How can Backpropagation neural network could be used to solve higher order non-linear system and to make it online?

Dear Sir or Madam;

I am a graduate student in Physics from NEPAL. I am going to study the effects of soil moisture content and its prediction using 15 minute remotely sensed data. I am going to design the prediction based on neural network, as it is said that it is good on forecasting. However, I need to test and validate the instantaneous data from the sensors. I need to do some experimental work. I need to know how can I used neural network model in my case?

I found in some literature that people use multi layer backpropagation neural network for the static design. And then, they performed sliding window algorithm or accumulated training algorithm to make the system online? Could you provide me some idea and if some sample of matlab how I could use sliding window algorithm or accumulated training and generalization of the neural network.

I have sample of remotely sensed data of 15 minute duration of 6 months and total sample data is 17,280. My input data are: soil temperature, air temperature, precipitation, soil adjusted vegetation index and land surface temperature. My output data is: soil moisture content.

Also, the effect of soil moisture content is polynomial equation based on the emphirical physical modeling.

I will be really appreciated for your valuable suggestion. Muna Adhikari

[Muna Adhikari]

Dear Greg,

The data is collected from the data acquistion system of 6 months duration. (Remotely sensed means it is transmitted from remote distance). Although from the sensors, the data is taken in (fraction of seconds), all the data are averaged for 15 minutes interval. It means that my data have a sampling interval of 15 minutes for 6 months duration. Yes, the data is averaged for every 15 minutes and the data is collection for 6 months duration.

What does it mean uniformly spaced in time? Does it mean the data is uniform i.e. 1Feb 12am, 12:15 am, 12:30 am and continuous. If it is, it is continuous.

I want online system. Can we perform the static neural network model and make it online using accumulated training or sliding window training system? OR Else i need to choose time series predictions ? I am not familar on choice of selection of this two methods...But what i want is after the generalization it should be the neural network should be adaptive.

Thanks.

Answer 1

[Greg Heath]

You have 5 input variables and 1 output variable.

Please explain what you mean by

"I have sample of remotely sensed data of 15 minute duration of 6 months ".

What is the sampling rate?

How many measurements per minute for each variable?

Is the data averaged ecvery 15 minutes?

Is the data collected continuously for 6 months?

Is the data uniformly spaced in time?

Do you want static predictions based on batches of 15 minute averages or do you want a time series of predictions based on a sliding 15-minute window?