Square Error (RMSE), Coefficient of Variation of Root Mean Square Error (CVRMSE), and Nash-Sutcliffe Efficiency (NSE). The results indicated that Random Forests exhibited superior performance in most cases, followed by KNN and GAIN.
Citation Formats
D. Gezgen, “Comparison of missing data imputation methods applied to daily temperature and precipitation data in Turkey,” M.S. - Master of Science, Middle East Technical University, 2023.