Neural Network to predict temperature profile

HI

i'm trying to train a nn with following training data

Pos  fan speed   power  temp
1     60           5     31
2
3
.
.
120
1     100         5       41
2
3
.
.
.
120

the positions go from 1-120

for each i have diff temperature reading for 5,10,15...35 kw power and 60,100,120 fan speed.

so my inputs are position, fan speed, power

output is temperature

I have around 2500 data rows for training.

will this train my network as a function of the position?

Answer 1

[Walter Roberson]

You will not be able to give the NN just position and have it output temperature or some kind of temperature profile. You should, though, be able to input position, fan-speed, and power, and have it predict the temperature.

[Greg Heath]

1. Don't forget to transpose the rows to columns to get

[ I N ] = size(input) % [ 3 2500 ]

[ O N ] = size(output) % [ 1 2500 ]

2. One of several ways to rank inputs:

a. Standardize all variables (e.g. ZSCORE or MAPSTD) to zero mean and unit variance

b. Obtain the all-input I-H-O net topology as the best of multiple random weight initialization designs for each candidate value of H.

c. Repeat for each input row

i. "remove" only that row only via indexing or setting the row to zeros(1,2500)and obtain the corresponding mse

ii. Compare the performances and permanently "remove" the input row corresponding to the best performance.

Hope this helps.

Greg

P.S. Some will argue that the net should be re-tuned after a row is "removed"; others will argue that the net should be completely re-trained ... your choice.

[Chinmay Date]

Thank you so much. firstly, I do not expect a temperature output until i give all the 3 inputs. i was trying to use the neural network toolbox on matlab and since i'm new to neural networks. I cant get the inputs and targets? how do i create datasets? for inputs and targets?

[Walter Roberson]

The targets are the temperatures. The inputs are the other columns.

[Chinmay Date]

okay. i have been able to train the network using inputs and targets and the nn fitting tool. after the training i have created a test where i give 3 inputs but there is no output.. am i using the wrong toolbox? should i use patter recognition or anything else?

Answer 2

[Greg Heath]

The NNTBX default technique for regression using FITNET or FEEDFORWARDNET, is to use 70% of the data for estimating weights, 15% of the data for validation stopping during training and the remaining nontraining data for obtaining an unbiased estimate of performance.

It's hard to say what you did wrong without more details. Can you post the commands?