The inspection of extensive and hard-to-access sewer systems is a challenging and expensive task. As these networks age and need to comply with stricter health and environmental regulations, the demand for effective inspection solutions has increased. The introduction of technologies like CCTV (closed-circuit television) and SSET (sewer scanner and evaluation technology) marked the initial automation steps in sewer inspections. Initially, images from these technologies were manually analyzed for defects, but over time, computer vision techniques have emerged as a highly promising method for automating image processing. However, these advanced computer vision methods are computationally intensive and typically rely on cloud-based architectures, which can be costly and sometimes impractical due to energy or communication limitations. A proposed solution is to shift the computational processes from the cloud to edge computing, which can address issues related to latency and scalability.

The Sewer-ML dataset, sourced from sewer inspection videos by Danish water utilities, comprises 1.3 million images annotated across 18 defect categories. This extensive multi-label dataset serves as a foundation for training and evaluating machine learning models within sewer system management. Model performance is evaluated using the $F2_{CIW}$ (class importance weight) score, which emphasizes recall and defect severity to ensure the accurate detection of critical defects, and the $F1_{Normal}$ score, which measures the model’s accuracy in identifying instances without defects, essential for efficient resource management.

The ResNet-101 and TResNet-L models, trained on the Sewer-ML dataset, underwent various compression methods including quantization, layer fusion, and pruning. Quantization was applied only to the ResNet-101, reducing its $F2_{CIW}$ and $F1_{Normal}$ scores by 2.52\% and 0.72\% respectively, while significantly boosting inference speed by up to 95\% on standard platforms and 174.50\% on L4 GPUs. Layer fusion was also implemented, further enhancing inference efficiency. Additionally, iterative pruning was performed, showing that while the TResNet-L could maintain performance up to an 80\% pruning rate, there was a noticeable initial drop in performance for both models.

The quantized ResNet-101, both the one with and without layer fusion,
even improve compared to the standard model in regards to correctly identifying
the highest CIW defect class present in pipes deemed defective in the validation
dataset. This model behaviour is positive because the priority of a sewer asset manager is the discovery of the defect that carry the highest risk with them if not treated
in time. This improved efficiency in defect recognition helps in optimizing repair
schedules and resource allocation, thus reducing operational costs as well.

Understanding the propagation of a flood is crucial for effective emergency response measures. While traditional numerical models provide reliable flood simulations, their high computational costs pose significant limitations during emergencies. Deep learning models have recently demonstrated potential in accelerating hydrological calculations while preserving high accuracy. Although various deep learning flood models have been developed, many are limited to specific case studies or neglect the dynamic propagation of flood waves, constraining their application during emergencies. To address this, Bentivoglio et al., 2023 proposes the use of a physics-based surrogate model for spatio-temporal flood modelling; the shallow water equation graph neural network (SWE-GNN). The model demonstrates promising results on small virtual landscapes, showcasing strong generalizability to unseen breach locations and domains, while achieving computational speed ups. In this research, the real-world applicability of the SWE-GNN for time-sensitive situations is analyzed. Two dike rings in the Netherlands are selected as our case study areas. The model is trained and tested within the same domain to evaluate its application during a crisis. Performance is assessed using statistical metrics and practical evaluations, including direct and indirect damage models. The SWE-GNN model is able to correctly predict the spatio-temporal evolution of floods for unseen breach locations. The mean average errors in time are of 0.027 m and 0.029 m for water depth and of 0.007 m^2/s and 0.006 m^2/s on units discharge. The resulting flood maps prove viable for practical applicability, presenting good results for both direct as indirect damage assessment. Additionally, the SWE-GNN demonstrates a speedup of roughly 5 to 6 times for the test case areas compared to a traditional numerical model. In this project, we affirm that the SWE-GNN represents a promising innovation for a new approach to time-sensitive flood modeling, providing a reliable alternative to numerical models in situations with time constraints.

This thesis investigates the efficacy of artificial intelligence (AI) models, particularly convolutional neural networks (CNNs) and U Net architectures, in reconstructing datasets with missing velocity data in river flow analysis. Optical flow and Particle Image Velocimetry (PIV) techniques have emerged as valuable tools for analyzing river flow patterns.
Through a comprehensive literature review, CNNs, and UNet models are identified as promising tools for this task due to their ability to capture intricate patterns in datasets. The study compares the performance of the U-Net model against a statistics-based hydrological benchmark model, revealing the superior performance of the U-Net model.
Furthermore, the analysis explores how the performance of AI models varies with differing quantities of missing data, by masking available data and comparing reconstructed values against the ground truth, highlighting the importance of data availability.
Additionally, the study investigates the influence of spatial patterns in training data on model performance, including patchy versus random missing data in the field of view, simulating more datasets more likely available in reality. This clarifies the challenges encountered in predicting grid points under different training dataset conditions.
Finally, the study identifies areas within the dataset that are particularly challenging to predict, shedding light on factors contributing to prediction errors. These findings underscore the potential of AI models in hydrological applications and provide valuable insights for future research in the field.
Our findings show that U net is capable of reconstructing velocity fields from a river flow better than an average benchmark that uses the average values, with varying accuracy depending on input data.
The average benchmark model had a relative error close to 0.2 in every instance, whereas the U-Net model showed relative errors ranging from 0.085 to 0.006. Errors from a patchy mask are ranging from 0.09 8 to 0.031.

