Automatic gap-filling of daily streamflow time series in data-scarce regions using a machine learning algorithm

摘要

Complete hydrological time series are crucial for water and energy resources management and modelling in a changing climate. The reliability of the non-parametric stochastic machine learning algorithm MissForest was assessed for gap-filling of daily streamflow time series in a data-scarce region with strong climatic variability. A total of 1,586 reconstructions of streamflows for 1970-2016 were analyzed. Overall, MissForest performed satisfactorily to well, allowing a precise and reliable simulation of the missing data quickly and automatically. MissForest performance increased with the number of predictor records and record length, achieving satisfactory results with 20 or more records having 15 or more years in length. Reconstructed daily streamflow time series of rivers with natural flow regimes were simulated with good performance, which slightly decreased for discharge magnitude alterations by runoff inputs from urbanized areas and water diversion for irrigation. In cases of severe alterations of the flow regime, such as by hydropeaking, MissForest failed at filling daily streamflow series gaps. Reconstructed hydrographs allow analysis of streamflow change and variability and their interactions with key climatic variables.