Você está na página 1de 32

Stochastic Analysis, Modeling and

Simulation (SAMS)
Anlisis Estocstico, Modelamiento y Simulacin

Stochastic
Stochastikos: Palabra Griega que hace
referencia a algo que contiene una
variable aleatoria.
En hidrologa se usa para describir un
sistema que tiene en el un elemento
aleatorio.

1. Introduccin
SAMS es un programa de cmputo, escrito en FORTRAN y C,
que permite analizar, modelar y simular las series de tiempo
hidrolgicas. Las principales capacidades y limitaciones son:
Analiza datos anuales y estacionales (hasta 12 estaciones o meses)
Incluye diferentes tipos de normalizacin o transformacin de datos
Permite trabajar con hasta 40 estaciones y para propsitos de
desagregacin multivariado con grupos de 10 estaciones
Incluye esquemas para modelar sistemas complejos de redes
Mximo nmero de aos de datos de entrada 600
El nmero de nuestras a generarse es ilimitado
El nmero de aos a generarse es ilimitado

2. Descripcin del SAMS


El SAMS consta de tres mdulos bsicos de aplicacin:
Anlisis Estadstico de Datos
Definicin del Modelo Estocstico
Generacin de Series Sintticas

2.1. Anlisis Estadstico de Datos


Este mdulo permite:
Plotear los datos como serie de tiempo
Verificar la normalidad y Transformar los datos
Determinar las caractersticas estadsticas de la serie

Anlisis Estadstico de Datos


Ploteo de datos

El grfico de serie de tiempo permite observar y


detectar saltos, tendencias y otros errores en los
datos. SAMS puede plotear los datos de forma
continua o en forma de barras. La escala es
determinado basado en los valores mximo y mnimo
de los datos, pero el usuario puede cambiar si lo
estima conveniente.

Anlisis Estadstico de Datos


Verificacin de la Normalidad y
Transformacin datos

SAMS verifica la normalidad de los datos ploteando


en papel de probabilidad normal y realizando pruebas
de asimetra de normalidad. Si los datos disponibles
no son normales pueden ser transformados mediante
tcnicas de transformacin logartmica, potencial o
de Box-Cox.

Anlisis Estadstico de Datos

Determinacin de las Caractersticas


Estadsticas de la Serie
SAMS puede calcular para los datos originales y transformados
los siguientes parmetros estadsticos:
Media, desviacin estndar, coeficiente de asimetra,
coeficiente de variacin y valores mximos y mnimos
Coeficientes de correlacin serial y coeficientes de correlacin
cruzada en el caso de datos anuales y coeficientes de
correlacin de estacin a estacin en el caso de datos
estacionales
Parmetros estadsticos referidos a sequas, avenidas y
almacenamiento

2.2. Fijacin del Modelo Estocstico


modelos especficos disponibles en SAMS
Para datos anuales:
Modelo Univariado ARMA(p,q)
Modelo Univariado GAR(1)
Modelo Multivariado AR(p); (MAR)
Modelo Contemporneo ARMA(p,q); (CARMA)
Modelo Desagregado Multivariado Anual (espacial)
Para datos estacionales:
Modelo Univariado PARMA(p,q)
Modelo Desagregado Univariado Estacional
Modelo Multivariado PAR(p); (MPAR)
Modelo Desagregado Multivariado Estacional

2.3. Generacin de Series Sintticas


La generacin de series sintticas es ejecutada una vez
definido el modelo y los parmetros estadsticos respectivos.
SAMS permite comparar los datos generados con los datos
originales con la finalidad de verificar si el modelo escogido
es el apropiado.

3. Generacin de Series Sintticas


La generacin de series sintticas es ejecutada una vez
definido el modelo y los parmetros estadsticos respectivos.
SAMS permite comparar los datos generados con los datos
originales con la finalidad de verificar si el modelo escogido
es el apropiado.

Purpose of using synthetic streamflows


instead of historical:
Reduce and more appropriately reflect
uncertainty of future reservoir inflows and
their effect on power marketing and reservoir
operating decisions

Uses for Stochastic Flows


Operations studies that determine inventory on
probabilistic basis in support of marketing decisions
Risk management-- studies that show the risk of
operating the system to meet specific criteria

The Problem
Produce sequences of monthly streamflows at 23
sites for forecast horizon of up to 2 years
Synthetic streamflows must behave statistically
similar to historical values and be consistent with
seasonal volume forecasts

The Solution
SAMS (Stochastic Analysis, Modeling, and
Simulation) -- developed as coop effort between CSU
and USBR
BPA funded enhancements to the model including the
capability to generate based on initial conditions
forecasts conditioned on NWRFC final volume
forecasts and antecedent flows

Configuration of Basin
Incremental flow
Total natural
flow
Seasonal volume forecast

Mica
Duncan
Revelstoke

Libby
Arrow

Bonneville
The Dalles

McNary
John Day
Lower Granite

Kootenay

Priest Rapids
Grand
Coulee

Albeni Falls
Boundary

Kerr

Long Lake

Lookout Point

Columbia Falls

Dworshak
Brownlee
Post
Falls

Hungry Horse

Partitioning a time series into its various components -- deterministic and stochastic
components

Systematic Approach to Hydrologic Time Series Modeling


Characteristics of overall water resource
system

Identify model composition:


univariate, multivariate,
disaggregation?

