Dr. R N Sankhua

1
HYDROLOGICAL TIME SERIES GENERATION IN INDIAN
CONTEXT
Dr. R N Sankhua Director
National Water Academy, Central Water
Commission India
2
STATUS OF WATER RESOURCES IN INDIA

• India occupies only 3.29 million Ha area, (2.4)
  of world's land area.
• 4 of water resources of the World.
• supports over 16 of the world's population.
• livestock population 500 million, 15 of world's
  total

WR documentation of India at www.cwc.nic.in www.in
dia-water.com/ffs/index.htm
3
River Basins in India
4
(No Transcript)
5
SWOT analysis of Water Resources

• Strength
• India is gifted with large number of rivers
• 4000 BCM of water available
• Long-term average annual rainfall is 1160 mm,
  which is the highest anywhere in the world for a
  country of comparable size
• Annual precipitation of about 4000 BCM, including
  snowfall. monsoon rainfall 3000 BCM
• Highest rainfall (11,690 mm) recorded at
  Mousinram near Cherrapunji in Meghalaya in
  northeast
• Weakness
• Spatial and temporal distribution
• 690 BCM is utilizable form
• Storage insufficient to meet the demand
• Monsoon failure or excess rainfall in one monsoon

6
Event-Based Models

• RAINFALL
• Usually based on statistical analysis
• Sometimes, historical storm information used
• WATERSHED CHARACTERISTICS
• Relationship between rainfall and runoff
  identified (e.g. Rational Method C factor,
  Runoff CN).
• coefficients depend on soil infiltration rate,
  vegetation, land use, soil type, imperviousness,
  etc

7
Continuous Simulation Models

• Use long term rainfall record (20-30 years) and
  simulate flows for entire period of record
• Incorporate ET0 and infiltration estimates
  simulate water balance
• HEC-HMS, SWMM, SWAT, HYMOS, Arc-CN runoff for
  predicting variability in flow based on
  event/long term observed hydrologic data

8
Using HEC-HMS

• Three components
• Basin model - contains elements of basin,
  connectivity, runoff parameters
• Meteorologic Model - contains rainfall ET0 data
  
• Control Specifications - contains start/stop
  timing and calculation intervals for the run

9
Using SWMM
SWMM Visual Objects
- distributed, dynamic rainfall-runoff simulation
model used for single event or long-term
(continuous) simulation of runoff quantity and
quality from primarily urban areas.
10
Conventional Models of Synthetic Stream flow
generation

• AR (Auto Regression)
• AR(1) -1st order
• AR(2) -2nd order
• ARMA (Auto Regression Moving Average)
• ARIMA (Auto Regression Integrated Moving Average)
  - EVIEWS
• THOMAS-FIERRING MODEL
• All the models use the statistical properties of
  the inflow
• Used for monthly, seasonal annual inflow
  prediction

                                                                                                                                  
11

• All stationary time series can be modeled as AR
  or MA or ARMA models
• constant ? and ?2
• If a time series is not stationary it is often
  possible to make it stationary by using fairly
  simple transformations

Forecasts can be either in-sample or
out-of-sample forecasts.
12
Conventional Models Stream flow generation

• Periodical component,(parameters show variation)
  
• Trend component (increase or decrease of process
  parameters (mean std deviation) with time)
• Independent (random) components dependant
  components

13
AR Models of Synthetic Stream flow generation

• Produce sequences of streamflows at multi sites
  for low forecast horizon
• Synthetic streamflows must behave statistically
  similar to historical values and be consistent
  with seasonal volume forecasts

14
THOMAS-FIERRING MODEL PARAMETERS
Back
15
River Flow Forecasting - Krishna
Learning a model from existing data (e.g.
observations of the period from 1987 to 2000)
The resultant model will be used to forecast
future behaviour
16
(No Transcript)
17
Non-stationary Time series

• Linear trend
• Nonlinear trend
• Multiplicative seasonality
• Heteroscedastic error terms (non constant
  variance)

18
Making them stationary

• Linear trend
• Take non-seasonal difference. What is left over
  will be stationary AR, MA or ARMA
• Non-Linear trend
• Exponential growth
• Take logs this makes the trend linear
• Take non-seasonal difference

19
Multiplicative seasonality Heteroscedsatic
errors

• Taking logs
• Multiplicative seasonality often occurs when
  growth is exponential.
• Take logs then a seasonal difference to remove
  trend

20
Soft techniques for Synthetic Streamflow
generation
Neural Network

• Using ANN technique
• Using daily flow data
• Training of network
• Validating network
• Predicting flow

Fuzzy logic
ANFIS
21
Input Layer
Hidden Layer
Output Layer
REF-ET in mm/day
ANN model developed for predicting daily Ref-ETr
22
Stage- Discharge (DIBRUGARH- Brahmaputra river)
Nash 0.9864 RMSE 0.816
1-4-1
23
Model Parameters (DIBRUGARH)
Training Target Output AE ARE
Mean 7.324 7.307 1.931 0.460
Std Dev 6.460 6.069 1.649 0.501
Min 0.928 1.958 0.026 0.004
Max 25.025 25.15 7.149 3.326
C 0.92
Validation Target Output AE ARE
Mean 10.194 10.63 1.823 0.480
Std Dev 7.462 6.313 2.345 1.014
Min 1.242 1.960 0.165 0.008
Max 25.36 25.15 9.102 4.44
C 0.92
Testing Target Output AE ARE
Mean 7.015 7.075 3.267 0.688
Std Dev 6.655 5.256 2.880 0.661
Min 1.448 1.956 0.481 0.072
Max 20.55 18.88 8.7225 2.22
C 0.85
C Correlation coefficient, AE Absolute Error
(Target value - desired value), ARE Absolute
Relative Error (Target value - desired
value)/(Target value)
24
MODELS FOR REAL-TIME STAGE FORECAST
Pancharatna
Brahmaputra
Pandu
Station Input data Desired output data Travel time
Pandu (t)th day stages (Pandu) (t-1)th day stages (Pandu) (t1)th day Stages (Pandu)
Pancharatna (t)th day stages (Pandu) (t)th day stages (Pancharatna) (t1)th day Stages (Pancharatna) 1 day (from Pandu)
Dhubri (t-1)th day stages (Pandu) (t)th day stages (Pancharatna) (t)th day stages (Dhubri) (t1)th day Stages (Dhubri) 1 day (from Pancharatna)
25
ARCHITECTURE OF MODELS FOR STAGE FORECAST
Neural Network model for Pandu
Neural Network model for Pancharatna
Neural Network model for Dhubri
26
REAL-TIME STAGE FORECAST (PANDU)
2002
27
Using Fuzzy logic Data-1998 to 2002
Stage-Discharge (h Q) Nine Gaussian MF - IP/OP
Pancharatna
Pandu
28
FUZZY LOGIC (MFS VALIDATION)
29
FUZZY LOGIC (RULE VIEWER)
30
(No Transcript)
31
Regression eqn between gauged ungauged locations
Digital Precipitation Model
calculated relationship P115.60.258h
Conclusion Data driven modelling, coupled with
physical insights about the system, will produce
more reliable results for medium-and long-term
predictions.
32
THANK YOU