Weather Forecasting Using Learning Strategies

1
Weather ForecastingUsing Learning Strategies

• 55145 Pattern Recognition Final Project
• Dec. 8, 2005

Daisy Espino Sangyeol Lee Dept. of Biomedical E
ngineering, The University of Iowa
2
Content of Presentation

• Preface
• Weather Data Collection
• Weather Forecasting
• Fuzzy Logic Method
• Neural Network Method
• Neuro-Fuzzy Method
• Experiment
• Conclusion

3
Preface

• Problem Statement
• weather data prediction using learning
  strategies
  
• Work Flow

Daisy Espino
Data Collection
Fuzzy Logic
Documentation
Sangyeol Lee
Neural Network
Neuro-Fuzzy
4
Weather Data Collection

• Historical Data/Current Data www.wunderground.co
  m , http//mesonet.agron.iastate.edu
  
• Historical Data 1996-2004, 11/30 to 12/15, 11
  Parameters
  
• Iowa City and 8 neighbors, Boston City and 4
  neighbors
  
• Current Global Data www.usatoday.com ,
  www.wundergound.com
  
• Primarily for Fuzzy Logic
• Low/High Pressure areas, Front locations
• Weather Rules Many sites, but best site was
  www.usatoday.com
  
• Primarily for Fuzzy Logic/Neuro-Fuzzy
• Many variations in rules

5
Fuzzy Logic Weather Forecasting

• Considerations
• Weather rules themselves are fuzzy (e.g. wind
  direction)
  
• Global data changes frequently
• Strategies involved
• Determining weather rules
• Assigning membership functions to weather
  parameters
  
• Adjusting weights in certain cases
• Extensive experimentation because of high
  interdependency

6
Matlab Fuzzy Logic Toolbox

• Easy to use, high flexibility
• Two types Mamdani
• Or Sugeno
• Our Model
• Sugeno
• Andmin Ormax
• Implicationmin
• Aggregationmax
• Defuzzificationcentroid
• Rule Viewer is excellent for troubleshooting

7
Determining Weather Rules

• Precipitation Type
• Inputs Pressure/Front/Temperature
• Precipitation Amount
• Inputs Humidity/Front
• Humidity
• Inputs Pressure/Front/Temperature
• Temperature
• Inputs Humidity/Front/PrevTemp
• Wind Direction
• Inputs Pressure/Front/PrevWind

8
Determining Member Functions

• Example Humidity

9
Final Adjustments

• Example1 PrecipitationType
• Need None category
• Example2 Temperature
• Must adjust weights for slight increase or
  decrease in temp
  
• If Humidity is Medium and Front is Cold and
  PrevTemp is Warm then Temperature is Warm
  (0.8)
  

10
Neural Network Weather Forecasting

• Considerations
• Weather parameters are inter-dependent.
• Weather parameters are daily averaged.
• Strategies involved
• Neural time-series prediction
• Statistical analysis of weather parameter
• Weather pattern matching

11
Neural Network Weather Forecasting

• Neural Time-Series Prediction

xisig(wijkj) i1, 2, , n j1, 2, , 10
Weather data k1TempHi k2TempLow k3Humid
ity k4WindDir k10S
kyCoverage
Output xi ???, n???
Weight wij ???
12
Neural Network Weather Forecasting

• Output Parameter Decision
• A process of selecting the most descriptive
  weather parameters from cross-correlated data
  
• The Principle component analysis (n, xi)

13
Neural Network Weather Forecasting

• Node Weight Decision
• Cross-correlation (wij) between the selected
  principle component and weather parameter

Iowa City, IA
14
Neural Network Weather Forecasting

• Pattern Matching
• The least mse sense
• Higher priority to neighbors condition
• Wind direction voting

NW Sioux Falls, SD W Omaha, NE SW Topeka,
KS
Iowa City, 12/6/2005
15
Neuro-Fuzzy Weather Forecasting

• Fuzzy Rules for Weight Adjustment

P median(P)
Y
H
N
Y
TemL median(TemL)
Rain
N
Rain ,Snow
H median(H)
Y
Rain ,Snow , TemH , TemL
N
Rain ,Snow
16
Experiment

• Programming Environment
• Matlab 7.0 R12 (Matlab GUI Fuzzy Toolbox)

17
Experiment
18
Conclusion

• Weather forecasting based on learning strategies
• Fuzzy logic
• Neural network
• Neuro-fuzzy system
• Limitation in data set
• Inter-dependency
• Daily average statistics
• For higher accuracy reliability
• Data cleaning
• Investigation of global factors
• and

19

• Thank You