on 04-04-2014 11:56 AM
Dear All,
I have a problem with one of PAL procedures!
Currently, I am using FORECASTSMOOTHING with Triple Exponential Smoothing from SAP HANA Predictive Analysis Library (PAL). I am working with hourly data and I am testing this procedure only on 6445 records.
I have a problem with cycle parameter that I can't overcome...
When the cycle parameter is setted to 24 (hourly data) I can use only 4267 records from my input table. Even one record more generates error: “Could not execute 'CALL _SYS_AFL.PAL_TSF_TMH(TMH.INPUT_DATA_TS_F, "#PAL_CONTROL_TBL_TS", TMH.PAL_OUT_TS_F, ...' in 383 ms 461 µs . SAP DBTech JDBC: [2048]: column store error: search table error: [2620] _SYS_AFL.AFLPAL:FORECASTSMOOTHING: [135] (range 2) AFLFunctionFatal exception: PAL error[73001502]:Ahead exception throw out, Check trace for details”.
But when I change cycle to 5 or less I can use as many records as I want. Have you ever had such problem? I hope you could help.
Regards,
Wiola
Wiola,
I've had the same PAL error when working with monthly data from '10 to '12 year.
The problem has gone when I added some observations - saying, I started not from '10, but from '09.
As I understand, TESM algorithm needs some "training" data to learn on (in my case, I don't have any predictions for '09 year).
Also, it's always referenced as on-the-year algo. I mean, CYCLE is "The length of a cycle (L > 1). For example, the cycle of quarterly data is 4, and the cycle of monthly data is 12" - from PAL Reference. In your case, 6445 may be even insufficient, imho... Have you tried giving MORE?
Best regards,
Oleg
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.
Dear Oleg,
Thank you for your message.
Unfortunately, in this particular case I can’t give more data… because I don’t have it. I’ve tried to do what you suggested and start prediction later (to give TESM algorithm some training data as you said) but it returns the same error.
Strange is that when I use LESS data in my procedure everything is fine, adding more data than exactly 4267 rows results with an error...
I really don't know what's wrong
Kind regards,
Wiola
Hi Wiola,
Are you using time hierarchy or having the system presume the time interval used!
My explanation would be generic ... [testing & training the dataset is not required in time series;]
for EG: if you have 24 lines and give the cycle as 2 then it takes that each row rep a 6month's aggregation; and if you provide 12 then it takes them as monthly value for the same data.
So in your case when you use hourly data, your should have only 24 rows for a day i.e if you use time hierarchy/ aggregated; else if you have only 8 hrs data for a day; do not define time hier, but just sort the rows and define cycle; can you just give the below details:
How is time defined
what does each of the row measure rep;
do you have more rows rep for the same hour [like 10:00, 10:05, 10:45 etc]
Regards,
Arvind E
Hi Arvind,
in my PAL code i don't let HANA to presume the time interval, like you said I just sorted rows in my table and I defined cycle (24).
Strange is that my code perfectly works with less than 4267 rows.
Maybe there is some problem with alpha, beta and gamma parameters optimization with more data... but I don't know how to manage it.
This problem only occurs when I try to use Forecast Smoothing with large amount of data, in other algorithms like Double Exponential Smoothing or Triple Exponential Smoothing everything works well. Major difference between this algorithms is alpha, beta and gamma parameters optimization. Only in Forecast Smoothing they are optimized.
Kind regards,
Wiola
User | Count |
---|---|
93 | |
10 | |
10 | |
9 | |
9 | |
7 | |
6 | |
5 | |
5 | |
4 |
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.