Model Calibration

In this study a combined approach of manual calibration and autocalibration has been chosen. Therefore the available discharge record has been used for a split sample test: The observed inflows to the reservoir (1995-2007) have been used to calibrate the model. This period has been chosen, due to three reasons. The calibration period is dominated by intermediate flows, wet and dry years, therefore all relevant hydrological processes can be assumed to occur in the calibration period. In contrast the mid-eighties are very dry, exhibiting very little discharge at all and the late eighties/early nineties are very wet, exhibiting frequent floods. Furthermore the irrigated areas used within the CropWat-module represent the mid-nineties conditions. Finally snow dynamics have only been studied since the year 2000 within the IMPETUS project.

5.4.1 Hydrological Model

During the first manual calibration phase, the parameters presented in Table 5-16 have been adapted and a rough match of model results with the observed seasonal flow, observed snow cover and applied irrigation water was obtained. Besides the matching with flow records as presented in Figure 5-23 and Figure 5-24, irrigation records and snow cover measurements have been used to calibrate the snow parameters and the irrigation parameter IRR_FRAC.

To compare innerannual dynamics of the

snow cover, TERRA-MODIS (Moderate

Resolution Imaging Spectroradiometer) image products (MOD09 GHK) have been used by Schulz (2007). The Normalized Difference Snow Index (NDSI; Salomonson & Appel 2004) classification method was applied to a time series of 500 images, with about 100 for each snow cover period for the years 2001 to 2006 that last from October to June. Analysis of the NDSI data shows that snow cover is present for only short periods in altitudes below 2000 masl. In elevations above 3000 masl snow cover is present from October through May (~ 150 days), exhibiting two distinct peaks in the early and the late rainy season. The annual variability of snow cover days is below 20% in altitudes above 3,000 masl, but rises with decreasing altitude towards the basin (Klose et al. 2010a).

Table 5-16: Parameters adapted during

manual calibration (seeAppendix 3 for details on parameters).

Parameter Value GW_DELAY [days] 90 SFTMP [°C] 1 SMTMP [°C] 3 SMFMX [mm/°C*Day] 5 SNOCOVMX [mm] 80 SNO50COV [%] 20 IRR_FRAC [%] 50

Figure 5-21: Snow dynamics in the higher elevations of the Upper Drâa catchment 2001- 2006. Measured: MODIS-based NDSI with courtesy of Oliver Schulz; Simulated: SWAT-MAROC

Figure 5-21 shows the comparison of measured and simulated snow covers for the period 2001-2006. Though absolute coverage differs in the elevated regions, the temporal extents of measured and simulated snow cover are similar in all elevation zones. Overestimations of absolute cover may derive from the simple degree-day-approach, which does not account for slope, exposition, wind deposition and other factors. Furthermore SWAT-MAROC uses only lumped parameters for snow-related processes; therefore it does not reflect topographic differences between high-montane and intermediate regions. For the future it should be tested whether the lumped snow module parameterization could be improved by spatially distributed parameters, for instance by relating snow cover thresholds to topographic features such as slope or terrain roughness that could easily be derived from the DEM.

However SWAT-MAROC modeling of interseasonal snow-melt is imperfect, as the distinct peaks in December and March/ snow melt in February, as identified in the satellite imagery are not reproduced by the model. The elevation zones 4 and 5 (2500- 3000 masl and >3000 masl) represent only 1% and 8% of the catchment but roughly half the snow falls in these zones. The other half falls in elevation zone 2 and 3 (1500-2000 masl and 2000-2500 masl), which exhibit lower snow depths but a larger areal extent. In the latter two

zones snow dynamics are adequately covered by the model. The model also represents the two peaks of the snow cover in the elevated regions badly.

Concerning irrigation, a good match of the seasonal cycle, as well as the total volume of measured and simulated irrigation has been obtained for the hydrological year 1995/96, which is the only monthly irrigation time series available (Figure 5-22).

Figure 5-22: Monthly irrigation amounts for the Upper Drâa in 1995/96. (Measured data taken from DRPE 1998)

In the autocalibration phase, the SUFI-2 algorithm has been used. According to Refsgaard (1997) the number of real calibration parameters should be kept low, hence only three parameters have been chosen for calibration. All three are explicitly intended for calibration and have no or little physical meaning (Table 5-17).

Table 5-17: Parameters adapted during autocalibration (seeAppendix 3 for details on parameters).

Parameter Initial range Calibrated Value

GW_REVAP[%] 0 – 1 0. 47

ESCO[-] 0 – 1 0.78

CN2[-] -3 – 3 -0.3

To assess the model goodness at the reservoir measured inflow (calculated from the reservoir stage as presented in section 5.3.7.2) has been compared to simulated inflow. On an annual basis the dimensionless criteria are within an acceptable range: Nash-Sutcliffe Coefficient of Model Efficiency (CME) is 0.81 and the Index of Agreement (IoA) is 0.95. Absolute error indices do well likewise: The Root Mean Square Error (RMSE) is 79.36 Mm³/year (≙ 41% of measured values SD (Standard Deviation)) and the Mean Absolute Error (MAE) is 57.72 Mm³/year (≙ 33% SD) as displayed in Figure 5-23. During the calibration period goodness of fit criteria are within a good range for monthly values as well: (CME: 0.83; IoA: 0.96; RMSE: 13.89 Mm³/month (≙42% SD); MAE: 8.44 Mm³/month (≙26% SD); Figure 5-24).

Figure 5-23: Simulated and measured annual discharge into the reservoir Mansour-Eddahbi (1978-2007) (14,988 km²). (Measured data: SE Ouarzazate)

Figure 5-24: Simulated and measured monthly discharge into the reservoir Mansour-Eddahbi (1978-2007) (14,988 km²). (Measured data: SE Ouarzazate)

Within the model structure developed and using the input data available, calibration results can hardly be improved. As indicated in Figure 5-23 and Figure 5-24 no general bias can be detected, peak discharge as well as recession periods are overestimated and underestimated to the same extent. Deviations are therefore most likely results of measurement errors (climate or discharge), or incorrect process representation by the model. Further parameter adjustments would imply the trial to compensate for errors in the model or in the data (Abbott & Refsgaard 1996), and should therefore be avoided.

5.4.2 Reservoir Water Balance

The main purpose of reservoirs water balance module is the use within the scenarios, it has been driven with simulated discharge data. The module has been calibrated and validated for the same years as the hydrological model. Adapted parameters are the residual volume of the reservoir (50 Mm³) and the buffer coefficient that governs water releases in low storage periods (0.5).

Figure 5-25: Simulated and measured final annual stored volume and water release of the reservoir Mansour-Eddahbi (1973-2007) (14,988 km²). (Measured data: SE Ouarzazate)

The simulated reservoir volume and water release are in synchronicity with measured values, as indicated by high r² values (0.80 for the reservoir filling level and 0.65 for the water release). MAE is well below 50% of the measured values standard deviation (38% for the reservoir filling level and 41% for the water release); hence calibration results can be considered good. Nevertheless an additive error can be stated, as the modeled reservoir storage is below the measured one throughout the 90ies (see Figure 5-25). As outlined in section 5.1.2.4 evaporation losses are subtracted from the reservoir at the end of the calculation step, therefore the volume at the end of the year can never equal the capacity of the reservoir.