Two-Step Calibration Method for SWAT

Transcription

1 SWAT 2005 Zurich, Switzerland July 14 th, 2005 Two-Step Calibration Method for SWAT Francisco Olivera, Ph.D. Assistant Professor Huidae Cho Graduate Student Department of Civil Engineering Texas A&M University College Station - Texas

2 Can we extract spatial SWAT 2005 Zurich, Switzerland July 14 th, 2005 information from temporal data? Francisco Olivera, Ph.D. Assistant Professor Huidae Cho Graduate Student Department of Civil Engineering Texas A&M University College Station - Texas

3 Objective Given: A calibration routine that adjusts the terrain parameters independently for each sub-basin and HRU (i.e., extracts spatial information). Hypothesis: A model that can reproduce the system s responses for a long period of time has the correct parameter spatial distribution.

4 Lake Lewisville Watershed Area: 2500 km 2 Four flow gauging stations: Six Four precipitation temperature one is an stations: inlet, one two stations: inside is used one the for inside watershed calibration, the and watershed two four are outside. and used three for (spatial) outside. validation. Period of record:

5 Lake Lewisville Watershed Curve number

6 Calibration and validation Three-year warm-up period. Calibration periods: One year: [ ] + [1995] Six years: [ ] + [ ] Validation period: Four years: [ ] + [ ] Calibration location Validation location

7 Calibration parameters Parameter Description CN2 SOL_AWC ESCO GWQMN GW_REVAP REVAPMN CH_K2 ALPHA_BF OV_N SLOPE SLSUBBSN SCS runoff curve number Soil available water capacity (mm H 2 O/mm soil) Soil evaporation compensation factor. Whenever the upper layer cannot meet the evaporative demand, SWAT extracts more soil water from the lower layer. Threshold depth of water in the shallow aquifer required for return flow to occur (mm H 2 O) Groundwater revap coefficient. Revap is the process by which water moves into the overlying unsaturated zone from the shallow aquifer by the capillary fringe or deep-rooted plants. Threshold depth of water in the shallow aquifer for revap or percolation to the deep aquifer to occur (mm H 2 O) Effective hydraulic conductivity in main channel alluvium (mm/hr) Baseflow alpha factor (days) is a direct index of groundwater flow response to changes in recharge. Manning s n value for overland flow Average slope steepness (m/m) Average slope length (m)

8 Calibration levels Watershed: All sub-basin and HRU parameters are adjusted by applying one single parameter-change rule over the entire watershed. Sub-basin: All sub-basin and HRU parameters are adjusted by applying a different parameter-change rule in each subbasin. Follows watershed calibration. HRU: Each HRU parameter is adjusted differently. Follows sub-basin calibration.

9 Parameter-change rules Method A (plus/minus): < α < 0.05 P = i 1 P + i (P + max P min ) α Method B (factor): 0.9 < α < 1.1 P = i 1 P + i (P + i P min ) α Method C (alpha): -0.5 < α < 0.5 P = i 1 P + i (P + max P i) α

10 Initial conditions Initial conditions for the iterative calibration process: Non-uniform: parameter values based on landuse and soil-type data. Uniform: average parameter values throughout the watershed let the calibration process extract the spatial information.

11 Objective functions Sum of the square of the residuals (SSR) = simulated SSR (Q Q ) observed Sum of the absolute value of the residuals (SAR) = simulated SSR Q Q observed No logq or Q in the denominator was used because some flows were zero. 2

12 Model efficiency Model efficiency was evaluated with the Nash- Sutcliffe coefficient: 2 (Qsimulated Q observed ) SSR NS = 1 = 1 (Q Q ) (Q Q ) observed 2 2 observed Q where is the long-term flow average (i.e., the predicted flows with no model ).

13 Hydrologic unit Calibration // SSR // Distributed // Plus-minus // 42 Nash-Sutcliffe coefficient Number of years used in calibration SGP The increase in NS between subbasin and watershed is small, and between HRU and subbasin is negligible.

14 Hydrologic unit Validation // SSR // Distributed // Plus-minus // 42 Nash-Sutcliffe coefficient Number of years used in calibration SGP The decrease in NS between subbasin and watershed is small, and between HRU and subbasin is negligible.

15 Objective function Calibration // Distributed // Plus-minus // 42 Nash-Sutcliffe coefficient SAR Nash-Sutcliffe coefficient Number of years used in calibration SGP Number of years used in calibration SSR SGP The increase in NS between subbasin and watershed is small, and between HRU and subbasin is negligible.

