Parallelizing SWAT Calibration in Windows using SUFI2 Program,

Transcription

1 Parallelizing SWAT Calibration in Windows using SUFI2 Program, Lessons learned from Black Sea Catchment study Elham Rouholahnejad 2012 International SWAT Conference July 16-20, 2012, Delhi, India

2 Content Introduction to the Black sea project BSC SWAT model - model set up - complication regarding river flows Calibration - Claibration Set up - Iteration 1 (rough decisions) - Iteration 2 a) parallelization b) parameterization

3 I. Black Sea Catchment

4 Black Sea

5 Black Sea Project, EnviroGrids The project aims at building: A high resolution calibrated water quantity and water quality model for the entire region accounting for agricultural activities and point source pollutions using SWAT.

6 Black Sea Model o Major challenges in calibration process o Lack of sufficient data o Large scale application o Calibration run time o Water managements in the basin o Inaccurate data (gauge locations) o Flow directions o Glacier and snow parameters

7 II. SWAT model of the Black Sea basin

8 Black Sea SWAT Model set up Arc SWAT was used to parametraize the whole area 2 Mkm2 drainage area subbasins Hrus About 600,000 SWAT input files CRU data sets were used as weather data Agricultural management for three major crops: wheat, Maize and Barely. ET Calculation based on Hargreaves Method Daily step SWAT run 36 yrs of simulation, 3 yrs warm up period ( )

9 Basin, Subbasin, Hrus Black Sea SWAT Model

10 Black Sea SWAT Model

11 Complication in flow directions BSC SWAT Model, Complications

12 Imprisice Basin mask or wrong rivers! BSC SWAT Model, Complications

13 DEM BSC SWAT Model, Complications

14 DEM BSC SWAT Model, Complications

15 DEM BSC SWAT Model, Complications

16 III. Calibration Of the hydrologic model of the Black sea basin

17 o SWAT Cup, SUFI2 Calibration_ SUFI2 algorithm

18 SUFI-2 algorithm Uncertainty Attributed to parameters Obtained by calibrating a set of parameter ranges against observed data (priors posteriors) Characterised by 95 % prediction uncertainty (95PPU) Dual aim: Maximise P-factor Minimise R-factor Process for each calibration iteration Calibration_SUFI2 Latin hypercube sampling of parameter space, model simulations, objective function evaluation and parameter updating

19 Observation stations Calibration_SUFI2

20 Calibration_First Run Set up First run , 3 years warm up period (34 years) Discharge (monthly, 125 stations) Default parameter 1 run- 37 hours

21 Observation station check_deleted Calibration_First Run

22 Observation station check_deleted Calibration_First Run

23 Observation station check_deleted Calibration_First Run

24 Calibration_First Run Observation station check_ kept to apply Elev band

25 Observation station check_deleted Calibration_First Run

26 Observation station check_dislocation Calibration_First Run

27 A few good satations Calibration_First Run

28 A few good satations Calibration_First Run

29 Calibration Set up_ Iteration 1 Iteration 1 Calibration and evaluation , 3 years warm up period, 34 years calibration Discharge (monthly, 87) Parameter selected based on sensitivity analysis, experience and litrature 23 parameters sensetive to flow_ global parameters, not paramerizd yet Initial parameter ranges from physically meaningful limits New ranges based on parameterisation with highest objective function 200 runs (Estimated run time in a single machine 350 days!!

30 Calibration_SUFI2 paralelization Running time issue Copy all files in TxtInOut to BACKUP directory SUFI2_LH_sample.exe par_inf.txt par_val.txt SUFI2_new_pars.exe BACKUP SWAT_Edit.exe no. of parallel processes Modified SWAT inputs swat2009.exe SWAT outputs SUFI2_extract_*.def SUFI2_extract_*.exe *.out observed.txt *.out SUFI2_goal_fn.exe SUFI2_95ppu.exe goal.txt 95ppu.txt Is calibration criteria satisfied? no yes stop

31 o parallel Sufi2 Spead up tested in Different projects Calibration_SUFI2 paralelization Speedup Server1 Ideal Danube Alberta Test Speedup PC 1 Ideal Danube Alberta Test Number of processors Number of processors Speedup PC 2 Ideal Alberta Test Speedup PC 3 Ideal Alberta Test Number of processors Number of processors Speedup Laptop 1 Ideal Alberta Test Speedup laptop 2 Ideal Danube Alberta Test Number of processors Number of processors More details: E. Rouholahnejada, K.C. Abbaspoura, M. Vejdanib, R. Srinivasanc, R. Schulind, A. Lehmanne: A parallelization framework for calibration of hydrological models Environmental Modelling and Software, 2011.

