CecileMallet
commented
3 years ago Thank you for your email. I will be out of the office until August 26th, I will review your message following my return.
Best regards
Maya
Le 2 août 2021 à 15:24, PeterLuenenschloss @.***> a écrit :
Hallo, i stumbled over your publication https://link.springer.com/article/10.1007/s41060-019-00193-1 concerning rainfall data allignment with dtw approach.
I am working on a rainfall data imputation project (10 min sampling) with a machine learning approach. The project is a cooperation of the research facility i work for, the center for environmental research https://www.ufz.de/index.php?en=33573 and the german weather service https://www.dwd.de/DE/Home/home_node.htmll My idea to improve prediction results, is to warp the predictor (meassurements from nearby stations) onto the target rainfall timeseries before training the model.
So i was happy when finding out, that there is already some research done in that direction.
I have some questions regarding the algorithm. Unfortunately, the warp path i get with the algorithm, includes associations of rainfall events that are really distant from each other (more than 2 Days for stations that are seperated by less than 30 km) - you referenced that unwanted effect in your paper and tackled it by setting the radius to zero - wich unforunately doesnt solve the issue for me.
i guess it doesnt work with my timeseries, because they are really long (10 years) and the local stretch constraint is somehow implicitly dependend on the timeseries length.
My questions are: 1)can this effect somehow be mitigated by tweaking the aggregation_step parameter? I somehow not really understand how this parameter works or how it is supposed to be interpreted.
2)Have you made some new findings/progress on rainfall data warping since your publication? In the paper you state that you would like to continue investigating the subject. I would be really thankfull for some ideas! Of course i could also give feedback about how it worked out with my data.
with regards Peter Lünenschloß
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/djallelDILMI/IMs-DTW/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADE3UAQJHQMST2456ESXKUDT22L7FANCNFSM5BMXRHRA.
Hallo, i stumbled over your publication concerning rainfall data allignment with dtw approach.
I am working on a rainfall data imputation project (10 min sampling) with a machine learning approach. The project is a cooperation of the research facility i work for, the center for environmental research and the german weather service
My idea to improve prediction results, is to warp the predictor (meassurements from nearby stations) onto the target rainfall timeseries before training the model.
So i was happy when finding out, that there is already some research done in that direction.
I have some questions regarding the algorithm. Unfortunately, the warp path i get with the algorithm, includes associations of rainfall events that are really distant from each other (more than 2 Days for stations that are seperated by less than 30 km) - you referenced that unwanted effect in your paper and tackled it by setting the radius to zero - wich unforunately doesnt solve the issue for me.
i guess it doesnt work with my timeseries, because they are really long (10 years) and the local stretch constraint is somehow implicitly dependend on the timeseries length.
My questions are: 1)can this effect somehow be mitigated by tweaking the aggregation_step parameter? I somehow not really understand how this parameter works or how it is supposed to be interpreted.
2)Have you made some new findings/progress on rainfall data warping since your publication? In the paper you state that you would like to continue investigating the subject. I would be really thankfull for some ideas! Of course i could also give feedback about how it worked out with my data.
with regards Peter Lünenschloß