Problem while developing a multivariate Regression model using neural network

Dear all,

I am trying to develop a multivariate regression model to predict some variable x which is a function of inputs such as, universal time (UT), latitude, longitude etc. I have used a feedforward network with one input layer, one hidden layer (40 neurons) and an output layer. I have used tansig as the activation function. I have completed the training and currently testing the network. I am facing a problem with the network.

At the boundaries of the UT, the values predicted by the model are not matching. I could see a clear 'jump' between 23.75 UT and 0UT. But, my data doesn't have any jump. I have checked with different data sets having the diurnal variation and I am facing the same issue. Why did the model fail to predict the values at the boundaries?

I didn't understand this problem clearly. Is the periodicity (means data repeat every 24 hours) of data causing the issue?

Kindly help in this regards.

Thanks in advance.

[Greg Heath]

Why can't you just unwrap UT to get a linear variable?

Greg

[gowtham sai]

I can't do that because the data has the variation with the time (UT). All the variables that I have mentioned earlier have a significant influence on the output. The current NN training has captured the variation with UT except at the boundaries.

[Nikhil Negi]

like greg said you should convert the UT into linear time and transform the data accordingly and also i think you should normalize all the variables in case you have not.

[gowtham sai]

@ Greg and @Nikhil

I have already normalized the data.

By the way, how to convert the UT into liner time? Could you please elaborate?