Dynamic long-term streamflow probabilistic forecasting model for a multisite system considering real-time forecast updating through spatio-temporal dependent error correction

Abstract

Long-term streamflow forecasting for a multisite river basin system can experience oscillation of spatio-temporal dependent forecast errors. Characterizing the dependent information and conducting spatio-temporal dependent forecast error correction can reduce the total uncertainty in real-time forecasting. On the basis of a single-site martingale model of forecast evolution (MMFE), which can simulate forecast uncertainty evolution over the temporal scale, this study developed a dynamic long-term streamflow probabilistic forecasting model that addresses spatio-temporal dependent error correction for a multisite system. Error characteristics and hybrid spatio-temporal forecast improvements with complex dependencies are captured by Copula function, which describes the systemwide evolution of forecast uncertainty. The proposed model was verified by forecasting the streamflow of the Hongze Lake–Luoma Lake system in China. The results revealed the following. 1) The spatio-temporal dependent error correction based on MMFE can refine the spatio-temporal dimension hybrid error information, which is helpful in improving forecast accuracy. 2) As the spatio-temporal dependence of forecast uncertainty is described using a Copula function, selection of the marginal distribution and connection function is flexible through combining statistical and informational characteristics of the errors, which is helpful in improving the fittedness of complex system error simulation. 3) Compared with the benchmark forecast model without correction, the proposed model greatly decreased forecast uncertainty by reducing the standard deviation of the errors, continuous ranked probability score (CRPS), and Brier score (BS) by nearly 40%, 11.4%, and 34.6%, respectively, and further reduce the CRPS by 2.1% on average (highest value: 3.7%) and reduce the BS by 1.2% on average (highest value: 2.0%) in comparison with only temporal correction. The results demonstrate that the proposed model improves the accuracy and reliability of probabilistic streamflow forecasting for complex water resource systems by reducing uncertainty.