WEKA: A Useful Tool for Air Quality Forecasting

Preview:

DESCRIPTION

 

Citation preview

Weka: A Useful Tool for Air Quality Forecasting

William F. Ryan

Department of Meteorology

The Pennsylvania State University

wfr1@psu.edu

2007 National Air Quality Conference, Orlando

Weka

The weka, or woodhen, is a birdnative to New Zealand. Weka is

also the name of a suite of machinelearning software tools, written in

Java, and developed at the Universityof Wiakato in New Zealand.

http://www.cs.waikato.ac.nz/ml/weka

Machine Learning

• Machine learning is a subfield of artificial intelligence (AI) concerned with the development of algorithms and techniques that allow computers to "learn".

• The machine learning algorithms in Weka include, among others, linear regression, classification trees, clustering and artificial neural networks (ANN).

Weka Can Be A Useful Tool

• Weka has the potential to be a useful tool to support local air quality forecasting efforts – particularly those operating on a limited budget. – Weka is open source (free) software - although the

purchase of the associated text book is strongly recommended.

– Weka is easily installed on standard PC's but can also run on Linux and other platforms.

– Only minimal modifications are necessary to prepare data files for use in Weka.

– The user interface is simple and intuitive.

Weka and PM2.5 Forecasting

• Of particular interest to air quality forecasters is the wide range of algorithms included in Weka.

• These algorithms may be useful to address shortcomings in statistical forecast guidance for fine particulate matter (PM2.5).

• Simple linear regression methods provide reasonable skill for O3 forecasting, due to the very strong and nearly linear ozone-temperature relationship, but linear regression methods have shown limited skill in forecasting PM2.5.

PM2.5 Forecasting

O3 (left panel) is well-behavedstatistically. Distribution is nearnormal with a strong associationwith maximum temperature. As a

result, linear techniques areuseful.

PM2.5 (right panel) is not well-behaved. Distribution is skewed,

no strong association with anyparticular weather variable.

Tools included in Weka, including ANN and classification

and regression trees (CART), are capable of addressing

non-linear problems posed by PM2.5.

Weka: Information

http://www.cs.waikato.ac.nz/ml/weka/

Input File Format

Weka uses its ownfile format called: *.aarf

All you need to dothough is provide a*.csv file with variablenames in the first lineand Weka will convert

aarf Format

aarf format is simple anyway:

ASCII fileList of variable and type

Then data follows, comma separated

Missing data marked as “?”

Data Editing

Data can be easily editedwithin Weka itself

Analyzing Data

Variables can be easilyscanned with basic

statistics and histogramsprovided by Weka

Quick Analysis Tools

Sampling and Test Data Set Options

Functions Available

WEKA includes a number of different techniques that can be useful for forecast development.

These include:

Linear and logistic regressionPerceptron models (Neural networks)

Linear Regression

Unfortunately, the “work horse” linearregression module in Weka is limited inusefulness:

-No automatic stepwise function-Poor diagnostics

Compare: SYSTAT, Minitab

Classification and Regression Trees (CART)

A variety of classificationalgorithms are available.

Standard algorithm isJ48, which is a souped up version of the lastfree version of CART(Version 4.5)

Commercial version iscurrently 5.0.

CART Options

A number of optionsare available tofine tune the CARTAnalysis:

-Minimum # of cases per node-Types of pruning: e.g., sub-tree raising-Confidence values for splitting nodes

CART Diagnostics

CART is notorious for usingCPU resources but the WEKAversion runs efficiently on mystandard PC.

Diagnostics are better forCART than linear regression.

Example on left is of a 4 categoryPM2.5 CART forecast.

CART Visualization

Artificial Neural Networks (ANN)

“Linear Regression by a mob”

Produces forecast bytaking the weightedsum of predictors andthen layering the process.

Artificial Neural Networks - Summary

Known samples (historical data) are used to “train” the network.

Input data (xi) are assigned weights (wi) and combined in the “hidden” layer – like a set of linearregressions. These sets are then combined in additional layers – like regressions of regressions.

The sum of data and weights are transformed(“squashed”) to the range of the training data and error is measured.

A supervised training algorithm uses output error to adjust network weights to minimize errors.

Artificial Neural Networks – Pros/Cons

• Pro: ANN’s are a powerful technique utilized across scientific disciplines.

• Pro: Theoretically well suited to non-linear processes like air quality.

• Con: Not transparent to users. Hard to integrate into forecast thinking.

• Con: Technically difficult to understand, raises risk of misuse.

Example: Neural Network Structure

www.doc.ic.ac.uk/~sgc/teaching/v231/

WEKA Neural Networks

WEKA provides user controlof training parameters:

# of iterations or epochs (“training time”)

Increment of weight adjustments in back propogation (“learning rate”) Controls on varying changes to increments (“momentum”)

Conclusions

• Weka is a low-cost forecasting tool that has the potential to be a useful for air quality forecasting – particularly in situations where non-linear effects dominate.

• Some Weka modules are not fully developed for forecast algorithm development.

• Patience, use of textbook and Weka listserv are required to get the most out of the program.

URLs of Interest

• Weka:– http://www.cs.waikato.ac.nz/ml/weka

• Mailing List: – https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

• Mailing List Archives– https://list.scms.waikato.ac.nz/mailman/htdig/wekalist/

• Informal FAQ:– http://www.public.asu.edu/~sksinghi/weka-faq.html

Acknowledgements

• The Delaware Valley Regional Planning Commission (DVRPC) – Mike Boyer and Sean Greene – and the member states (PA, DE and NJ) for supporting air quality forecast development.

• Dr. George Young of Penn State for his advice, patience and teaching skill.

Recommended