1. McGill University
Montreal, Quebec, Canada
Brace Centre for Water Resources Management
Global Environmental and Climate Change Centre
Department of Civil Engineering and Applied Mechanics
School of Environment
A SPATIAL-TEMPORAL DOWNSCALING APPROACH TO CONSTRUCTION OF INTENSITY-DURATION-FREQUENCY RELATIONS IN CONSIDERATION OF GCM-BASED CLIMATE CHANGE SCENARIOS
Van-Thanh-Van Nguyen (and Students)
Endowed Brace Professor Chair in Civil Engineering

2. OUTLINE
INTRODUCTION
Design Rainfall and Design Storm Concept – Current Practices
Extreme Rainfall Estimation Issues?
Climate Variability and Climate Change Impacts?
OBJECTIVES
DOWNSCALING METHODS
Spatial Downscaling Issues
Temporal Downscaling Issues
Spatial-Temporal Downscaling Method
APPLICATIONS
CONCLUSIONS

3. INTRODUCTION
Extreme storms (and floods) account for more losses than any other natural disaster (both in terms of loss of lives and economic costs).
Damages due to Saguenay flood in Quebec (Canada) in 1996: $800 million dollars.
Average annual flood damages in the U.S. are US$2.1 billion dollars. (US NRC)
Information on extreme rainfalls is essential for planning, design, and management of various water-resource systems.
Design Rainfall = maximum amount of precipitation at a given site for a specified duration and return period.

4. Design Rainfall Estimation Methods
The choice of an estimation method depends on the availability of historical data:
Gaged Sites  Sufficient long historical records (> 20 years?)  At-site Methods.
Partially-Gaged Sites  Limited data records  Regionalization Methods.
Ungaged Sites  Data are not available  Regionalization Methods.

5. Design Rainfall and Design Storm Estimation
At-site Frequency Analysis of Precipitation
Regional Frequency Analysis of Precipitation
⇒ Intensity-Duration-Frequency (IDF) Relations
⇒ DESIGN STORM CONCEPT for design of hydraulic structures
(WMO Guides to Hydrological Practices: 1st Edition → 6th Edition: Section 5.7, in press)

6. Extreme Rainfall Estimation Issues (1)
Current practices:
At-site Estimation Methods (for gaged sites): Annual maximum series (AMS) using 2-parameter Gumbel/Ordinary moments method, or using 3-parameter GEV/ L-moments method.
⇒ Which probability distribution?
⇒ Which estimation method?
⇒ How to assess model adequacy? Best-fit distribution?
Problems: Uncertainties in Data, Model and Estimation Method

7. Extreme Rainfall Estimation Issues (2)
Regionalization methods
GEV/Index-flood method.
Index-Flood Method (Dalrymple, 1960):
Similarity (or homogeneity) of point rainfalls?
How to define groups of homogeneous gages? What are the classification criteria? 4 3 2 1
Geographically contiguous fixed regions
Geographically non contiguous fixed regions
Hydrologic neighborhood type regions
Proposed Regional Homogeneity:
PCA of rainfall amounts at different sites for different time scales.
PCA of rainfall occurrences at different sites.
(WMO Guides to Hydrological Practices: 1st Edition → 6th Edition: Section 5.7, in press)

8. Extreme Rainfall Estimation Issues (3)
The “scale” problem
The properties of a variable depend on the scale of measurement or observation.
Are there scale-invariance properties? And how to determine these scaling properties?
Existing methods are limited to the specific time scale associated with the data used.
Existing methods cannot take into account the properties of the physical process over different scales.

9. Extreme Rainfall Estimation Issues (4)
Climate Variability and Change will have important impacts on the hydrologic cycle, and in particular the precipitation process!
How to quantify Climate Change?
General Circulation Models (GCMs):
A credible simulation of the “average” “large-scale” seasonal distribution of atmospheric pressure, temperature, and circulation. (AMIP 1 Project, 31 modeling groups)
Climate change simulations from GCMs are “inadequate” for impact studies on regional scales:
Spatial resolution ~ 50,000 km2
Temporal resolution ~ (daily), month, seasonal
Reliability of some GCM output variables (such as cloudiness  precipitation)?

