# data prediction/forecast

#### Do you have a question? Post it now! No Registration Necessary

Translate This Thread From English to

•  Subject
• Author
• Posted on
Hi all,

I am looking for an algorithm (If there are a one.) to predict/forecas
for next data, according to the history data.

I have a slow reponse sensor, I would like to find a math method t
predict the final value according to the history value, and then I ca
show the predict value (which close to the final value) to user earlier.

For example: The calibration value for the instrument is 100.
My sensor will measurment in 5sec interval and give the data below.

50
51
52
..
..
..(10min later)
80
81
82
..(more than 1 hour later)
97
97
97
98
98
98
98
100 (It takes more than 1 hour to come to the final value.)

I want to predict the final value, according to the data in first 10min o
more, depends on the algorithm requirement. Would you help me to figure ou
where can I find this kinds of algorithm?

have a good day
Ken

Re: data prediction/forecast

It would be a big advantage to know what kind of the dependence is
expected. If you have a mathematic model of the sensor then you can try
to estimate the parameters of the model which make the best fit with
your data. This is what is called Proni method.

If the dependence is supposed to be linear, then you can apply Kalman or
Viener filtering or autoregression.

If nothing is known, then you can try to build a polynomial approximation.

DSP and Mixed Signal Design Consultant

http://www.abvolt.com

Re: data prediction/forecast

There is no such thing as a 'generic prediction algorithm'.

As a basis, you need to know the characteristics of the typical response
curve over time. This would be in the form of some sort of expression
whose actual values are dependant on some initial values/samples.

Exactly how you derive this typical response depends on what you're
doing. Are you able to use previous history (ie. many many hours) in
some sort of long-term adaptive algorithm in the device itself? Or can
you derive a theoretical response (from physics and basic principles)
that is valid for *all* sensors and just dependant on samples for the
1st 10 minutes?

I'm guessing you need to record a large number of observations (over
many hours) and fit a response curve to the sample data and derive the
parameters for an algorithm from that, which you can subsequently

Regards,
Mark

Re: data prediction/forecast

Do you mean it reaches (approaches) a steady state? If this is like
heating a mass or storing a charge, the form is:

f(t) = A + B( 1 + exp(-kt) )

You can find better and better approximations for the factors as time goes by,
if this is a one time event. Otherwise, calibrate.

Correction

f(t) = A + B( 1 - exp(-kt) )