Final Report for cs5601(Fuzzy Sets: Theory and Applications):

Try To Pridict The Weather

³¯«Å¾§ 853739

Table of Contents

Abstract	Problem Definition	Data Set Description	Our Approach
Simulation Results	Conclusions	Computer Programs	Division of labor
References

---

Abstract

This report describes our attempt to apply ANFIS (Adaptive Neuro-Fuzzy Inference Systems) for the prediction of weather. The preliminary results show that ANFIS is an accepatible approach to pridict the weather.

Problem Definition

This project tries to make pridiction of weather using the ANFIS approach. The problem of weather prediction is a diffucult job.We try to use the ANFIS to solve the very much nonlinear and chaotic problem.This is an approach to make the pridiction of weather without using supercomputer to solve the Lorenze's equation.We just use a personal computer.

Data Set Description

• Origin: This dataset was original from the ¤¤¥¡®ð¶H§½·s¦Ë®ð¶H¯¸.It contains the data of temperature,moisture and rainfall from 95/Jan/1 to 97/Apr/30.
• Description: :
• Input 1: temperature
• Input 2: moisture
• Output: rainfall
• Visulization :

Approach

Our approach to this problem can be explained in two aspects:

• Classification: ...
• ANFIS Training: ...

Simulation Results

• Data Preprocessing:
• Data screening: Input-output pairs with missing values were eliminated and there remains 392 valid instances for NN training. Also input 8 was deleted from the data set since it is a unique string for each instance, which is not usable for NN training.
• Input normalization: none.
• Output normalization: All outputs are linearly normalized to [-1, 1].
• Data reduction: none
• Others: none
• Input selection: ...
• ANFIS Structure:
  Describe your network architecture, such as
• 2-input 4-rule ANFIS (fcm1.m)
• 4-input 8-rule ANFIS (fcm2.m , fcm3.m , fcm4.m)
• Training method:
  Describe your training methods, such as
• Vanilla backpropagation:
• steepest descent: defaulted by ANFIS.M program.
• Training Results:

fcm1.m : Data fitting of the rainfall and temperature+moisture in the same day. We find that the RMSE keep high after many epochs . So, the temperature+moisture only are not enough to guess the rainfall condition. More information is needed.
(See also picture : fcm1fig1.gif and fcm1fig2.gif )

fcm2.m : This program has the same goal as fcm1.m . Since it was metioned above that 2 conditions -- temperature and moisture -- are not enough. Their differences are introduced as the input of ANFIS. The RMSE can go down to a factor of about 0.2 now.
(See also pictures : fcm2fig1.gif and fcm2fig2.gif )

fcm3.m : The success of fcm2.m encouraged us to try the prediction of the rainfall one day after the day that the temp. and moist. data were got. The temp. and moist. and their diff. of day 1 are used,and the rainfall of day 2 are predicted. But the result is not ideal at the part of checking data. This method is not good without on-line learning.
(See also pictures : fcm3fig1.gif and fcm3fig2.gif )

fcm4.m : If the data of 2 days ago are use, the prediction quality may be improved.But,data of 1 data have 4 terms(temp. ,moist. and their diff. respectively),2 days,8 terms. That cause an 'Out of memory' error.So, only moisture and its diff. of 2 days before the day predicted are used as inputs of the ANFIS.the result becomes better than fcm3.m
(See also pictures : fcm4fig1.gif and fcm4fig2.gif )

Concluding Remarks and Future Work

• Concluding Remarks:
  ¤Ñ¦³¤£´ú­·¶³
  ¤HÃþ¤@«ä¦Ò,¤W«Ò´Nµo¯º
  It is an useful idea to use ANFIS or the fuzzy and NN to pridict nonlinear(even very chaotic)problem.Though the results are not ideal, they are acceptable.
• Future Work:
• Try to use new methodsof training and classification .
• Try to introduce Gaussion or LM method into ANFIS.
• If memory and calculation speed allow, try bigger training data sets. ex. if the data of some whole years cam be include, the 'quasi-periodic' phenomena will be taken into effective.It will helpful in the prediction of the rough trend.
• More history data before the day which will be predicted are used,better the result will be.

Computer Programs

fcm1.m

fcm2.m

fcm3.m

fcm4.m

Division of Labor

• ³¯«Å¾§: µ{¦¡¼¶¼g,¹w´ú¤èªk
• ´¿±R­õ: ¸ê®Æ¦¬¶°¾ã²z,¼Ò«¬«Ø¥ß
• ªL©s¾§: html¼¶¼g,¤fÀY³ø§i,ÃD¥Ø§Î¦¨

References

• [Jang97] J.-S. R. Jang, C.-T. Sun, and E. Mizutani, `` Neuro-Fuzzy and Soft Computing: A Computational Approach to Learning and Machine Intelligence,'' Prentice Hall, 1997.
  
  
  ¯S§O·PÁ·s¦Ë®ð¶H¯¸´£¨Ñ¸ê®Æ