Selection of Model type:


Statistical characteristics of hydrologic
time series

AR, ARMA, etc

Identification of Model Form: order,


periodicity?

Estimation of Model Parameters: method of


moments, method of maximum likelihood

Testing goodness of Fit of the Model: how


well do results represent history?

Evaluation of Uncertainties: model


uncertainty, parameter uncertainty

Modeler input: knowledge, experience,


bias, limitations

Characteristics of Hydrologic
Physical Processes

Stochastic models reproduce basic statistics and


probability distributions of the historic data
assumption that variable is normally distributed

Approach
MPAR(3) used for 6 key sites
Spatial disaggregation of seasonal data for other 17 sites-generation of subkey sites (and subsequent sites)
dependent on generated values at key sites
Used modified flows for 1929-1983 and Kuehl Moffitt
synthetic seasonal volume forecasts

Autoregressive model
univariate, stationary case
p

y j (y j )
j1

Multivariate case and periodic case more complicated--matrix form used

Dissagregation models
Developed for reproducing statistics at more than one
level of aggregation
General Form: Y = A X + B Z + C

Stochastic Flow Forecast Generation Process

antecedent flows
at 24 sites

Seasonal
volume
forecasts

Exogenous
parameters
(SOI, MEI,
PDI)

Model parameter files

Post processor:

Pre-processor:
facilitate input data
checking and prepare
formatted input file
to SAMSF

SAMSF
generate flow and
seasonal volume forecasts

HYDSIM inputs:
1.flow forecasts
2. seasonal vol forecasts
for VECCS

summarize 1000
sequences into a smaller
subset that
defines the distribution
of flow forecasts.
Output necessary files

Volume forecasts
for VURC
computations

Output used
to generate
graphic summaries

Results

forecasted vs observed at TDA


10/29-9/46

forecasted (3 mos ahead)

observed

forecasted (6 mos ahead)


45000
40000

30000
25000
20000
15000
10000
5000

date

Oct-45

Oct-44

Oct-43

Oct-42

Oct-41

Oct-40

Oct-39

Oct-38

Oct-37

Oct-36

Oct-35

Oct-34

Oct-33

Oct-32

Oct-31

Oct-30

0
Oct-29

mo avg volume (kaf)

35000

forecasted vs observed at TDA


10/46-9/63
forecasted (3 mos ahead)

observed

forecasted (6 mos ahead)


50000
45000

35000
30000
25000
20000
15000
10000
5000

date

Oct-63

Oct-62

Oct-61

Oct-60

Oct-59

Oct-58

Oct-57

Oct-56

Oct-55

Oct-54

Oct-53

Oct-52

Oct-51

Oct-50

Oct-49

Oct-48

Oct-47

0
Oct-46

mo avg volume (kaf)

40000

date

Oct-81

Oct-80

Oct-79

forecasted (6 mos ahead)

Oct-78

Oct-77

Oct-76

Oct-75

Oct-74

Oct-73

Oct-72

Oct-71

Oct-70

Oct-69

Oct-68

Oct-67

Oct-66

Oct-65

Oct-64

Oct-63

mo avg volume (kaf)

forecasted (3 mos ahead)


forecasted vs observed at TDA
10/63-9/82
observed

60000

50000

40000

30000

20000

10000

month-yr

Apr-02

Mar-02

Feb-02

Jan-02

Dec-01

Nov-01

Oct-01

Sep-01

Aug-01

Jul-01

Jun-01

May-01

Apr-01

Mar-01

Feb-01

Jan-01

Dec-00

Nov-00

Oct-00

Sep-00

Aug-00

Jul-00

Jun-00

May-00

avg monthly volume (kaf)

observed
stochastic forecasts at TDA
May-00, Oct-00 fcsts 50%tile forecasts
with 5% and 95% tile forecasts
historic mean, max,
min

30000

25000

20000

15000

10000

5000

month-yr

Apr-02

Mar-02

Feb-02

Jan-02

Dec-01

Nov-01

Oct-01

Sep-01

Aug-01

Jul-01

Jun-01

May-01

Apr-01

Mar-01

Feb-01

Jan-01

Dec-00

Nov-00

Oct-00

Sep-00

Aug-00

Jul-00

Jun-00

May-00

avg monthly volume (kaf)

observed
stochastic forecasts at TDA
Jan-01, Mar-01, May-01 50%tile forecasts
with 5% and 95% tile forecasts
historic mean, max,
min

30000

25000

20000

15000

10000

5000

study date: 5/23/0


1

m onthly stochastic flow fcsts, cond'd on obs flow s thru Apr 30;
May FF vol fcst; char'd by 60 uniform quantiles (using May-Sep vol at TDA );
vert bars are 5%, 50%, 95%-tiles of fcst dist (using mo'ly vol at site)

fcst period: may 2001 - apr 2002

TDA( 365)

based on FF May TDA jan-jul fcst=56.5 MAF


400000

350000

forecast avg
historic m edian

300000

(compare to 50% fcst


as depicted by upper
center of vertical bars)

flow (cfs)

250000

historic extrem a
(com posite)

200000

observed
150000

100000

50000

Feb-01

Mar-01

Apr-01

May-01

Jun-01

Jul-01
month

Aug-01

Sep-01

Oct-01

Nov-01

Dec-01

Future opportunities

gain experience using stochastic forecasts


disaggregate into daily/weekly flows
produce forecasts at more sites
use of exogenous variables
condition on new NWRFC volume forecasts
compare to ESP