flood forecasting
Hello! I'm working on a multivariate time series, 7 independent variable as rainfall depth (mm/hr) and one dependent variable as stream-flow (cumecs), so i want to develop a flood forecasting model by using past values of streamflow time series, together with joint time series of the observed present and past, as well as anticipated future values of rainfall time series. I want to forecast rainfall the from the forecasted rainfall I can be able to predict the stream flow.
Let y be the variable (scalar/vector-valued) to be forecasted let Yt be the joint time series of the
present and past values of y.
Yt = [yt, yt-1,…, yt-n],
Let u be the variable (scalar or vector-valued) that is in causal relationship with y, and let Ut be
the joint time-series of the observed present and past, as well as any anticipated future values
(denoted by hat) of u, such that,
Ut = [Ūt+α, ut, ut-1,…,ut-n], and let
Zt = [Yt, Ut],
α > 0,
Lead time forecast of the y variable is p(yt+α|Zt),
α= 1
yt+1 = [yt, yt-1, …, yt-n; Ūt+1, ut, ut-1,…,ut-n]
Inwould like to use neural networks and svm to carry out this task
Answers
@thabo_bafie are you trying to make a more responsive HEC-RAS? There was an article that @mschmitz sent me (can't find it right now) about using deep learning for flood modeling.
hi @thabo_bafie - I would assume you've read the KB article on the Time Series extension by @tftemme? He is our resident time series data guru at RapidMiner.
https://community.rapidminer.com/t5/RapidMiner-Studio-Knowledge-Base/Time-Series-Extension-new-version-0-1-2-with-ARIMA-Trainer/ta-p/45151
Scott
yes! something like that but using data-driven modeling methods rather that HEC-RAS, when you find the article you were talking about please feel free to share it with me.
Thank you!
@Thomas_Ott
yes! something like that but using data-driven modeling methods rather that HEC-RAS, when you find the article you were talking about please feel free to share it with me.
Thank you!
@thabo_bafie I think this is the one: https://ieeexplore.ieee.org/document/7966716/, but it's toll gated by IEEE.
Forecasting of precip is going to get a lot trickier with all the additional moisture in the air (climate change).
I do a lot of runoff analysis with the SCS and Rational Methods but not so much percip forecasting. I suspect that the unit hydrographs are going to change for many areas.
@Thomas_Ott
Hello Thomas! Thanks for your ansa, one more thing can you please clarify something for me here, was running my model here using sliding window validation together with SVM and I got
How do you interpret this results, and what is mikro?
One more, when using windowing operator and sliding window validation operator, do you have to set all the parameters the same, eg... window size for windowing operator should it be the same as the training and test window width, same thing for step size moreover, how about horizon, should it be the same across all operators "windowing, sliding window validation and forecasting performance"?
Thank You!
@thabo_bafie a discussion of mikro is here: https://community.rapidminer.com/t5/RapidMiner-Studio-Forum/mikro-in-perfomancevector/td-p/13872
W.R.T. to your results, you're trend accuracy is only 34.5%, which is terrible. I usually grid optimize training/testing/horizon parameters as well as SVM's C and gamma values. There's a great example here on how to optimize parameters for volatility: http://www.neuralmarkettrends.com/predicting-historical-volatility-for-the-sp500/
@Thomas_Ott it's so funny how we get so used to see "mikro" after all these years. I just filed a ticket with dev team as of course it should be "micro" (or maybe "micro average" as per Ingo's post).
Scott
@sgenzer it never bothered me cause it was 'mikro' is the german word for 'micro.'
@Thomas_Ott, Thanks so much, but you didn't answer my last question about how to set the parameters of the windowing, sliding window validation and forecast performance operator.
@thabo_bafie yes I did, via parameter optimization. So you'll start out with a range that *you* think makes sense. Obviously, your initial paramater values gave you a very poor trend accuracy. So I suggested parameter optimizatopn where you enter a range (i.e. 2-20) for the training/test/horizon parameters and it the process then tests all combinations and gives you the trend accuracy for each combination.
Doing that helps you understand the time frames (windows) that you should investigate to get a good model.
@Thomas_Ott
Thank you so much, but you haven't answered the last question about the setting of windowing, sliding window validation and
forecast performance operator....
Regards!
@Thomas_Ott
Thank you so much, Thomas sorry about that last message.