Abstract
<jats:p>An analisis comparing modern methods for forecasting water regimes with different lead times was performed, contrasting neural networks with advanced statistical methods. The comparison utilized data of routine hydrometeorological observations for five test watersheds located in different physiographic zones of river runoff formation according to the classification by B.D. Zaikov (type II (rivers with flooding during the warm season), Far Eastern type II, East European type I and two catchments of North Caucasian type III). Despite limited hydrometeorological data, which constrained the effectiveness of neural network modeling, the standard forecast error criterion (the ratio of the root-mean-square error to the root-mean- square change over the forecast lead-time) based on deep learning (DL) turned out to be significantly better than the statistical methods used. For three test catchments DL models gave satisfactory skill scores differentiated by forecast lead times from one to ten days. For time-tested statistical methods such a result was obtained only for one test catchment (type II (rivers with flooding during the warm season)) with a forecast lead time of one day. The completed work, utilizing neural network technologies, demonstrates the validity of expanding the scope of scientific research related to physico-statistical mathematical modeling of streamflow generation. The prospects for developing this approach are highlighted in the transition toward a fundamentally new automated neural network system for producing operational hydrological forecasts with varying lead times for all gauged watersheds in the Russian Federation that are of economic importance.</jats:p>