Research on groundwater level prediction of northern karst spring of China based on LSTM-Attention neural network

doi:10.13745/j.esf.sf.2025.10.1

Abstract

Abstract:

Karst groundwater serves as a critical water source in northern karst regions of China, and accurately forecasting its water level dynamics is essential for the scientific management and conservation of groundwater resources. However, the inherent strong heterogeneity and anisotropy of karst aquifer systems often result in non-stationary and nonlinear groundwater fluctuations, posing significant challenges for reliable prediction and frequently leading to considerable errors. This study proposes a multivariate groundwater level forecasting model for the Baotu Spring, based on the integration of an attention mechanism and a long short-term memory (LSTM) neural network. The model is trained and validated using daily data from 2013 to 2024, including precipitation (as a proxy for recharge), vapor pressure, air temperature, and groundwater extraction volume (as indicators of discharge). The main findings are as follows: (a) The BEAST (Bayesian Estimator of Abrupt Change, Seasonality, and Trend) algorithm was applied to the 1958-2024 Baotu Spring groundwater level time series, identifying four abrupt change points. Based on these, the time series was segmented into four distinct stages. (b) Cross-correlation analysis revealed a time lag of approximately 2-3 months between precipitation and groundwater level variations, suggesting a strong dynamic coupling between the two variables. (c) The proposed model incorporates multiple input variables (precipitation, vapor pressure, temperature, and extraction volume). The interplay among these variables enhances model interpretability and contributes to improved forecasting accuracy. (d) Daily temperature data were smoothed using a sinusoidal fitting function, which effectively reduces measurement noise and improves prediction performance. (e) Compared to standard LSTM and gated recurrent unit (GRU) models, the LSTM-Attention model demonstrates superior predictive capability. By leveraging the attention mechanism to prioritize influential features, it achieved a root mean square error (RMSE) of 0.13 m and a coefficient of determination (R²) of 0.94. In summary, the LSTM-Attention-based forecasting model exhibits robust accuracy and stability, offering valuable insights into the precise prediction of groundwater dynamics in karst environments.

Key words: northern karst, prediction of groundwater dynamics, multivariate modeling, LSTM-Attention neural network

CLC Number:

P641.2
TP183

HUANG Linxian, XU Zhenghe, ZHI Chuanshun, LI Shuang, LIU Zhizheng, XING Liting, ZHU Henghua, WANG Xiaowei, BI Wenwen, HU Xiaonong. Research on groundwater level prediction of northern karst spring of China based on LSTM-Attention neural network[J]. Earth Science Frontiers, 2026, 33(1): 419-431.

Figures/Tables 15

Fig.1 Hydrogeological map of Jinan Spring area. Modified after [19].

Fig.2 Geological profile of II'. Adapted from [19].

Fig.3 Time series decomposition and abrupt change detection in Baotu Spring groundwater levels

Fig.4 Comparison of daily groundwater level and rainfall in Baotu Spring from 2013 to 2024

Fig.5 Cross-correlation analysis between rainfall and groundwater level in Baotu Spring

Fig.6 Raw data and sinusoidal fitting curve of daily average temperature from 2013 to 2024

Fig.7 Structure of LSTM neural network

Fig.8 Self-attention mechanism structure diagram

Fig.9 Framework of LSTM_Attention prediction model

Fig.10 Prediction results of LSTM_Attention model (with fitted temperature)

Fig.11 Spring groundwater level prediction results of raw and fitted temperature

Fig.12 Prediction error of raw and fitted temperature(RMSE和R2)

Fig.13 Prediction results of Baotu Spring groundwater level using different methods

Fig.14 Prediction errors of different methods(RMSE和R2)

