Evaluation of shallow groundwater dynamics after water supplement in North China Plain based on attention-GRU model

• Published in: Journal of hydrology (Amsterdam) 2023-10, Vol.625, p.130085, Article 130085
• Main Authors: Nan, Tian, Cao, Wengeng, Wang, Zhe, Gao, Yuanyuan, Zhao, Lihua, Sun, Xiaoyue, Na, Jing
• Format: Article
• Language: English
• Subjects: Attention-GRU; Deep learning models; Ecological water supplement; Groundwater evaluation; North China Plain
• ISSN: 0022-1694; 1879-2707
• DOI: 10.1016/j.jhydrol.2023.130085

•Driving factors of groundwater level variation have been evaluated and ranked.•Attention-GRU model shows the best performance on groundwater level prediction.•Ecological water supplement shows variance effects in the North China Plain.•Specific yield updates exhibit a significant influence during specific period.•Groundwater storage variation presents different features in the North China Plain. The continuous ecological river supplement has enhanced the recharge of groundwater in the North China Plain, and alleviated the depletion of groundwater resources. In the study of ecological river supplement, efficient quantitative prediction of groundwater level (GWL) and groundwater storage (GWS) is a crucial issue. To predict GWL and estimate the variation of GWS in the North China Plain, this paper uses the attention-gated recurrent unit (Attention-GRU) model as a primary method and coefficient of determination (R2) and root mean square error (RMSE) as the indices to evaluate model performance. Compared to traditional recursive neural network (RNN) and long short-term memory (LSTM) model, Attention-GRU model demonstrated an outstanding performance, as the RMSE is respectively 0.5 m and 2.28 m in the training and testing periods, and the R2 is 0.87 and 0.68. With the usage of the convolution neural network (CNN) model, the specific yield field was continuously updated from 2018 to 2021. The average annual variation of GWS estimated by this method is respectively −15.19 × 104 m3, −13.29 × 104 m3, 2.55×104 m3 and 100.05×104 m3 from 2018 to 2021 in the North China Plain. Besides, the semiannual GWS variation, calculated using the updated specific yield, is −59.56 × 104 m3 in 2022. Furthermore, the influence of the reference time length and the driving factors were evaluated. Results show that the precipitation, evaporation and human activities are the most important temporal features influencing the changes in GWS, and the changes in different areas have varying local characteristics. In the piedmont plain, river ecological recharge played a crucial role for groundwater storage recovery. When it moved to the coastal plain, the contribution rate of recharge decreased from 23% to 11%. In addition, the optimal reference time length used in Attention-GRU model varies in different hydrogeology divisions, and the specific yield variation would result in substantial differences for the GWS estimation under different climate conditions and human activities. The systemati