10. …
How to develop Climate Change scenarios for impacts studies in hydrology?
Spatial scale ~ a few km2 to several 1000 km2
Temporal scale ~ minutes to years
A scale mismatch between the information that GCM can confidently provide and scales required by impacts studies.
“Downscaling methods” are necessary!!!
GCM Climate Simulations
Precipitation (Extremes) at a Local Site

11. IDF Relations At-site Frequency Analysis of Precipitation
Regional Frequency Analysis of Precipitation
⇒ Intensity-Duration-Frequency (IDF) Relations
⇒ DESIGN STORM for design of hydraulic structures.
Traditional IDF estimation methods:
Time scaling problem: no consideration of rainfall properties at different time scales;
Spatial scaling problem: results limited to data availability at a local site;
Climate change: no consideration.

12. Summary Recent developments:
Successful applications of the scale invariant concept in precipitation modeling to permit statistical inference of precipitation properties between various durations.
Global climate models (GCMs) could reasonably simulate some climate variables for current period and could provide various climate change scenarios for future periods.
Various spatial downscaling methods have been developed to provide the linkage between (GCM) large-scale data and local scale data.
Scale Issues:
GCMs produce data over global spatial scales (hundreds of kilometres) which are very coarse for water resources and hydrology applications at point or local scale.
GCMs produce data at daily temporal scale, while many applications require data at sub-daily scales (hourly, 15 minutes, …).

13. OBJECTIVES
To review recent progress in downscaling methods from both theoretical and practical viewpoints.
To assess the performance of statistical downscaling methods to find the “best” method in the simulation of daily precipitation time series for climate change impact studies.
To develop an approach that could link daily simulated climate variables from GCMs to sub-daily precipitation characteristics at a regional or local scale (a spatial-temporal downscaling method).
To assess the climate change impacts on the extreme rainfall processes at a regional or local scale.

14. DOWNSCALING METHODS
Scenarios

15. (SPATIAL) DYNAMIC DOWNSCALING METHODS
Coarse GCM + High resolution AGCM
Variable resolution GCM (high resolution over the area of interest)
GCM + RCM or LAM (Nested Modeling Approach)
More accurate downscaled results as compared to the use of GCM outputs alone.
Spatial scales for RCM results ~ 20 to 50 km  still larges for many hydrologic models.
Considerable computing resource requirement.

16. (SPATIAL) STATISTICAL DOWNSCALING METHODS
Weather Typing or Classification
Generation daily weather series at a local site.
Classification schemes are somewhat subjective.
Stochastic Weather Generators
Generation of realistic statistical properties of daily weather series at a local site.
Inexpensive computing resources
Climate change scenarios based on results predicted by GCM (unreliable for precipitation)
Regression-Based Approaches
Results limited to local climatic conditions.
Long series of historical data needed.
Large-scale and local-scale parameter relations remain valid for future climate conditions.
Simple computational requirements.

17. APPLICATIONS
LARS-WG Stochastic Weather Generator (Semenov et al., 1998)
Generation of synthetic series of daily weather data at a local site (daily precipitation, maximum and minimum temperature, and daily solar radiation)
Procedure:
Use semi-empirical probability distributions to describe the state of a day (wet or dry).
Use semi-empirical distributions for precipitation amounts (parameters estimated for each month).
Use normal distributions for daily minimum and maximum temperatures. These distributions are conditioned on the wet/dry status of the day. Constant Lag-1 autocorrelation and cross-correlation are assumed.
Use semi-empirical distribution for daily solar radiation. This distribution is conditioned on the wet/dry status of the day. Constant Lag-1 autocorrelation is assumed.

18. Statistical Downscaling Model (SDSM) (Wilby et al., 2001)
Generation of synthetic series of daily weather data at a local site based on empirical relationships between local-scale predictands (daily temperature and precipitation) and large-scale predictors (atmospheric variables)
Procedure:
Identify large-scale predictors (X) that could control the local parameters (Y).
Find a statistical relationship between X and Y.
Validate the relationship with independent data.
Generate Y using values of X from GCM data.