This thesis investigates the integration of algorithm unrolling and genetic algorithms (GA) for optimizing pump scheduling in water distribution systems (WDS), a critical component for ensuring energy-efficient water delivery. In the context of modern civilization’s reliance on clean, affordable water for diverse uses, the operation of a WDS, particularly through energy-intensive pumps, presents significant challenges. Traditional optimization techniques often resort to hydraulic solvers like EPANET, which, while accurate, are computationally intensive for large-scale applications. Our methodology introduces a meta-model based on algorithm unrolling, building upon prior work and extending it to address pump scheduling with a multi-objective function focusing on both cost and energy efficiency. This approach significantly reduces the computational load, offering a faster alternative to EPANET while maintaining considerable accuracy. The meta-model demonstrated promising results in the Fossolo network, achieving comparable schedules 20 times faster than traditional methods. However, its applicability to more complex networks and its ability to capture detailed system behaviors are limited, highlighting the need for further enhancements in model stability and reproducibility. Despite these limitations, the study emphasizes the potential of meta-models as a complementary tool to traditional methods, especially in scenarios requiring rapid decision-making under computational constraints. This research contributes to the broader field of water utility management, offering insights into more sustainable and efficient operation strategies.

Fully distributed hydrological models take into account the spatial variability of a catchment, and allow for assessing its hydrological response at virtually any location. However, these models can be time-consuming when it comes to model runtime and calibration, especially for large-scale catchments. Meanwhile, deep learning models have shown great potential in the field of hydrological modelling, but a multivariable, fully distributed hydrological deep learning model is still lacking. To address the aforementioned challenges associated with fully distributed models and deep learning models, we explore the possibility of developing a fully distributed multivariable deep learning model by using Graph Neural Networks (GNN), an extension of deep learning methods to non-Euclidean topologies. We develop a surrogate model of wflow_sbm, a fully distributed, physics-based hydrological model, by exploiting the similarities between its underlying functioning and GNNs. The GNN model uses the same input as wflow_sbm: gridded static parameters based on physical characteristics of the catchment and gridded dynamic meteorological forcings. The GNN model is trained to approximate wflow_sbm outputs, consisting of multiple gridded hydrological variables such as streamflow, actual evapotranspiration, subsurface flow, saturated and unsaturated groundwater storage, snow storage, and runoff. Our results show that the GNN model accurately predicts multiple hydrological variables in unseen catchments (median KGE=0.76), and can serve as an emulator of wflow_sbm with a shorter runtime. We furthermore demonstrate how the GNN model can function up to a prediction horizon of a full year, using physical system states to account for system memory, as well as a curriculum learning strategy combined with a multi-step ahead loss function during training. Overall, this study contributes to the field of fully distributed modelling using a deep learning approach.

In a world with accelerating climate change, rapid population increase and urbanization, urban water systems are under a growing stress. Thus precise short- and medium-term water demand forecasts are needed to optimize water supply operations. Water demand is influenced by human behavior and industrial activities which bring uncertainty, hence it is useful to utilize probabilistic methods to forecast water demand. This thesis provides an overview of probabilistic methods to predict water demand 24 hours ahead, highlighting their advantages, disadvantages, the accuracy of their interval and point forecasts. The case study uses the dataset from the Battle of Water Demand 2024, covering two years and two months of data across 10 districts in Ferrara, Italy, including residential, hospital, countryside, city center, and industrial districts. Three commonly used probabilistic extensions of neural networks were applied: QR (Quantile Regression), MDN (Gaussian Mixture Density Network), and an adapted CQR (Conformal Quantile Regression) method with online updating.

First, forecast models were developed to obtain probabilistic predictions. Three neural network architectures were investigated: a linear model, an MLP (multi-layer perceptron), and an LSTM (Long Short-Term Memory) model, along with a seasonal moving average as a benchmark. These neural network models were trained per district and collectively across districts. When trained per district, the linear model performed most accurately. When trained together, the MLP model performed best, but the linear model generalized the best overall, with the MLP generalizing second-best. The LSTM model had the worst performance. In districts with less heteroscedasticity in the demand pattern, the benchmark model performed on par with the neural networks in the end of the forecasting horizon, indicating that complex models are not always necessary. A categorical variable to determine the DMA did not improve the point forecasts.

Because the MLP with solely lagged features ultimately had the best performance for point forecasts, this model was used for probabilistic extensions to estimate the 0.95 prediction interval. The MLP was extended with the aforementioned probabilistic extensions. The probabilistic models were assessed in terms of reliability with where a probability was computed that tells whether the 0.95 prediction interval is reached as well as sharpness which tells how wide the interval is. Finally the Winkler Score is used that computes a trade-off between both.

Both models that fully learn the prediction interval (QR and MDN) were more difficult to calibrate, and further research is needed to calibrate them accordingly. By training these models jointly on the 10 DMAs, the coverages did vary per DMA. This can potentially be solved by training one model per DMA or by using regularization per DMA to push the model to have a similar coverage per DMA.

