Self-Organizing Map and Nonlinear Autoregression Networks for Forecasting Groundwater Levels at Pingtung Plain

In this study, Pingtung Plain was the study area and the investigative data collected during 1999 and 2014 consisted of rainfall, flow, groundwater extraction and groundwater level. First, we extracted groundwater characteristics in dry seasons from historical data by using the principle component analysis (PCA). The results showed that two principal components could explain 70% of data characteristics. Besides, the analytical results on the weight distribution and scores of principal components indicated that there were distinguishable features between the eastern and the western zones of the study area. Second, we investigated the correlations between different hydrological factors and groundwater levels. The results indicated that it was difficult to explain the regional groundwater level variations based on one single hydrological factor and we only could infer rainfall was an important factor affecting groundwater level variations. Third, we used the Self-Organizing Map (SOM) to classify monthly groundwater level data for constructing groundwater level distribution maps. We found that significant differences in groundwater levels occurred around 2005 in the study area, which were consistent with the results of the PCA, and therefore we could obtain the spatial distribution of groundwater levels. Fourth, we used the Nonlinear Autoregressive with Exogenous Inputs (R-NARX) to forecast the average groundwater level in each layer of the aquifer. The results showed that the R^2 value reached as high as 0.81 at T+1 and remained around 0.70 at T+2 in the testing phrases for each zone. It demonstrated that the R-NARX could well forecast the average groundwater levels by using the feedback information of groundwater level. Finally, based on the SOM groundwater level distribution maps we interpolated and extrapolated the forecasted average groundwater levels obtained from the R-NARX to derive the groundwater level of each monitoring well in the study. We then used the Kriging method to estimate the spatial distribution of groundwater levels in the whole study area for completing the construction of the regional groundwater level forecast model, which can provide valuable information for the management of water resources.