Abstract:
Forecasting the fluctuations in groundwater levels in advance with satisfactory
accuracy will help in conjunctive use and proper planning and management of
groundwater resources in a sustainable manner. The main objective of this paper is to
forecast the groundwater levels in different geological formations of Khondalite, Sand
stone, Crystalline and Alluvium in a coastal aquifer of East Godavari district in Andhra
Pradesh using wavelet neural networks (WNN). Six hourly of available data, monthly
groundwater levels and rainfall data of 44 piezometers of 10 years (2002 to 2011) were
used. From the monthly groundwater level contours, it has been observed that average
fluctuation of groundwater level is 5-10 m for agency area, 10-15 m for upland area and
2-5 m for delta areas. The predictions were made for 6, 12, and 24 hours, weekly and one
month interval using 6 hourly piezometric data. Using monthly data, predictions were
made for 1, 2 and 3 months. The input data for WNN models are antecedent groundwater
levels and rainfall. Output is the grounwater level to be forecasted at one step advance.
The number of hidden neurons was arrived at by trial and error. The model consists of
Daubechies mother wavelet of order 5 (db5) at level 3 and was run with LM algorithm
to train the neural network structure. The model performance was monitored with
different statistics. The forecasts were made in three cases namely 1. One month
advance for district average, upland area and Prathipadu station 2. Six hourly forecasts
3. Monthly forecasts were made for various piezometers in different formations. From
first case, it was observed that the performance of WNN was much better than ANN
model in the prediction of groundwater levels. It may be due to the inputs used in
wavelet based models are decomposed sub time series, Second case 6-hourly data was
given an acceptable accuracy upto 1-day (24 hours) forecast and for 1-week and 1-
month it gives poor result due to longer duration of forecast with shorter interval data.
Third case observation wells in Khondalite, crystalline and sandstone have better
forecast upto 3-months, whereas alluvium formations show the poor result. It may be
due to the variation in hydrogeology and characteristics of these geological formations.
