A Multi-imputation Method to Deal With Hydro-Meteorological Missing Values by Integrating Chain Equations and Random Forest

Imputing hydro-meteorological missing values is essential in time series modeling. Imputation of missing values was traditionally performed after an observation period, which cannot effectively support water resources management in time. Therefore, it is necessary to deal with the missing data online. Moreover, traditional imputation methods usually consider only one observation variable and generate one set of imputations, which cannot describe the imputation uncertainty. Thus, a multiple-imputation method is proposed in this paper by integrating chain equations and random forest, namely, MICE-RF, to deal with the hydro-meteorological missing values. MICE-RF first simulates multiple imputation series to obtain the optimal imputations using the evaluation results of multiple imputation series. The traditional linear imputation, mean imputation, spline imputation, and k nearest neighbor imputation are compared to illustrate the robustness, reliability, and accuracy of the MICE-RF. According to the results, the MICE-RF provides the best imputation accuracy and can be easily implemented online.

Automated summary

This paper proposes a multiple-imputation method, MICE-RF, which integrates chain equations and random forest to deal with hydro-meteorological missing values. MICE-RF provides the best imputation accuracy, supports water resources management in time, and describes imputation uncertainty.