Analysis of forecast shortterm groundwater level data different geological formations in a coastal aquifer using ANN And WNN

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.