19. Geographical locations of sites under study.
Geographical coordinates of the stations

20. DATA:
Observed daily precipitation and temperature extremes at four sites in the Greater Montreal Region (Quebec, Canada) for the period.
NCEP re-analysis daily data for the period.
Calibration: ; validation:

21. Evaluation indices and statisticsNo
Code
Unit
Time scale
Description 1
Prcp1 %
Season
Percentage of wet days (daily precipitation  1 mm) 2
SDII
mm/r.day
Daily Mean: sum of daily precipitations / number of wet days 3
CDD
days
Maximum number of consecutive dry days (daily precipitation < 1 mm) 4
R3days mm
Maximum 3-day precipitation total 5
Prec90p
90th percentile of daily precipitation amount 6
Precip_mean
mm/day
Month
Sum of daily precipitation in a month / number of days in that month 7
Precip_sd
Standard deviation of daily precipitation in a month
Evaluation indices and statistics

22. The mean of daily precipitation for the period of 1961-1975
BIAS = Mean (Obs.) – Mean (Est.)

23. The mean of daily precipitation for the period of 1976-1990
BIAS = Mean (Obs.) – Mean (Est.)

24. The 90th percentile of daily precipitation for the period of 1976-1990
BIAS = Mean (Obs.) – Mean (Est.)

25. From CCAF Project Report by Gachon et al. (2005)
GCM and Downscaling Results (Precipitation Extremes )
1- Observed
2- SDSM [CGCM1]
3- SDSM [HADCM3]
4- CGCM1-Raw data
5- HADCM3-Raw data
From CCAF Project Report by Gachon et al. (2005)

26. SUMMARY Downscaling is necessary!!!
LARS-WG and SDSM models could provide “good” but generally “biased” estimates of the observed statistics of daily precipitation at a local site.
GCM-Simulated Daily Precipitation Series
Is it feasible?
Daily and Sub-Daily Extreme Precipitations

27. The Scaling Concept

28. The Scaling Generalized Extreme-Value (GEV) Distribution.
The scaling concept
The cumulative distribution function:
The quantile:

29. The Scaling GEV Distribution

30. The first three moments of GEV distribution:

31. APPLICATION: Estimation of Extreme Rainfalls for Gaged Sites
Data used:
Raingage network: 88 stations in Quebec (Canada).
Rainfall durations: from 5 minutes to 1 day.
Record lengths: from 15 yrs. to 48 yrs.

32. Scaling of NCMs of extreme rainfalls with durations: 5-min to 1-hour and 1-hour to 1-day.
red: 1st NCM; blue: 2nd NCM; black: 3rd NCM; markers: observed values; lines: fitted regression

33. Results on scaling regimes:
Non-central moments are scaling.
Two scaling regimes:
5-min. to 1-hour interval.
1-hour to 1-day interval.
Based on these results, two estimations were made:
5-min. extreme rainfalls from 1-hr rainfalls.
1-hr. extreme rainfalls from 1-day rainfalls.

34. 5-min Extreme Rainfalls estimated from 1-hour Extreme Rainfalls
markers: observed values – lines: values estimated by scaling method
markers: observed values – lines: values estimated by scaling method

35. markers: observed values – lines: values estimated by scaling method
1-hour Extreme Rainfalls estimated from 1-day Extreme Rainfalls
markers: observed values – lines: values estimated by scaling method

36. The Spatial-Temporal Downscaling Approach
GCMs: HadCM3 and CGCM2.
NCEP Re-analysis data.
Spatial downscaling method: the statistical downscaling model SDSM (Wilby et al., 2002).
Temporal downscaling method: the scaling GEV model (Nguyen et al. 2002).

37. The Spatial-Temporal Downscaling Approach
Spatial downscaling:
calibrating and validating the SDSM in order to link the atmospheric variables (predictors) at daily scale (GCM outputs) with observed daily precipitations at a local site (predictand);
extracting AMP from the SDSM-generated daily precipitation time series; and
making a bias-correction adjustment to reduce the difference in quantile estimates from SDSM-generated AMPs and from observed AMPs at a local site using a second-order nonlinear function.
Temporal downscaling:
investigating the scale invariant property of observed AMPs at a local site; and
determining the linkage between daily AMPs with sub-daily AMPs.

