Presentaci

1
Self-Organizing Maps for Weather Forecasting
Application Antonio S. Cofiño
The 1st EU CrossGrid ConferenceMarch 17 - 20,
2002,   Cracow, POLAND
http//grupos.unican.es/ai/meteo
2
SOM Application for DataMining
Adaptive Competitive Learning
Downscaling Weather Forecasts
Sub-grid details scape from numerical models !!!!!
3
Atmospheric Pattern Recognition
Prototypes for a trained SOM. Close units in the
lattice are associated with similar atmospheric
patterns.
T 1000mb
T 500mb
Z, U, V 500mb
4
Neural Networks for Unsupervised Clustering
Competitive neural networks perform clasification
with a single requirement specifying the number
of desired clusters.
Two output Neurons (two classes)
Three output neurons (three classes).
5
Self-Organizing Maps. Preserving Data Topology
SOMKMeans Preserving Neigbourhood
Cluster units are located on a 2D lattice, each
one associate with a pattern prototype (dimension
500).
Dimension 500
Adaptive Competitive Learning
Topology preserving transformation. Close
clusters in the lattice correspond to close
prototypes in the high dimensional data space.
6
Self-Organizing Maps. A simple example.
7
User Interface.
Monitoring
Interactive Graphic
DATASET Dictionary (Classes) Basic
Object Derived Procedures/Methods
Alphanumeric Output
Analysis Scripts
Work Persistency
8
Data Computing Flow in SOM Application
SOM Trainning Flow
UI
Broker
Master
Database Storage
SOM training in each Data Cluster
Database Splitting
9
XML Schema for Data Patterns.
lt?xml version"1.0" encoding"UTF-8"?gt ltClass
xmlnsxsi"http//www.w3.org/2001/XMLSchema-instan
ce" xsinoNamespaceSchemaLocation"meteo.xsd"
gt ltStream model"Wave model"gt ltVersion
value"4096"gt ltType type"Forecast"gt
ltDategt1967-08-13lt/Dategt
ltTimegt1200lt/Timegt ltStepgt24lt/Stepgt
ltNumbergt0lt/Numbergt ltLevel type"Pressure
level"gt1000lt/Levelgt ltParameter
table"ECMWF"gtZlt/Parametergt lt/Typegt
lt/Versiongt lt/Streamgt lt/Classgt
XML Instance