Sequence-based statistical downscaling and its application to hydrologic simulations based on machine learning and big data

Abstract

In this study, a recurrent neural network (RNN) was used to perform statistical downscaling, and its advantages were showed compared to the traditional artificial neural network (ANN). The hydrological response to the downscaled meteorological data was evaluated using the Soil and Water Assessment Tool (SWAT) model. The results indicated that the temperature downscaled in southeastern China was better than that in northwestern China, while precipitation was the opposite. Although RNN and ANN model had different feasibility in different regions of China, the performance of RNN model for maximum and minimum temperature downscaling was about 6% and 10% better than that of ANN model overall, respectively. And RNN model was better for extreme temperature conditions simulation. Regarding precipitation, the performance of RNN and ANN model was similar when simulating precipitation amount. However, the use of RNN model improved the prediction accuracy of dry and wet days. In order to improve the accuracy of extreme precipitation downscaling, a new model, RNN-RandExtreme, was proposed. Compared with the ANN and single RNN model, RNN-RandExtreme model improved the prediction accuracy of extreme precipitation by 28.32% and 16.56%, respectively. The hydrological simulation results of SWAT model showed that the RNN and RNN-RandExtreme model significantly improved the accuracy of hydrological simulations of flow and evapotranspiration compared to the ANN model. However, as the time scale became rougher (from daily to annual scale), the improvement effect of RNN and RNN-RandExtreme model would weaken. The results of this study may help improving the accuracy of statistical downscaling, and support choosing downscaling models in different areas.