基于人工神经网络模型的地下水水位动态变化模拟

Dynamic Variation Simulation of Ground Water Table Based on Artificial Neural Network Model

地下水水位的预测在流域地表水和地下水资源的综合规划管理中起着非常重要的作用。在该研究中,人工神经网络模型被应用于希尼尔水库周边地下水水位的预测中。采用研究区6口地下水观测井资料,用人工神经网络模型进行模拟预测1周后的地下水水位。模型输入因子包括此前1周蒸发量、水库水位、排渠水位、抽水量和观测井地下水位,因此模型有15个输入节点和6个输出节点。将3种不同的神经网络训练算法,即自适应学习速率动量梯度下降反向传播算法(GDX)、LM算法和贝叶斯正则化算法(BR)用于地下水水位预测,并对模拟结果进行了评估。结果表明:3种神经网络训练算法在研究区地下水水位预测中表现均较好。然而,BR算法的性能总体略优于GDX和LM算法。将BR算法训练的人工神经网络模型用于预测研究区未来2、3和4周的地下水水位,虽然地下水位预测的准确性随着时间的增加有所降低,但模拟效果仍然较好。

Predication of the ground water table plays an important role in planning management of catchement surface and ground water resources.In this study, the artificial neural network model is applied in predication of the ground water table around the Xinier reser-voir.By application of data from 6 monitoring wells in the study area and of the artificial neural network model, the ground water table af-ter one week is predicated by simulation.The factors input the model include evaporation, reservoir level, escape canal level, water pumped volume and ground water table of the monitoring wells in last week.Therefore, the model is with 15 input points and 6 output points.Three different neural network methods of GDX, LM and BR methods are applied for the predication of the ground water table. The study shows that all three methods perform well in the predication.Generally, BR performance is better than these of GDX and LM. The artificial neural network model trained by BR method is applied for the predication of the ground water table in future 2nd, 3rd and 4th weeks in the study area.The simulation results are still better although the accuracy of the predication of the ground water table slight-ly decreases with time increment.