|
Flood Prediction
Since
June of 2001, the SMU Database Research Group has been attempting to use
Data Mining techniques to make predictions using data that is spatially and
temporally distributed.
Challenge:
Our
specific application is to try and predict flood events on a river
using data about the current and past weather conditions. The techniques we
have explored so far include time-series analysis, neural networks, and
Markov models.
Research:
We have
focused on applying the principles of Hidden Markov Models (HMMs) and derived
models to solve the problem above. A HMM is made up of two stochastic (or
probabilistic) processes that can produce a sequence of observable symbols.
HMMs can be trained to 'recognize' a particular sequence of symbols using
specialized algorithms. In one of my proposals for solving the flood
problem, several HMMs are trained to recognize the weather conditions that
precede a flood event. The hope is that they will be able to predict a
flood before it happens based on the current weather conditions.
Proposals:
We are
also attempting to model the water level of a river using the internal
states of a HMM. The weather conditions that influence the water level are
allowed to influence the transitions between the states in the HMM. The
hope is that, given a set of weather conditions that precede a flood, the
HMM will make the state transitions that lead to a 'flood state'.
The
proposals are quite simple conceptually but, as they say in speech
recognition research, the devil is in the details. Both models have been
implemented in Java and are currently being tested using simple data sets
and looking for ways to improve their performance.
In
preliminary tests, the models have performed well using best-case or
trivial scenarios but have produced large errors on the more realistic
datasets. By making some modifications to the models, we hope to reduce
these errors so that we can eventually decide how appropriate HMMs are for
modeling river conditions. HMMs are efficient at modeling real world
processes that are characterized by periods of steady behavior and periods
of gradual change.
This
research project was initially motivated by the SIVAM
project -- an effort to monitor the conditions of the Amazon River in
Brazil. Our long-term goal is to adapt some of our models to make flood
predictions on the Amazon River. In the mean time, we hope to submit some
papers describing our models and their performance.
|