Streamflow Forecasting without Models

Abstract

The authors explore persistence in streamflow forecasting based on the real-time streamflow observations. They use 15-min streamflow observations from the years 2002 to 2018 at 140 U.S. Geological Survey (USGS) streamflow gauges monitoring the streams and rivers throughout Iowa. The spatial scale of the basins ranges from about 7 to 37 000 km2. Motivated by the need for evaluating the skill of real-time streamflow forecasting systems, the authors perform quantitative skill assessment of persistence schemes across spatial scales and lead times. They show that skill in temporal persistence forecasting has a strong dependence on basin size, and a weaker dependence on geometric properties of the river networks. Building on results from this temporal persistence, they extend the streamflow persistence forecasting to space through flow-connected river networks. The approach simply assumes that streamflow at a station in space will persist to another station which is flow connected; these are referred to as pure spatial persistence forecasts (PSPF). The authors show that skill of PSPF of streamflow is strongly dependent on the monitored versus predicted basin area ratio and lead times, and weakly related to the downstream flow distance between stations. River network topology shows some effect on the hydrograph timing and timing of the peaks, depending on the stream gauge configuration. The study shows that the skill depicted in terms of Kling–Gupta efficiency (KGE) > 0.5 can be achieved for basin area ratio > 0.6 and lead time up to 3 days. The authors discuss the implications of their findings for assessment and improvements of rainfall–runoff models, data assimilation schemes, and stream gauging network design.

Multi‐Scale Hydrologic Evaluation of the National Water Model Streamflow Data Assimilation

“…We used persistence-based predictions as reference to assess the DA-based prediction results. The persistence method incorporates streamflow observations from the same upstream stations used in DA and its concept is rather simple but efficient (e.g., Krajewski et al 2020). We found that DA outperforms persistence, particularly at catchment scales smaller than 5,000 km 2 (the number might be different at different regions depending on the configuration of stream gauge network), where the coverage fraction is not as good as the one for larger scales as shown in Figure 5c.…”

“…Persistence leads to underestimations in volume and peak discharge, and early peak timing, as illustrated in Figure 3; drainage areas (represented by single or multiple upstream gauging stations) that are smaller than the area represented by the downstream evaluation station yield the observed underestimations and early peak. However, the overall performance (KGE) of persistence seems better at many locations than that of model simulation with NoDA, implying that the forecasting approach without models can provide useful guidance if there are reliable gauging stations upstream (see Krajewski et al 2020). Overall, the NWM with DA outperforms persistence and NoDA based on KGE.…”

“…NWM retrospective analysis with streamflow DA requires meteorological forcing products (e.g., precipitation) and streamflow observations, and we collected these data for the period of 2015-2018. We note that several earlier studies (Seo et al 2018;Krajewski et al 2020;Seo and Krajewski 2020) include a variety of evaluation (e.g., precipitation) and analyses of these data for the common temporal and spatial domain used in this study.…”

A Regional Semi-Distributed Streamflow Model Using Deep Learning

“…As a benchmark in hydrology, persistence forecast is also known as a hard to beat model for large watersheds at short to medium forecast range. Many studies suggest that persistence can be an alternative for evaluating future estimates of streamflow at short to medium range (Ghimire and Krajewski, 2020;Krajewski et al 2020).…”

Nitrogen legacies in anthropogenic landscapes: a case study in the Mondego Basin in Portugal

“…There is no unique threshold value of KGE that fits all applications (Knoben et al, 2019). Krajewski et al (2020) considers that model provides good results if KGE > 0.5. A prior application of ELEMeNT also considers that results are reliable for a KGE > 0.5 (Liu 2020).…”