Flow forecasting

In data-scarce environments and inaccessible mountainous terrain, like in the Chilean Andes, it is often difficult to install instrumentation and retrieve real-time physical data from these instruments. These real-time data can be useful to capture the hydroclimatic variability in this region, and improve the forecasting capability of hydrological models. Although statistical models can provide skillful seasonal forecasts, using large-scale climate variables and in situ data (Piechota and Chiew 1998; Grantz et al. 2005; Regonda et al. 2006; Bracken et al., 2010), a particular hydropower company in Chile was mainly interested in the potential use of an integrated system, using measurements derived from both Earth observation (EO) satellites and in situ sensors, to force a hydrological model to forecast seasonal streamflow during the snow melting season. The objective of the INTOGENER (INTegration of EO data and GNSS-R signals for ENERgy applications) project was therefore to demonstrate the operational forecasting capability of the SPHY model in data-scarce environments with large hydroclimatic variability.

During INTOGENER, data retrieved from EO satellites consisted of a DEM and a time series of snow cover maps. Snow cover images were retrieved on a weekly basis, using RADARSAT and MODIS (Parajka and Blöschl 2008; Hall et al. 2002) imagery. These images were used to update the snow storage (SS (mm)) in the model in order to initialize it for the forecasting period. Figure 10 shows the snow storage as simulated by the SPHY model during the snow melting season in the Laja basin. These maps clearly show the capability of SPHY to simulate the spatial variation of snow storage, with more snow on the higher elevations, and a decrease in snow storage throughout the melting season. Discharge, precipitation and temperature data were collected using in situ meteorological stations. In order to calculate the lake outflow accurately, the SPHY model was initialized with water level measurements retrieved from reflected Global Navigation Satellite System (GNSS) signals in Laja Lake. Static data that were used in the SPHY model consisted of soil characteristics derived from the Harmonized World Soil Database (HWSD) (Batjes et al. 2009) and land use data obtained from the GLOBCOVER (Bontemps et al. 2011) product. The SPHY model was set up to run at a spatial resolution of 200m.

Figure 11 shows the observed vs. simulated daily streamflow for two locations within the Laja River basin for the historical period 2007–2008. It can be seen that model performance is quite satisfactory for both locations, with volume errors of 4 and 9.4% for the Abanico Canal (downstream of Lake Laja) and Rio Laja en Tucapel, respectively. The NS coefficient, which is especially useful for assessing the simulation of high discharge peaks, is less satisfactory for these locations. Hydropower companies, however, have more interest in expected flow volumes for the coming weeks/months than in accurate day-to-day flow simulations, and therefore the NS coefficient is less important in this case. If the NS coefficient is calculated for the same period on a monthly basis, then the NS coefficients are 0.53 for the Abanico Canal and 0.81 for Rio Laja en Tucapel. It is likely that SPHY model performance would even have been better if a full model calibration would have been performed.

Figure 12: Snow storage (mm) as simulated by the SPHY model on 12 August (left) and 01 October (right) during the snow melting seasons of 2013 in the Laja River basin.

The hydropower company’s main interest is the model’s capacity to predict the total expected flow for the coming weeks during the melting season (October 2013 through March 2014). To forecast streamflow during the snow melting season, the SPHY model was forced with gridded temperature and precipitation data from the European Centre for Medium-range Weather Forecasts (ECMWF) Seasonal Forecasting System (SEAS) (Andersson 2013). The SEAS model provided daily forecasts at a spatial resolution of 0.75, 7 months ahead, and was used to forecast streamflow up till the end of the melting season. Figure 12 shows the bias between the total cumulative forecasted flow and observed flow for the 23 model runs that were executed during operational mode. Although there are some bias fluctuations in the Rio Laja en Tucapel model runs, it can be concluded that the bias decreases for each next model run for both locations, which is a logical result of a decreasing climate forcing uncertainty as the model progresses in time. It can be seen that the SPHY model streamflow forecasts for Canal Abanico, which is downstream of Laja Lake, are substantially better than for Rio Laja en Tucapel (the most downstream location). The reason for this has not been investigated during the demonstration study, but since model performance for these two locations was satisfactory during calibration, a plausible explanation could be the larger climate forecast uncertainty in the higher altitude areas (Hijmans et al. 2005; Rollenbeck and Bendix 2011; Vicuña et al., 2011; McPhee et al. 2010; Mendoza et al., 2012; Ragettli and Pellicciotti 2012; Ragettli et al. 2014) in the northeastern part of the basin that contributes to the streamflow of Rio Laja en Tucapel. Additionally, only two in situ meteorological stations were available during operational mode, whereas during calibration, 20 meteorological stations were available. Moreover, these operational meteorological stations were not installed at higher altitudes, where precipitation patterns tend to be spatially very variable (Wagner et al. 2012; Rollenbeck and Bendix 2011).

Figure 13: Daily observed vs. SPHY simulated streamflow (period 2007-2008) for the streamflow stations Canal Abanico (ID19) and Rio Laja en Tucapel (ID23). The Nash-Sutcliffe (NS) and bias model performance indicators are shown as well.

Figure 14: Bias between total cumulative forecasted flow and observed flow for the 23 model runs that were executed between the end of September 2013 and March 2014. Results are shown for the locations Canal Abanico (ID19) and Rio Laja en Tucapel (ID23).