16 Objective function SSR Validation // Distributed // Plus-minus // 42 Nash-Sutcliffe coefficient Number of years used in calibration 0.4 SAR Nash-Sutcliffe coefficient The increase in NS between subbasin and watershed is small, and between HRU and subbasin is negligible. SGP Number of years used in calibration SGP

17 Parameter change function 1.0 Calibration // SSR // Distributed // 42 Nash-Sutcliffe coefficient Number of years used in calibration HGA WGA WGF SGF SGP HGF SGA The NS values for the factor parameter-change-function are slightly lower than for the other functions.

18 Parameter change function Validation // SSR // Distributed // 42 Nash-Sutcliffe coefficient Number of years used in calibration HGA WGA WGF SGF SGP HGF SGA

19 Spatial variability 1.0 Calibration // SSR // Plus-minus // 42 Nash-Sutcliffe coefficient Number of years used in calibration WAP SGP SAP HAP The NS values are not significantly affected by the initial assumed spatial variability.

20 Hydrographs - Calibration 250 Calibration // SSR // Plus-minus // 42 Flow (m 3 /s) WAP Observed 0 1/1/1990 1/1/1991 1/1/ /31/ /31/ /31/1994 The simulated hydrographs are fundamentally equal even though the initial conditions before the calibration were very different.

21 Parameter values 100 SSR // Average // Plus-minus // HRU Curve number Curve number (base) SSR // Distributed // Plus-minus // Watershed Same results were obtained from significantly different sets of initial parameters.

22 Spatial variability 1.0 Validation // SSR // Plus-minus // 42 Nash-Sutcliffe coefficient Number of years used in calibration WAP SGP SAP HAP The NS values are not significantly affected by the initial assumed spatial variability.

23 Hydrographs - Validation 160 Validation // SSR // Plus-minus // 42 Flow (m 3 /s) WAP Series1 0 1/1/ /31/ /31/ /31/1998 The simulated hydrographs are fundamentally equal even though the assumptions during calibration were very different.

24 Spatial validation Calibration // SSR // Plus-minus // 38 Nash-Sutcliffe coefficient Flow (m 3 /s) WAP SGP SAP HAP WAP Observed /1/1990 1/1/1991 1/1/ /31/ /31/ /31/1994 Number of years used in calibration The initial assumed spatial variability does not make a significant difference; however, the watershed-based calibration is more accurate.

25 Spatial and temporal validation Validation // SSR // Plus-minus // 38 Nash-Sutcliffe coefficient Flow (m 3 /s) /1/ /31/ /31/ /31/1998 Number of years used in calibration WAP SGP SAP HAP WAP Observed In validation over space and time, the watershed-based calibration with an average initial spatial variability has the highest NS values.

26 Spatial validation Calibration // SSR // Plus-minus // 37 Nash-Sutcliffe coefficient Flow (m 3 /s) WAP SGP SAP HAP WAP Observed /1/1990 1/1/1991 1/1/ /31/ /31/ /31/1994 Number of years used in calibration Not even the watershed-based calibration using the spatial variability defined by the soil and land use data produces a good NS value.

27 Spatial and temporal validation Validation // SSR // Plus-minus // 37 Nash-Sutcliffe coefficient Flow (m 3 /s) WAP SGP SAP HAP WAP Observed /1/ /31/ /31/ /31/1998 Number of years used in calibration Not even the watershed-based calibration using the spatial variability defined by the soil and land use data produces a good NS value.

28 Discussions Dispersion is the hydrodynamic process by which some water particles flow faster than others. Because of dispersion, Q it is difficult to know exactly when and where a particle entered the system. dispersion The effect of dispersion (i.e., response width) increases proportionally to the square root of the flow time; while the effect of advection (i.e., location of the response centroid) increases linearly with the flow time. t In small watersheds, dispersion mixes all responses (i.e., unit advection hydrograph); while in large watersheds, advection keeps responses separate (i.e., flow-time area diagrams).

29 Conclusions It was not possible to extract hydrologic information from temporal data for the 2,500-km 2 Lake Lewisville watershed. The effect of the spatial variability was small compared to the effect of hydrodynamic dispersive processes in the system. The number of years used for calibrating the model was fundamental for determining the parameter values. The parameter-change rule and the selected objective function did not significantly affect the calibration process.

30 Questions?