Groundwater dynamics clustering and prediction based on grey relational analysis and LSTM model: A case study in Beijing Plain, China

Abstract

Study region

Beijing Plain, China

Study focus

The traditional zoning of groundwater systems relies on lithological characteristics and hydraulic connections. However, the influence of climate change and human activities has led to the emergence of diverse dynamic patterns within these systems. Additionally, the differences in groundwater dynamic characteristics are often ignored when modeling with machine learning methods. In this study, a new clustering method named GRA-CLU was proposed to classify dynamic types of groundwater. Then, regional models are built using LSTM based on the dynamic zoning, aiming to incorporate hydrogeological significance into the pure data-driven model.

New hydrological insights for the region: The GRA-CLU method classified the study area into six groundwater dynamic types. Compared with the traditional hydrogeological units, the new zoning result identified two dynamic types that have never been considered: urban influence zone and mountain-plain junction zone where surface water interacts with groundwater. Through analysis, the significant rise of groundwater levels in 2021 was more influenced by heavy rainfall events rather than human activities. This study further explored the performance of LSTM regional models based on the six dynamic zones. The results indicated that the NSE of models increased by 5.7–50.0 %, the improvement was more obvious in zones with irregular fluctuations. The RMSE decreased by 31.5–59.8 %, particularly noticeable in zones with regular annual fluctuations and large samples.