Adaptive spatiotemporal deep learning with transfer learning for multi-well concentration prediction in in-situ...

Accurate prediction of state variables in multi-well pumping systems (MWPS) is critical for real-time management of groundwater-related operations, yet remains challenging under complex hydrogeological conditions and limited observational data. This study presents a lightweight CNN-LSTM model integrated with transfer learning for short-term but high-accuracy prediction of MWPS dynamics. The framework is evaluated using field data from an in-situ uranium leaching aquifer (China), where pumping rates from 42 wells are used to predict the uranium concentrations in the mixed solution recovered from 20 extraction wells. The proposed model contains fewer than 1000 weights and can be robustly trained using 120 pairs of historical observations. Model performance degrades with increasing prediction time when relying solely on pre-trained weights. Incorporating new observations through transfer learning significantly improves prediction accuracy. It is found that fine tuning the weights in LSTM layer adaptive to new observations yields better prediction of uranium concentration than those models updating the weights only in CNN layers. This indicates the dominant role of temporal dynamics, such as time-varying pumping rates and progressive uraninite depletion, in controlling the variation of system response features in comparison to the spatial factors including the well placement and subsurface heterogeneity. Prediction accuracy can be further improved under shorter update intervals and longer input sequences, while prediction periods ranging from 1 day to 1 week have minimal influence on the model performance. The proposed framework provides an efficient and adaptive approach for high-accuracy prediction in data-limited MWPS, with potential applications in in-situ leaching, groundwater remediation, and managed aquifer systems.