32 o parallel Sufi2 Efficiency tested in Different projects Calibration_SUFI2 paralelization Efficiency (%) Efficiency (%) Efficiency (%) Ideal Danube Alberta Test Server Number of processors Ideal Alberta Test Ideal Laptop 1 Alberta Test PC Number of processors Number of processors Efficiency (%) Efficiency (%) Efficiency (%) Ideal Ideal Laptop 2 Alberta Test PC 1 Danube Alberta Test Number of processors Ideal Alberta Test PC Number of processors Number of processors More details: E. Rouholahnejada, K.C. Abbaspoura, M. Vejdanib, R. Srinivasanc, R. Schulind, A. Lehmanne: A parallelization framework for calibration of hydrological models Environmental Modelling and Software, 2011.

33 o parallel Sufi2 Efficiency tested in Different projects Calibration_SUFI2 paralelization 100 Server PC 1 CPU Usage (%) Danube Alberta Test Number of processors CPU Usage (%) Danube Alberta Test Number of processors CPU Usage (%) PC 2 Alberta Test Number of processors CPU Usage (%) PC 3 Alberta Test Number of processors 120 Laptop Laptop 2 CPU Usage (%) Alberta Test Number of processors CPU Usage (%) Alberta Test Number of processors More details: E. Rouholahnejada, K.C. Abbaspoura, M. Vejdanib, R. Srinivasanc, R. Schulind, A. Lehmanne: A parallelization framework for calibration of hydrological models Environmental Modelling and Software, 2011.

34 CERN super computer center Calibration_Gridification

35 Gridification on CERN Grids Calibration_Gridification 1th Parameter Set 2nd Parameter Set CERN Center SWAT RUN 3rd Parameter Set... nth Parameter Set.. SWAT RUN SWAT RUN SWAT RUN Output Collection

36 Calibration_Gridification Gridification scheme 200 runs_estimated run time in a single machine 350 days!! 200 runs using the parallel processing and Grid infrastracture 8days!! (4 blocks of 50 worker node)

37 Snow melt, glaciers, elevation band Calibration_Parameterization

38 Calibration_Parameterization Model consistently over pridicts the flow High Surface flow Solutions - Curve number for different land uses- decrease (CN in *.mgt) - Soil available water- increase (SOL_AWC in *.sol) - Soil evaporation compensation factor- increase up to 1.0 (ESCO in *.sub

39 Calibration_Parameterization Model consistently over pridicts the flow High base flow Too little evapotranspiration Solutions Increase deep percolation loss (adjust threshold depth of water in shallow aquifer required for the base flow to occur) (max 100 mm, GWQMN in *.gw) Increase grouindwater revap coefficient (max of 0.4, GW_Revap in *in.gw) Decrease threshold depth of water in shallow aquifer for revap tp occur (min of 0.0, REVAPMN in.gw)

40 Calibration_Parameterization Simulated flow follows the observed pattern but lags the actual flow consistently Time of concentration is too long Less than actual slope for overland flow Over estimated surface roughness Snow melt parameters Flood routing coefficients Solutions: Increase slope (up to 20%) for overland flow (SLOPE) Lower Manning`s roughness coefficient (OV_N) Lower the value of overland flow length to 4-10 m, if necessary (SLSUBBSN)

41 Calibration_Parameterization New parameter set

42 What comes next? Once the model is calibrated, the impcat of climate change and landuse change on water resources will be evaluated.

43 Acknowledgments This project has been funded by the European Commission s Seventh Research Framework through the envirogrids project (Grant Agreement n ).