Using NARX neural network to forecast droughts and floods over Yangtze River Basin

Drought and flood events are two extreme climate phenomena which usually bring enormous economic and social loss. For meeting the goal of flood and drought prevention, the nonlinear autoregressive with exogenous input (NARX) neural network is employed to bridge the data gap between the Gravity Recovery and Climate Experiment (GRACE) and GRACE Follow On (GRACE-FO) over Yangtze River Basin (YRB). The precipitation data from NASA Global Precipitation Measurement, temperature data from Global Historical Climatology Network and the Climate Anomaly Monitoring System, and terrestrial water storage anomalies (TWSA) from Global Land Data Assimilation System (GLDAS) are considered as the external inputs. Meanwhile, the performance of NARX models is evaluated for all possible combinations of time delays and neurons in order to find the optimal model structures. Then total storage deficit index (TSDI) is constructed based on TWSA reconstructions to assess drought and flood events over YRB, along with forecasting the extremes during the data gap period. The results show that when the number of time delays and neurons equals one and nine, respectively, the NARX model has an optimal performance with root mean square error (rmse), scaled rmse $$R^{ * }$$ , Nash-Sutcliff Efficiency (NSE) and correlation coefficient r of 1.34 cm, 0.34, 0.95 and 0.94, respectively. As indicated by TSDI and comparisons with previous studies, YRB has switched from drought periods to increased flood risks with a moderate correlation to global warming and El Nino-Southern Oscillation (ENSO). Finally, the most important conclusion that we successfully predict the flood events during the data gap period suggests that NARX neural network is promising for forecasting short-term hydrological extremes over YRB.