Are reasonable targetless multistep ahead predictions in a NARXNET even possible?

1 次查看（过去 30 天）
PetterS
PetterS 2015-6-28
编辑： Greg Heath 2016-1-4
Hi!
I want to forecast the future of a time series that I believe is depending on several other external inputs. Now since I want to predict something occurring in the future I can model and train a NARXNET on historical dependencies between external and target variables but for the time steps I want to predict I won’t have the actual external inputs available.
My initial idea on how to tackle this problem was to design one individual NARNET for each of my external input time series – predict ahead into the future and then use those forecasted results as external inputs into the NARXNET to simulate the target value I want with the hope that that would be more accurate compared to simply doing one NARNET on the time series I want to predict. This approach would require many different nets depending on each other and it would be much less complicated and time consuming if I could do the prediction simply by using one NARXNET alone. My concern is simply whether or not this is possible. I’ve looked at the guide in the newsgroup called: NARXNET CODE FOR MULTISTEP AHEAD PREDICTIONS
Which supposedly does precisely what I want – predicts beyond the known range completely targetless into the future using empty cells as inputs. However when I run the code of that tutorial and plot the resulting predictions they just collapse as soon as then become targetless and don’t even look realistic at all. And this tutorial is based on the “simplenarx_dataset” which could potentially be a lot easier to predict compared to a real life problem. I find it a bit puzzling that a tutorial on predictions would result in a completely collapsed prediction, does it also do that for everyone else or am I doing something wrong?
Even if I try the tutorial code on a problem of my own using a network architecture (input delays, feedback delays, number of hidden nodes) that has proven to have really good performance during historical training on my problem – I’ve never been able to acquire targetless NARXNET predictions even resembling anything realistic. So to summarize my main questions:
1. Is the tutorial on targetless NARXNET predictions giving other people reasonable results?
2. Would it be a good idea to create many individual NARNET predictions on my external inputs to then use as inputs to a NARXNET?
3. Are there any other threads/guides/recommendations on targetless NARXNET predictions that are successful that could be of help? I’ve looked at many threads here on the forum but so far I have yet to see any real evidence of good targetless NARXNET predictions.
  1 个评论
Charlie M
Charlie M 2015-11-10
Hi Petter,
I was wondering if you have solved this problem as I have got the exact same issue? I am predicting 7 steps ahead but beyond the inputs too. Can you please tell me what you have done?
Thank you!

请先登录，再进行评论。

请先登录，再回答此问题。

采纳的回答

Greg Heath
Greg Heath 2015-6-29
编辑：Greg Heath 2016-1-4
AS I have stated in
Subject: HELP NARXNET BUGS, COMMENTS and SUGGESTIONS
From: Greg Heath
Date: 7 May, 2015 13:06:58
Message: 3 of 3
---SNIP--
8. Extension of results beyond the original specified target can be
accomplished with the closed loop configuration
provided the extension of the original input is known. Filling an
unknown input extension with NaNs and relying on the closeloop
feedback will not yield valid results. However, either predicting
the input or using the inputless NARNET should be considered.
Hope this helps
Thank you for formally accepting my answer
Greg
  2 个评论
PetterS
PetterS 2015-6-29
Ok, thanks I didn’t see that thread. I guess I’ll have to design an awful lot of NARNETS on my inputs then!
Greg Heath
Greg Heath 2015-6-29
编辑：Greg Heath 2016-1-4
1. You may be more successful with the targetless prediction
for the simpleseries_dataset because the original input is
nearly linear.
2. Currently ID, FD and H were all defaults. Perhaps the
following have merit:
a. ID = 0
b. Significant auto and cross correlation delays
c. Minimizing H subject to NMSEo < 0.005.
d. Using NARNET for predicting X and T
3. Keep in touch via the NEWSGROUP.
Greg

请先登录，再进行评论。

更多回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 Sequence and Numeric Feature Data Workflows 的更多信息

标签

产品

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by