Using back-propagation neural network to estimate groundwater level and pumping quantity

Due to the densely population and uneven rainfall in Taiwan, the management of water resources is an important issue, especially the excessive extraction of groundwater situation which may cause land subsidence problem. Traditional hydrological analysis methods commonly include some assumptions that are difficult to conform to the actual situation, and the neural network model can construct a relational model that is different from the traditional one. Therefore, this study collects the hydrological observation data and develops a supervised back propagation neural network model. By predicting the groundwater level and estimating the pumping quantity, the results can provide government to effectively manage future groundwater resources. This study includes three parts: data supplements, groundwater level predictions, and groundwater pumping estimations. First, this study uses the neural network model to supply the loss data, and selects the Zhongshan groundwater monitoring well in Taichung area with small human influence as an example. The meteorological data and the observed groundwater level data are adopted, and a certain section of the continuous observation data is removed to do supplement and compare with the observations. The result shows that the supplement data using neural network interpolation obtains better results than using the inverse distance interpolation method. The same concept is adopted to estimate the groundwater level. The result shows that the neural network model well predicts the groundwater level, though still includes some discrepancies. The estimated groundwater level variations have accumulative errors, but the result looks reasonably. After realizing the correlation between climatic hydrological data and groundwater level data, this study continued to build a neural network model based on the observed pumping data of a well group and the groundwater level monitoring data from a nearest well in southern Taiwan. The pumping quantity is estimated by using the groundwater level data, and considering if the rainfall be the input data or not to compare the advantages and disadvantages of these two situations. The results show that the pumping quantity estimated by the neural network model can capture the pattern but have a discrepancy with the observed pumping quantity, which may because the monitoring well is located in the cone of depression of the pumping well group. Therefore, it is difficult to establish a stable pumping quantity estimation model with climate conditions. The study results provide the tool for groundwater level supplements and predictions and groundwater pumping assessments in the study area, which can be the reference in future planning and management of groundwater resources.