Hybrid machine learning framework for hydrological assessment

Abstract

This study introduces a novel hydrological assessment tool (HAT) based on hybrid machine learning (HML) framework. The HML framework combines an unsupervised clustering technique and a supervised classification technique, to determine reasonable performance ratings (unsatisfactory, satisfactory, good, and very good) and build a practical assessment tool. Hydrologically significant error indices are used to cluster the performance rating groups and train the HAT. The HAT was applied to the National Water Model (NWM), which is operated in real time for the continental United States (CONUS). For establishing, training, and validating the HAT, data from October 2013 to February 2017 were used, and a performance assessment was conducted on the NWM in the San Francisco Bay Area. As a result, the HAT determined the performance ratings that were reliable in terms of the statistics and hydrograph. It was confirmed that the HAT could perform an accurate hydrograph assessment as the concordance rate of the performance ratings was 98%. The NWM was evaluated against 57 USGS streamflow gauges using the HAT and was found to perform with 46% on average, good and very good ratings. The HML framework, an integral part of the HAT, is expected to be useful not only in hydrological analysis but also across all geophysical fields that deal with physical processes.