dc.description.abstract | To respond to climate change and urbanization, water management systems will need to
adapt in the next decades all over the world, including the Netherlands. Hydrological modelling and
the simplification of real-world processes are vital for managing water resources and systems. In the
future decades, machine learning (ML), deep learning (DL), and neural networks (NN) are projected to
be critical in supporting humans in handling increasing volumes and diversity of data, extracting
relevant information for a specific variable, and offering viable answers to crucial issues. Numerous
articles have showed over the last decade that ML can help hydrologists to model transdisciplinary
and complex systems that are challenging to simulate using standard numerical modelling methods.
Machine learning and neural networks are becoming essential tools for hydrological analysis since
they allow us to handle large amounts of data and extract significant and hidden information, as well
as correlations between hydrological variables. The objective of this study is to enhance the
performance and prediction skill of an existing groundwater level model by evaluating the impact and
relevance of ML model selection and input datasets. For this purpose, a process-based ML approach
was implemented, using the National Hydrological Model for physical consistency along with different
types of input features including meteorological, hydrological, and environmental variables. The
findings reveal that both applied methods are capable of predicting groundwater levels and boosting
the numerical model's capabilities. To better represent and visualize these results a groundwater map
was created for average summer conditions in 250m resolution for the whole area of the Netherlands.
Furthermore, in order to facilitate future groundwater management and research, the feature
importance was evaluated in various situations to examine the overall picture of variable relevance.
The estimated feature importance values and the model’s error results were further examined to
determine whether there are any spatial pattern or trend in the outcomes. From this, it can be
concluded that the model is suitable for modelling typical groundwater levels, but it suffers from
significant error when predicting groundwater extremes, despite the fact that the errors and results
are still more closely related to actual groundwater levels than the numerical model's results. As a
result, the approach works poorer in the southern areas of the Netherlands, such as Limburg and
Maastricht. Additionally, a model was also conducted to explore if the difference between the
numerical model's outputs and real groundwater levels could be estimated. Different scenarios were
investigated, and a generic, simplified model was developed which can predict the errors between
simulation values and actual groundwater observations with an adequate accuracy. This simplified
model might help to model hydrological and environmental processes, since by using this model
groundwater level predictions can be generated without knowing any actual groundwater level values.
In summary, the work includes a detailed description of methodology, demonstrating the required
steps in creating a machine learning model that can predict hydrological processes. The findings can
be used in future study to improve groundwater level predictions and, as a result, water management
strategies in order to reduce the detrimental effects of future groundwater level extremes that could
result in severe droughts or floods. | |