Fig.15 Violin plot of forecast error

References 29

[1]	管清花, 李福林, 陈学群, 等. 济南趵突泉泉域泉群生态基流量研究[J]. 中国农村水利水电, 2021(4): 75-80, 91.
[2]	邢立亭, 周娟, 宋广增, 等. 济南四大泉群泉水补给来源混合比探讨[J]. 地学前缘, 2018, 25(3): 260-272. DOI
[3]	袁学圣, 邢立亭, 赵振华, 等. 济南四大泉群流量衰减过程及其指示意义[J]. 干旱区资源与环境, 2022, 36(7): 126-132.
[4]	孙虹洁, 赵振华, 黄林显, 等. 多变量LSTM神经网络模型在地下水位预测中的应用[J]. 人民黄河, 2022, 44(8): 69-75.
[5]	WUNSCH A, LIESCH T, CINKUS G, et al. Karst spring discharge modeling based on deep learning using spatially distributed input data[J]. Hydrology and Earth System Sciences, 2022, 26(9): 2405-2430. DOI URL
[6]	周长松, 邹胜章, 冯启言, 等. 岩溶关键带水文地球化学研究进展[J]. 地学前缘, 2022, 29(3): 37-50. DOI
[7]	陈发家, 肖琼, 胡祥云, 等. 典型岩溶小流域碳酸盐岩风化过程及其碳汇效应[J]. 地学前缘, 2024, 31(5): 449-459. DOI
[8]	张郑贤, 刘艺, 张锋贤. 基于时间序列模型的济南趵突泉地下水位预测[J]. 中国水利水电科学研究院学报, 2019, 17(1): 51-59.
[9]	姜宝良, 陈宁宁, 李小建, 等. 河南某大型裂隙岩溶水源地地下水位动态分析[J]. 水文地质工程地质, 2021, 48(2): 37-43.
[10]	肖竞, 万军伟, 成建梅, 等. MODFLOW-CFPv2模型在岩溶隧道突涌水及对地下水环境影响中的应用:以云南鹤庆锰矿沟岩溶水系统为例[J]. 地质科技通报, 2024, 43(3): 301-310.
[11]	李志超, 姜宝良, 潘登, 等. 灰色理论在新乡百泉泉水流量动态分析中的应用[J]. 水文地质工程地质, 2023, 50(2): 34-43.
[12]	RAHBAR A, MIRARABI A, NAKHAEI M, et al. A comparative analysis of data-driven models (SVR, ANFIS, and ANNs) for daily karst spring discharge prediction[J]. Water Resources Management, 2022, 36(2): 589-609. DOI
[13]	许亮, 郭高轩. 典型北方山前岩溶泉历史流量序列重建研究[J]. 水文, 2023, 43(3): 88-92.
[14]	齐欢. 济南市趵突泉与白泉地下水位相关性研究[J]. 水文, 2020, 40(4): 79-84, 32.
[15]	LI J, LU W, LUO J. Groundwater contamination sources identification based on the long-short term memory network[J]. Journal of Hydrology, 2021, 601: 126670. DOI URL
[16]	HOCHREITER S, SCHMIDHUBER J. Long short-term memory[J]. Neural Computation, 1997, 9(8): 1735-1780. DOI PMID
[17]	WU X, WANG H Y, SHI P, et al. Long short-term memory model-a deep learning approach for medical data with irregularity in cancer prediction with tumor markers[J]. Computers in Biology and Medicine, 2022, 144: 105362. DOI URL
[18]	WUNSCH A, LIESCH T, BRODA S. Groundwater level forecasting with artificial neural networks: a comparison of long short-term memory (LSTM), convolutional neural networks (CNNs), and non-linear autoregressive networks with exogenous input (NARX)[J]. Hydrology and Earth System Sciences, 2021, 25(3): 1671-1687. DOI URL
[19]	于苗, 邢立亭, 吴吉春, 等. 基于时间序列分形的济南岩溶大泉动态研究[J]. 地质学报, 2020, 94(8): 2509-2519.
[20]	孙斌, 彭玉明. 济南泉域边界条件、水循环特征及水环境问题[J]. 中国岩溶, 2014, 33(3): 272-279.
[21]	侯新宇, 邢立亭, 孙蓓蓓, 等. 济南市岩溶水系统分级及市区与东西郊的水力联系[J]. 济南大学学报(自然科学版), 2014(28):300-305.
[22]	ZHAO K, WULDER M A, HU T, et al. Detecting change-point, trend, and seasonality in satellite time series data to track abrupt changes and nonlinear dynamics: A Bayesian ensemble algorithm[J]. Remote sensing of Environment, 2019, 232: 111181. DOI URL
[23]	迟光耀, 邢立亭, 侯新宇, 等. 基于小波分析与Mann-Kendall法的岩溶大泉动态研究[J]. 中国岩溶, 2018, 37(4):515-526.
[24]	MA S, DING W, ZHENG Y, et al. Edge-cloud collaboration-driven predictive planning based on LSTM-attention for wastewater treatment[J]. Computers & Industrial Engineering, 2024, 195: 110425. DOI URL
[25]	YANG Y, HAN L, QIU C, et al. A short-term wave energy forecasting model using two-layer decomposition and LSTM-attention[J]. Ocean Engineering, 2024, 299: 117279. DOI URL
[26]	任欢, 王旭光. 注意力机制综述[J]. 计算机应用, 2021, 41(增刊1): 1-6.
[27]	SHEWALKAR A, NYAVANANDI D, LUDWIG S A. Performance evaluation of deep neural networks applied to speech recognition: RNN, LSTM and GRU[J]. Journal of Artificial Intelligence and Soft Computing Research, 2019, 9(4): 235-245. DOI URL
[28]	SHEN G, TAN Q, ZHANG H, et al. Deep learning with gated recurrent unit networks for financial sequence predictions[J]. Procedia Computer Science, 2018, 131: 895-903. DOI URL
[29]	LI W, WU H, ZHU N, et al. Prediction of dissolved oxygen in a fishery pond based on gated recurrent unit (GRU)[J]. Information Processing in Agriculture, 2021, 8(1): 185-193. DOI URL