Transcript P100301816

Artificial Neural Network using for climate
extreme in La-Plata Basin: Preliminary results
and objectives
David Mendes*
José Antonio Marengo*
Chou Sin Chan+
*Centro de Ciência do Sistema Terrestre – CCST/INPE
+Centro de Previsão de Tempo e Estudos Climáticos – CPTEC/INPE
CLARIS Annual Meeting -WP 5- Rome, 22nd to 26th of February 10
OBJECTIVE
The objective of this study is to identify climate extreme (RClimDex), using Artificial Neural Network (ANN) that
can capture the complex relationship between selected large-scale predictors and locally observed meteorological
variables for temporal scale (predictands).
Using Artificial Neural Network to diagnose extremes in La-Plata Basin using Eta-HadCM3;
For Control Period (20c3m) and A1B scenarios;
 20c3m (1961-2000) or (1978-2000)
 A1B scenarios (2071-2100)
Artificial Neural Network (ANN) in Meteorology
Recently, non-linear approaches have been developed (in particular, the Artificial Neural Network ANN) and adopted as tools to downscale local and regional climate variables and extreme (climate) from
large-scale atmospheric circulation variables (e.g. Crane and Hewitson, 1998; Trigo and Palutikof, 1999).
Reference:
Gardner and Dorling (1998) – Review of applications in the atmospheric sciences.
Trigo and Palutikof (1999) – Simulation of Temperature for climate change over Portugal.
Sailor et al., (2000) – ANN approach to local downscaling of GCMs outputs.
Olsson et al., (2001) – Statistical atmospheric downscaling of short-term extreme rainfall.
Boulanger et al., (2006/2007) – Projection of Future climate change in South America.
Artificial Neural Network (RNA)
The ANN approach can be viewed as a computer system that is made up of several simple to the highly interconnected
processing elements similar to the neuron architecture of human brain (McClelland et al., 1986).
Supervised: observed precipitation (CRU);
No-supervised: auto-organization.
In this work, input Nodes is data base surface station or
Data interpolated
e.g. by: Mendes and Marengo (2009).
Multilayer Perceptons
O Multilayer Perceptons - The following diagram illustrates a perceptron network with three layers:
Input nodes
output nodes
Hidden nodes
Training Multilayer Perceptron Networks
 Selecting how many hidden layers to use in the network.
 Deciding how many neurons to use in each hidden layer.
 Finding a globally optimal solution that avoids local minima.
 Converging to an optimal solution in a reasonable period of time.
 Validating the neural network to test for overfitting.
Model
Extreme Climate
Extremes indices for La-PlataBasin have already been calculated from these daily station. Indices are calculated
using standard software produced on behalf of the ETCCDMI by the Climate Research Branch of the
Meteorological Service of Canada.
Preliminary results and objectives
Initial Results for exemple:
R25 (Number of very heavy precipitation day) – Annual count when prp > 25 mm day
Control (20c3m)
From 1978-1990
From 1978-1990
Future