Forcasting Groundwater Level in SAADAT-SHAHR PLAIN Using Artificial Neural Networks

A proper architectural design of Artificial Neural Network (ANN) models can provide a robust tool in water resources modeling and forecasting. The performance of different neural networks in a groundwater level forecasting was examined in order to identify an optimal ANN architecture that can provide accurate predictions up to 24 months ahead. Saadat-shahr Plain in Fars Province was chosen as the study area. Utilizing 4-month history of precipitation, temperature, runoff, and groundwater level as the input data, groundwater level at the next time step was predicted. All networks were trained for an 8 year period of data and calibrated for a 24-month period. Networks were verified based on groundwater level observations in 16 wells located in the plain. Five different types of network architectures consisting of three networks (Feed Forward Neural Network (FNN), Recurrent Neural Network (RNN), and Radial Basis Function (RBF)) and two training algorithms (Levenberg-Marquardt (LM) and error Back Propagation with momentum (BP)) were investigated and compared in terms of the prediction accuracy. Experiment results showed that the most accurate forcast (for up to 24 months ahead) is achieved with an FNN trained with LM algorithm.

Keywords:

Artificial neural network , Groundwater level forecasting , Saadat-shahr plain