基于多变量LSTM神经网络的地下水位预测方法研究

Multivariable LSTM model-based groundwater level prediction

地下水位的准确预测对于黑河流域水资源可持续开发及合理利用具有重要意义。针对物理模型、数值模型参数获取困难、建模过程复杂、建模数据时间跨度较短等问题,提出一种多变量LSTM算法来准确、快速预测地下水位。通过将黑河干流中游地区1986年至2018年月均地下水位观测数据作为数据集,构建了两层每层30个神经元的LSTM模型,并对单一动态类型、混合动态类型两种数据训练方式的预测结果进行比较。结果表明:1)构建的多变量LSTM神经网络模型准确预测了地下水位的月度变化,预测精度较高,RMSE均值为0.18,R^(2)均值为0.8以上。2)相同动态类型输入的预测正确率略高于混合动态类型输入,证明区分不同类型地下水并以相同动态类型的数据作为输入,有助于提高LSTM模型在时序数据预测中的准确性。该研究实现了地下水位的快速精确预测,确定了多变量LSTM模型在预测甘肃张掖盆地地下水位的可行性,进而为张掖盆地水资源初步判断、方案决策提供方法支撑。

The accurate prediction of groundwater level is of great significance for the sustainable development and rational utilization of water resources in the Heihe River Basin. In order to solve problems as difficulty in obtaining parameters, complex modeling process, and short time span of modeling data during numerical modeling, the multivariate LSTM algorithm was used to simulate monthly changes in groundwater level, and a LSTM model, that consists two layers, and each of layer consist 30 neurons, was constructed by using long-term observation data of groundwater level in the middle reaches of the Heihe River from 1986 to 2018 as a dataset. The prediction results of two data input modes, single dynamic type and mixed dynamic type were compared. The results show that: 1)The constructed LSTM model can effectively predict the dynamic changes of groundwater level, and it fits well with the observed water level curve, with high prediction accuracy.2)The prediction accuracy of the same dynamic type input is slightly higher than that of mixed dynamic type input, proving that distinguishing different types of groundwater and using data of the same dynamic type as inputs can be helpful to improve the accuracy of deep learning models.This study has achieved rapid and accurate prediction of groundwater level, and provides a scientific basis for the development of groundwater resources in the Zhangye area.