On the challenges of global entity-aware deep learning models for groundwater level prediction.

The application of machine learning (ML) including deep learning models in hydrogeology to model and predict groundwater level in monitoring wells has gained some traction in recent years. Currently, the dominant model class is the so-called single-well model, where one model is trained for each well separately. However, recent developments in neighbouring disciplines including hydrology (rainfall–runoff modelling) have shown that global models, being able to incorporate data of several wells, may have advantages. These models are often called "entity-aware models", as they usually rely on static data to differentiate the entities, i.e. groundwater wells in hydrogeology or catchments in surface hydrology. We test two kinds of static information to characterize the groundwater wells in a global, entity-aware deep learning model set-up: first, environmental features that are continuously available and thus theoretically enable spatial generalization (regionalization), and second, time-series features that are derived from the past time series at the respective well. Moreover, we test random integer features as entity information for comparison. We use a published dataset of 108 groundwater wells in Germany, and evaluate the performance of the models in terms of Nash–Sutcliffe efficiency (NSE) in an in-sample and an out-of-sample setting, representing temporal and spatial generalization. Our results show that entity-aware models work well with a mean performance of NSE >0.8 in an in-sample setting, thus being comparable to, or even outperforming, single-well models. However, they do not generalize well spatially in an out-of-sample setting (mean NSE <0.7 , i.e. lower than a global model without entity information). Strikingly, all model variants, regardless of the type of static features used, basically perform equally well both in- and out-of-sample. The conclusion is that the model in fact does not show entity awareness, but uses static features merely as unique identifiers, raising the research question of how to properly establish entity awareness in deep learning models. Potential future avenues lie in bigger datasets, as the relatively small number of wells in the dataset might not be enough to take full advantage of global models. Also, more research is needed to find meaningful static features for ML in hydrogeology. [ABSTRACT FROM AUTHOR]