Uncertainty in operational hydrological forecasting: Insights from SMHI’s services

Contributed by Ilias Pechlivanidis  (SMHI), member of the SMHI Guest Columnist Team

Background

The production of hydrological forecasts generally involves the selection of model(s) and their setup, calibration and initialization, verification and updating, generation and evaluation of forecasts. However, the precision of hydrological forecasts is often subject to both epistemic and aleatory uncertainties, with the former being related to various components of the production chain and the data used.

Aleatory uncertainty refers to quantities or natural phenomena that are inherently variable over time and space, and hence characterised as random or stochastic, and epistemic uncertainty is related to the lack of understanding of a hydrological system (e.g. model structure and parameters), being further propagated to the description of the system.

In operational systems, we commonly use field observations to calibrate and initiate hydrological models; however, recent technological advancements have allowed us to use additional information, i.e., remote sensing data and meteorological ensemble forecasts, to improve hydrological forecasts (Olsson and Lindström, 2008). For instance, the Ensemble Prediction System (EPS) approach is used to acknowledge the uncertainty in the meteorological initial conditions and to generate probabilistic forecasts.

SMHI operationally produces hydrological short-term forecasts (10 days; deterministic and probabilistic) in Sweden based on the HBV (Lindström et al., 1997) and the S-HYPE (Lindström et al., 2010) hydrological models (Table 1). Although some techniques and methods are commonly shared between the two models, uncertainties can impact differently the performance from the two services.

My objective in this blog post is to present some of these uncertainties in the production chain and some ways to reduce them.

Table 1. Datasets used for hydrological forecasting at SMHI. 

	Period	HBV	S-HYPE
Hindcast datasets	1961 – 1 month prior today	Archive pthbv*	Archive pthbv*
	1 month prior – Today 06:00	Realtime pthbv*	Realtime pthbv*
Forecast datasets	Today 06:00 – +10 days 06:00	Deterministic 51 ensemble EPS	Deterministic -**

*pthbv: gridded 4 km data set of temperature and precipitation over Sweden

**Current operational service includes EPS for S-HYPE

Model setup, structure and parameters

S-HYPE (its 2012 version, named here as SH2012) simulates the same processes as HBV, but includes more water pathways and its parameters are more linked to physiographical characteristics in the landscape (e.g. HRUs). S-HYPE is also setup for a resolution of some 37000 subbasins, while HBV for some 1000 subbasins. S-HYPE also explicitly models the routing through a large number of lakes.

Both systems operate at a daily time step; however this can limit the understanding of processes at the subdaily level, which could occasionally result into low forecast performance for some events. This is particularly observed during the days/periods of 0^oC temperature, and hence small temperature deviations within the day would in reality result into mixed processes of snow melting and accumulation, while on the aggregated modelled daily time step this would either be melting or accumulation (this is partially compensated by using a temperature interval over which the snowfall fraction decreases from 100% to 0%).

Nevertheless, Fig. 1 shows that both models (HBV Non-AR and SH2012 Non-AR) achieve a good performance (Kling-Gupta Efficiency KGE > 0.6) in all lead times. S-HYPE outperforms the HBV since the former comprises a real-time updating of the water discharge downstream of gauges for some of the stations, a feature which is lacking in the HBV forecasting system.

Updating methods

The operational flood forecasts at SMHI are primarily updated using an autoregressive (AR) forecast of the error. State updating and/or corrections of the input data are also sometimes used. Measured discharge and water levels in lakes are used for correction of modelled values by replacement of calculated values by the observed ones.

Fig. 1 shows the effect of AR updating of the forecasted discharge (without versus with AR). Introduction of AR updating can significantly improve the forecasts accuracy (KGE close to 0.9 for lead day 1 and 2), particularly due to the contribution of the volume and peak correction (Pechlivanidis et al., 2014). S-HYPE seems to achieve slightly better performance than HBV after lead day 4, again due to the update in the water level in the lakes and upstream discharge, which has a longer term impact on the performance rather than the AR updating.

Ensembles forecasts

Fig. 1 shows that, overall, the deterministic and the ensemble HBV systems are fairly similar. Digging deeper by decomposing the performance into terms describing different attributes of the flow signal, we show that this is due to the lack of adequately capturing the distribution in observed discharge (Fig. 2).

Note that the deterministic HBV forecasts used here are based on a high-resolution model, whilst the ensemble median is here used to represent the ensemble spread. However, Arheimer et al. (2011) showed that ensemble forecasts can be of added value, particularly if the difference between the probabilistic and the deterministic forecasts is large.

Summary

Overall, both HBV and S-HYPE services are capable of producing adequate forecasts with the performance steadily decreasing in lead time.

1. In the systems without AR-updating, S-HYPE shows superiority to its HBV counterpart during all lead times, highlighting the importance of updating the water level in the lakes and upstream discharge.
2. The AR updating method can reduce the epistemic uncertainty and improve the performance of the two systems. This is mainly because the method has a significant impact on the improvement of discharge volume. S-HYPE seems to perform slightly better than HBV in the longer lead time, probably because the S-HYPE system is capable of updating the lake water level, which has an impact on the longer lead times.
3. Moreover, the deterministic and ensemble HBV systems with AR updating perform fairly similar for all lead times. This could be subject to the high quality archive national dataset (pthbv), which was used to drive the deterministic model; however ensemble forecasts can be of added value, when the difference between the probabilistic and deterministic forecasts is large.

Acknowledgements: This was SMHI’s last contribution as guest columnist team to the HEPEX blog for 2016. I am very grateful to the HEPEX co-chairs and community for giving us the opportunity to share insights and challenges in operational forecasting. I would also like to thank my colleagues Göran Lindström, David Gustafsson, Chantal Donnelly and Jonas Olsson for acting as guest authors. I hope the community enjoyed reading our contributions.