38. Application Study Region Data
Precipitation records from a network of 15 raingages in Quebec (Canada).
Data
GCM outputs:
HadCM3A2, HadCM3B2,
CGMC2A2, CGCM2B2,
Periods: , 2020s, 2050s, 2080s.
Observed data:
Daily precipitation data,
AMP for 5 min., 15 min., 30 min., 1hr., 2 hrs., 6 hrs., 12 hrs.
Periods:

39. Daily AMPs estimated from GCMs versus observed daily AMPs at Dorval.
Calibration period:
CGCMA2
HadCM3A2

40. Residual = Daily AMP (GCM) - Observed daily AMP (local)
Calibration period:
CGCMA2
HadCM3A2

41. Daily AMPs estimated from GCMs versus observed daily AMPs at Dorval.
Validation period:
CGCMA2
HadCM3A2
Adjusted Daily AMP (GCM) = Daily AMP (GCM) + Residual

42. CGCMA2
HadCM3A2

43. CONCLUSIONS (1)
Significant advances have been achieved regarding the global climate modeling. However, GCM outputs are still not appropriate for assessing climate change impacts on the hydrologic cycle.
Downscaling methods provide useful tools for this assessment.
Calibration of the SDSM suggested that precipitation was mainly related to zonal velocities, meridional velocities, specific humidities, geopotential height, and vorticity.
In general, LARS-WG and SDSM models could provide “good” but “biased” estimates of the observed statistical properties of the daily precipitation process at a local site.

44. CONCLUSIONS (2)
It is feasible to link daily GCM-simulated climate variables with sub-daily AMPs based on the proposed spatial-temporal downscaling method. ⇒ IDF relations for different climate change scenarios could be constructed.
Differences between quantile estimates from observed daily AMPs and from GCM-based daily AMPs could be described by a second-order non-linear function.
Observed AMPs in Quebec exhibit two different scaling regimes for time scales ranging from 1 day to 1 hour, and from 1 hour to 5 minutes.
The proposed scaling GEV method could provide accurate AMP quantiles for sub-daily durations from daily AMPs.
AMPs derived from CGCM2A2 outputs show a large increasing trend for future periods, while those given by HadCM3A2 did NOT exhibit a large (increasing or decreasing) trend.

45. Thank you
for
your attention!

46. Slides required for presentations

49. DESIGN STORM CONCEPT Watershed as a linear system
Stormwater removal  Qpeak  Rational Method: Qpeak = CIA  Uniform Design Rainfall
Watershed as a nonlinear system.
Environmental control  Entire Hydrograph Q(t)  More realistic temporal rainfall pattern (or Design Storm) for more realistic rainfall-runoff simulation.
A design storm describes completely the distribution of rainfall intensity during the storm duration for a given return period.

50. DESIGN STORM CONCEPT Two main types of “synthetic” design storms:
Design Storms derived from the IDF relationships.
Design Storms resulted from analysing and synthesising the characteristics of historical storm data.
A typical design storm:
Maximum Intensity: IMAX
Time to peak: Tb
Duration: T
Temporal pattern

51. Design Storm Estimation Issues
Different synthetic design storm models available in various countries:
US Chicago storm model (Keifer and Chu, 1957)
US Normalized storm pattern by Huff (1967)
Czechoslovakian storm pattern by Sifalda (1973)
Australian design storm by Pilgrim and Cordery (1975)
UK Mean symmetric pattern (Flood Studies Report, 1975)
French storm model by Desbordes (1978)
US storm pattern by Yen and Chow (1980)
Canadian Atmospheric Environment Service (1980)
US balanced storm model (Army Corps of Engineer, 1982)
Canadian temporal rainfall patterns (Nguyen, 1981,1984)
Canadian storm model by Watt et al. (1986)
No general agreement as to which temporal storm pattern should be used for a particular site ⇒ How to choose? How to compare?

52. Intensity-Duration-Frequency curves for Montreal area.

53. ⇓ ⇓
Chicago ⇒
IDF
Design Storm

54. Design Storm Patterns for southern Quebec (Canada)

55. Design Storm Patterns for southern Quebec (Canada)

56. SUMMARY Results indicated: For runoff peak flows: For runoff volumes:
the Canadian AES design storm
the Desbordes model (with a peak intensity duration of 30 minutes)
For runoff volumes:
the Canadian pattern proposed by Watt et al.
None of the eight design storms was able to provide accurate estimation of both runoff parameters.

57. The 1-hr optimal storm pattern for southern Quebec (Canada)

58. Assessment of the Proposed Optimal Storm Pattern
Probability distributions of runoff peak flows and volumes for a square basin of 1 ha
Similar results of probability distributions for all tested basins.

