Impact of data partitioning in groundwater level prediction using artificial neural network for multiple wells
Jamel Seidu, Anthony Ewusi, Jerry Samuel Yaw Kuma, Yao Yevenyo Ziggah, Hans-Jürgen Voigt
Abstract
Information on groundwater level (GWL) fluctuation is very important for general planning and water resources management. In recent times, Artificial Neural Network (ANN) has gained popularity as alternative modelling tool to the existing physical-based models in groundwater level forecasting. In ANN modelling, the most widely and successfully used data partitioning approach of the hold-out cross validation has been a critical factor that affects the performance of the model. This study is focused on evaluating and comparing the impact of different data size partitioning in groundwater level prediction using ANN methods. The study considered all the possible partitioning percentages for the train-test sets which are (90–10), (80–20), (70–30), (60–40) and (50–50). These five different partitions have been analysed and tested on 13 different boreholes (BH) using Backpropagation Neural Network (BPNN), Radial Basis Function Neural Network (RBFNN), Generalized Regression Neural Network (GRNN) and Group Method of Data Handling (GMDH). The results indicated that (70–30) and (80–20) were the most dominant partitions to produce optimum GWL prediction models for the BHs. Overall, it was noticed that RBFNN was superior when (70–30) and (80–20) partitions were employed, whereas BPNN was the best for the (60–40) and (50–50) partitions. This study demonstrates that the performance of ANN method is heavily dependent on the partition percentage used in the model formulation. Therefore, the main contribution of this study was to bring to light how different data partitioning size impacts on the effectiveness of the ANN in GWL prediction.