hi roshni,

take time seires of groundwater level ,take probably last 20 years data, the
well or any other site frm which the groundwater level has been measured , does
it gives the correct indication of waterlevel for that rigion, do u check it or
do u have any reference for it,if its correct then divide groundwater level time
seried into 3 subsets on the basis of marked fluctuation, then used first subset
for training second for calibration and third for forecasting.


roshni_balan <roshni_balan@...> wrote:

Sir,

I am doing my M.Tech at IIT Kharagpur. M.Tech project is forecasting
watertable fluctuation using ANN.


I am taking monthly rainfall, previous month water table depth,
monthly river stage and monthly average temperature as inputs. I
tried with trainbr algorithm. I am not changing the parameters, only
I am changing the number of epochs, since they are taking the
default as 100. I am scaling the data in to 0-1 domain. Can I use
the same algorithm for 1-, 2-, 3- ....month ahead up to 1 year ahead
forecasting. According to Daliakopoulos et al., (2005), feedforward
NN with LM algorithm provides best results up to 18months forecast
for an area in which groundwater has been steadily decreasing due to
over exploitation. Same situation is in my study area (Konan aquifer
in Japan) also. I have seven years of data (1998-2004). I considered
1998-2001 for training and 2002-2003 for validation. Sir, I thought
that, If I have the data up to 2004, I can forecast for 2005 Jan,
Feb, etc. So for training the data from 1998-2001, for 1 year ahead
forecasting, what i have to give the target values for training? Is
it the same groundwater values from 1998-2001 or from 1999-2002.

Expecting a reply
Thanking you
Roshni T.