59. Assessment of the Proposed Optimal Storm Pattern

60. Climate Trends and Variability 1950-1998
Maximum and minimum temperatures have increased at similar rate
Warming in the south and west, and cooling in the northeast (winter & spring)
Trends in
Winter
Mean Temp
(°C / 49 years)
Trends in
Spring
Mean Temp
(°C / 49 years)
Trends in
Summer
Mean Temp
(°C / 49 years)
Trends in
Fall
Mean Temp
(°C / 49 years)
From X. Zhang, L. Vincent, B. Hogg and A. Niitsoo, Atmosphere-Ocean, 2000

61. Validation of GCMs for Current Period (1961-1990)
Winter Temperature (°C)
Model mean =all flux & non-flux corrected results (vs NCEP/NCAR dataset)
[Source: IPCC TAR, 2001, chap. 8]

62. Impact models require ... GCMs or RCMs supply...
Climate Scenario development need: from coarse to high resolution
A mismatch of scales between what climate models can supply and what environmental impact models require.
300km
Impact models require ...
50km
10km 1m
Point
GCMs or RCMs supply...
P. Gachon

63. Choice of distribution model for fitting annual extreme rainfalls
Common probability distributions:
Two-parameter distribution:
Gumbel distribution
Normal
Log-normal (2 parameters)
Three-parameter distributions:
Beta-K distribution
Beta-P distribution
Generalized Extreme Value distribution
Pearson Type 3 distribution
Log-Normal (3 parameters)
Log-Pearson Type 3 distribution

64. Choice of distribution model for fitting annual extreme rainfalls
Generalized Gamma distribution
Generalized Normal distribution
Generalized Pareto distribution
Four-parameter distribution
Two-component extreme value distribution
Five-parameter distribution:
Wakeby distribution
No general agreement on the choice of distribution for extreme rainfalls!!!

65. Choice of distribution model for fitting annual extreme rainfalls
A three-parameter distribution can provide sufficient flexibility for describing extreme hydrologic data.
A two-parameter distribution could be adequate for prediction.
The choice of a distribution is not as crucial as an adequate data sample. Discrepancies increase for extrapolation beyond the length of record (model error is more important than sampling error).

66. Estimation of model parameters
Graphical method (Probability plots)
Different plotting-position formulas
Frequency factor method
Method of moments
Sample mean, variance, and skewness.
Sample mean, variance, 1st and/or 2nd moments in log-space (method of mixed moments)
Sample mean, variance, and geometric and/or harmonic mean (generalized method of moments)
Should we use higher-order moments?

67. Estimation of model parameters
Method of maximum likelihood
Optimal estimators (unbiased, minimum variance) of the parameters.
Iterative numerical methods.
It could give bad estimators for small samples.
Method of L-moments
Linear combination of order statistics
Sample L-moments are found less biased than traditional moment estimators  better suited for use with small samples?
Other methods
Maximum entropy method
Etc.

68. MODEL ASSESSMENT Descriptive Ability Predictive Ability
Graphical Display: Quantile-Quantile Plots
Numerical Comparison Criteria
Predictive Ability
Bootstrap Method

69. Numerical Comparison Criteria
Root Mean Square Error
Relative Root Mean Square Error
Maximum Absolute Error
Correlation Coefficient

70. BOOTSTRAP METHOD
A nonparametric approach that repeatedly draws, with replacement,
n observations from the available data set of size N (N >n) and yields
multiple synthetic samples of the same sizes as the original observations.

71. Location of the 20 Climatological Stations
Record Length
Max: 52 yrs
Min: 24 yrs

72. Goodness-of-fit on the Right Tail
Quantile-Quantile Plots for the Distributions Fitted to
5-Minute Annual Precipitation Maxima at St-Georges Station
Fitted Precipitations (mm)
Observed Precipitation (mm)

73. Extrapolated Right-Tail Quantiles
Box Plots of Extrapolated Right-Tail Bootstrap Data for
5-Minute Annual Precipitation Maxima at McGill Station

74. Results for At-site Frequency Analysis of Extreme Rainfalls in Quebec
Comparable performance for all distributions in terms of Descriptive and Predictive abilities.
Top three distributions – WAK,GEV and GNO
Computational simplicity
GUM>GPA>BEP>BEK>GEV>GNO>PE3>WAK>LP3
Theoretical basis of GEV
⇒ GEV is recommended as the most suitable for representing annual maximum precipitation in Southern Quebec