Novel methodology for prediction of missing values in river flow based on convolution neural networks: Principles and application in Iran country

Abstract

In the present study, for the first time, a novel methodology has been introduced for constructing missed runoff data in Iran country. To this end, the convolution neural network (CNN) is developed based on various types of data, including basin characteristics, time data, the geography of each station, the statistical characteristics of river flow that do not have missing data, and the statistical characteristics of river flow. Furthermore, quantile mapping is used to correct bias in CNN results. Seven CNN structures were defined, and the results were compared with deep neural networks and machine learning algorithms. The obtained results of runoff modeling in the 1666 hydrometric station indicated the superiority of the best CNN structure (CNN4) with mean absolute error = 5.95m3/s, root mean square error = 25.61m3/s, relative root mean square error = 0.44, and Nash Sutcliffe efficiency coefficient = 0.81. In addition, the distribution of runoff modeled with CNN4 was more similar to observed runoff than those for other algorithms. Finally, the runoff time series for all stations was constructed, even for stations with 100% missing data. This study's methodology can potentially estimate missing data in runoff river data from other countries.