2:30 PM - 2:45 PM
[AHW17-04] The Development of a Configuration-flexible Artificial Neuro Network Model for Groundwater Modeling
Keywords:groundwater numerical simulation, artificial neural network, MODFLOW, linear superposition
Artificial neural networks (ANN) have been widely applied to different engineering disciplines including groundwater modeling, and the ANN models can have good performance for simulation or prediction. However, the configuration of a trained network cannot be changed during the prediction stage. This configuration-fixed restriction limits the application of the ANN model on many of the planning or design problems. The network configuration, for example, can be the numbers and locations of pumping or observation wells. A pumping or monitoring network design is to define the best network and can be an example to illustrate this issue. For a network design problem, the experiment of different networks, and different configurations, is necessary during the search process and a conventional ANN model is difficult to apply.
This study aims to develop a configuration-flexible ANN model (CF-ANN) for groundwater modeling. The CF-ANN model is an integration of various ANN components using the principle of linear superposition. Each ANN component corresponds to the influence of pumping or other boundary conditions alone on the water level. This study has conducted linear superposition tests, to confirm the influences of pumping, recharge, and river interaction can be linearly superimposed indeed. However, the influence of specified head boundary conditions on water levels cannot be directly linearly superimposed. Therefore, in this study, a linear approximation method was developed for specified head boundary conditions. This allows the proposed CF-ANN model can consider all types of boundary conditions and be a universal model applicable to various problem types.
The training, validation, and testing datasets for the various neural network components, were generated using the MODFLOW model under various hypothetical scenarios. The simulation results demonstrate that the novel CF-ANN model can rapidly and effectively simulate groundwater level responses with various network configurations. This model can be applied to real-time groundwater management problems and is flexible enough to be adapted to configuration-search problems such as network design.
This study aims to develop a configuration-flexible ANN model (CF-ANN) for groundwater modeling. The CF-ANN model is an integration of various ANN components using the principle of linear superposition. Each ANN component corresponds to the influence of pumping or other boundary conditions alone on the water level. This study has conducted linear superposition tests, to confirm the influences of pumping, recharge, and river interaction can be linearly superimposed indeed. However, the influence of specified head boundary conditions on water levels cannot be directly linearly superimposed. Therefore, in this study, a linear approximation method was developed for specified head boundary conditions. This allows the proposed CF-ANN model can consider all types of boundary conditions and be a universal model applicable to various problem types.
The training, validation, and testing datasets for the various neural network components, were generated using the MODFLOW model under various hypothetical scenarios. The simulation results demonstrate that the novel CF-ANN model can rapidly and effectively simulate groundwater level responses with various network configurations. This model can be applied to real-time groundwater management problems and is flexible enough to be adapted to configuration-search problems such as network design.