摘要
针对传统地下水位预测模型未考虑相关影响因素而导致的预测精度较低的问题,考虑多源因素如气象数据和归一化植被指数(NDVI)等对地下水位的影响,提出基于注意力机制的多因素长短期记忆(M-LSTM)神经网络模型,旨在通过融合多个关键因素提高预测准确性。算法的核心思想在于改造经典的长短期记忆(LSTM)网络输入端,使之能够学习多因素的数据特征,并在LSTM层之间加入注意力机制,使其能够更好地学习并强调多因素数据中的时间特征。以山西省大同市作为研究区域进行预测,输入数据集包括2015—2020年的逐月植被指数(NDVI)数据、地下水位观测数据以及气象数据,并进行嵌套交叉验证,以均方根误差(RMSE)作为性能评价指标。实验结果表明,引入注意力机制的M-LSTM模型能够有效预测地下水位,其最小均方根误差(RMSE)仅为0.4367,精度较高。
Addressing the issue of low prediction accuracy in traditional groundwater level forecasting models due to a lack of consideration for relevant influencing factors,this study incorporates multiple sources such as meteorological data and the Normalized Difference Vegetation Index(NDVI)to examine their impact on groundwater levels.A multi-factor Long Short-Term Memory(M-LSTM)neural network model with an attention mechanism is proposed to enhance prediction accuracy by integrating multiple key factors.The core idea of the algorithm is to modify the input end of the classic LSTM network to enable it to learn features from multi-factor data and to include an attention mechanism between LSTM layers to better learn and emphasize the temporal features within the multi-factor data.Datong,Shanxi Province,serves as the study area was predicted with a dataset comprising monthly NDVI data,groundwater level observations,and meteorological data from 2015 to 2020,using nested cross-validation with root mean square error(RMSE)as the performance evaluation metric.The experimental results demonstrate that the M-LSTM model with introducing attention mechanism can effectively predict groundwater levels,achieving a minimum RMSE of 0.4367 and high precision.
作者
刘小蝶
张红月
芮小平
孙文
LIU Xiaodie;ZHANG Hongyue;RUI Xiaoping;SUN Wen(School of Earth Sciences and Engineering,Hohai University,Nanjing 211100,China;Shanxi Hydrology and Water Resources Survey Station,Taiyuan 030001,China)
出处
《水文》
北大核心
2025年第2期73-79,共7页
Journal of China Hydrology
基金
国家自然科学基金项目(42376180)
山西省水利科学技术研究与推广项目(2023ZF17)。
