Deep Learning Models for Seasonal Groundwater Level Prediction

The development of purely data-driven approaches for groundwater level prediction is crucial for sustainable groundwater management, offering the ability to predict groundwater levels across numerous monitoring wells and large geographical regions. Especially in arid regions, groundwater resources are under pressure, as seen in areas like Brandenburg, Germany, which is characterized as the driest federal state with the highest number of lakes. Here, data-driven approaches can enable fast and accurate seasonal groundwater level predictions, supporting local authorities in managing sustainable utilization.

Unlike traditional numerical models, which are computationally expensive and require complex parameterization when applied to large geographical areas, data-driven models provide a scalable solution. Recent studies have demonstrated the potential of machine learning (ML) approaches, particularly different deep neural network architectures, to provide accurate groundwater level predictions. In these studies, recurrent neural networks, such as Long Short-Term Memory networks (LSTMs), as well as recently developed architectures like the Temporal Fusion Transformer (TFT), which combines LSTMs with the self-attention mechanisms, and Neural Hierarchical Interpolation for Time Series Forecasting (N-HiTS), which is a time-series decomposition algorithm based on multilayer perceptrons (MLPs), have been used. Another recently developed architecture, the Time-series Dense Encoder (TiDE), which is based on MLPs and residual blocks, has further expanded the toolkit for time-series prediction.

In this study, we evaluate and compare the performance of four deep learning (DL) architectures (LSTM, TFT, N-HiTS, and TiDE) in predicting groundwater levels up to 16 ahead, using a wealth of spatial and temporal information for over 1,000 monitoring wells across Brandenburg. Input features to our models include historical groundwater level measurements, climatic variables, and static physical characteristics, such as groundwater recharge and land cover. Our analysis identifies the environmental conditions under which these models achieve a good predictive performance accuracy and assesses their ability to capture varying groundwater dynamics, thereby testing their alignment with hydrogeological system understanding. Furthermore, we assess whether the static features enhance the models performance and facilitate generalization across monitoring wells with similar static features levels, which we test through ablation studies and spatial out-of-sample cross-validation.

Our findings provide valuable insights into the strengths and limitations of different DL architectures for groundwater level prediction, highlighting their potential to support sustainable groundwater management in regions facing water scarcity.