Prevent ANFIS from negative or unrealistic outputs

Hi all, I'm using ANFIS in order to forecast load values based on several inputs. However, many outputs are negative or sometimes very high valued and the accuracy is very bad. It is to be noted, that nor negative neither high values are included in the learning data.

I tried several things to address these problems, e.g. - preprocessing of the inputs - enlarge or shrink the learning data - adapting learning algorithm - adapting and/or methods - ...

Whatever I did, the outputs didn't get any better. I am aware of the fact, that the curve to predict is highly fluctuating and therefore a very complex case for ANFIS. Nevertheless, I hope to improve my results with further adaptions.

Has anyone any idea? I'm rather helpless at this point and really really stuck.

Thank you very much in advance! Martin

[Heleno]

Hi,Martin

I am a new user of ANFIS and I have some problem you posted above.

Neverthless, I would like to know if you could give me some tips to solve my problems.

Thank you very much

[Martin]

If you could describe your problems a bit more precisely, I would be happy to help you if I am able to.

Regards,

Martin

Answer 1

[Win co]

Hi, Could you show me your data file (train and check data) ? Winn

[Martin]

Hi, I use data like stored in the .xlsx file linked below. The load values to be predicted are in the first column, the rest (except of date) are input parameters for anfis. However, I do not use all of the parameters. The most important seem to be

- Hour

- Weekday

- PrevWeekSameHourLoad

- prevDaySameHourLoad

- Awake and present

The train and checking data varies: I split up the complete dataset to several subsets and use about 90% for training and 10% for checking. Then I try to forecast the next few values (from 24 to 168 values). In the next step I use recent values to forecast the next and so on. I also varied the range of training data from few to many.

However, in any approach I tried, the problems I mentioned above appear.

http://dl.dropbox.com/u/7967224/Example.zip

Thank you very much for your help!

Martin

Answer 2

[Martin]

One more thing: I used genfis to generate the anfis system and I also tried to develop an own underlying FI system. However, in each case the problems appeared.

Answer 3

[Win co]

Firstly, I'm sorry about not replying soon because this forum has not sent me notification of your comment. Secondly, your link seems not downloadable. Thirdly, could you show us the figures of training and checking phase ? Regards, Winn

Answer 4

[Harry]

Hello,

i have exactly the same problem as Martin described above. Does anyone know how to solve the problem?

Thank you very much for your answer!