The MCD model had difficulty to adapt the prediction interval over the forecasting horizon, causing the coverage to reduce. The QR and MDN models also have trade off imbalances between reliability and sharpness over the forecasting horizon on the testing set, which have more random patterns. Interestingly the Conformal Prediction algorithm maintains its coverage best over the forecasting horizon and increases the sharpness, which is due to the online updating procedure. When allowing small decreases up to 0.02 probability in coverage on the testing set, the CQR model performs best according to the Winkler Score. The MDN performs best when larger drops of coverage are allowed.

Analyzing the rolling coverage over time shows improvements are possible, especially in late spring and summer periods there is under-coverage. There is also still a difference of coverage between weekdays and weekends. This indicates there is still epistemic uncertainty left to reduce, which is the uncertainty of data and model parameters. More features are recommended to reduce this which are categorical features as well as future weather data. This is recommended to investigate for non-industrial DMAs, by assuming a perfect forecast. A larger dataset may also be beneficial to obtain better performing models.

Braided rivers are among the most dynamic natural Earth systems, with a rapid and complex morphological evolution. Limited understanding and inadequate algorithm implementation of specific processes affect the accuracy of physics-based models. This leads to uncertainties that complicate the effective design of interventions and protection measures. In recent years, artificial intelligence techniques rapidly advanced and the availability of satellite imagery products increased. This research sets a novel attempt to predict the planform evolution of braided rivers with deep learning using satellite images. The study focuses on the middle and lower reaches of the Brahmaputra-Jamuna River in India and Bangladesh. We developed JamUNet, a U-Net-based convolutional neural network (CNN). The model is trained with the Global Surface Water Dataset (GSWD) to classify each pixel as "Non-water" or "Water". Four images from the same month over four consecutive years were used as input. The fifth year image served as target. JamUNet demonstrates a general capability in capturing the planform evolution. Morphological processes like meander migration, channel abandonment, and confluence and bifurcation development are often well captured. However, temporal patterns are lacking. More complex phenomena, like channel formation and channel shifting, remain unpredicted. JamUNet also tends to underpredict the total areas of erosion and deposition. Overall, JamUNet achieves a 5-6% improvement compared to the benchmark method for which no morphological change occurs in metrics such as precision, recall, F1-score, and critical success index (CSI). Among these, recall is the most meaningful metric for evaluating the model performance. JamUNet can serve as a preliminary tool for water management authorities in India and Bangladesh. It can assist in prioritising bank protection in erosion-prone areas and support land reclamation projects and inland navigation. Caution is always advised due to the model tendency to underpredict erosion. More research is required to improve the current model. Nonetheless, deep-learning modelling seems a promising field of research. Testing alternative model architectures and incorporating additional data, such as water levels or river discharge, could improve the model performance.

Understanding morphodynamic processes and structures is essential for effective river management and enhancing our knowledge of river systems. River bars can be investigated through theoretical analyses, field measurements, experimental studies, or numerical modelling. While numerical modelling offers accuracy, it is computationally intensive, making multiple simulations or parameter calibrations both time-consuming and impractical. The numerical simulations follow physical laws and equations which make the runtime significantly high making instantaneous results in times of emergencies unfeasible.
Convolutional neural networks (CNNs), a type of data-driven modelling, have been employed to study the physical parameters defined by linear stability analysis. This study utilizes Delft3D simulations to generate diverse datasets, facilitating easier access and variability. A specific type of riverbar pattern, and alternate bars were chosen for simplicity. The CNN model takes initial bed levels as inputs and provides predictions for the next step of bed level or a time series, with velocity included an additional parameter to assess its influences on the model performance. The model is able to predict the bar behaviour with R2 being 0.99. The model can predict bar suppression or migration solely based on the initial bed levels provided. The model performance did not improve with an additional input parameter although this possibility can be explored with other architectures. However, the model currently lacks accuracy in making one-step-ahead predictions, potentially due to boundary issues within the numerical model or the CNN itself. Further optimization and exploration of additional methods are necessary. The integration of physical parameters into the training process may improve prediction accuracy. Strong conclusions cannot be drawn until additional research is conducted.

Under future warmer climates, drought events are projected to occur more frequently with increasing impacts in many regions and river basins. This study focuses on exploring the potential of the LSTM deep learning (DL) approach for operational streamflow drought forecasting for the Rhine River at Lobith with a lead time (LT) of up to 46 days.
The research investigates optimal spatial resolution, input and target variables, and loss functions. Four LSTM-based model architectures are developed and tested, incorporating both historical observation and forecast data to generate 46-step forecasts simultaneously. The robustness and stability of the models are assessed through cross-validation, and their performances are compared. Subsequently, the performance of the LSTM-based model is compared to the physically-based models, namely Wflow-Rhine and FEWS-Rhine, in forecasting streamflow drought.
The results suggest that utilizing a subbasin spatial resolution, including historical discharge as input, and training the model on time-differenced data enhance the forecast skill. Among the evaluated models, the model architecture with two LSTMs in cascade exhibits stable and robust performance across the forecast horizon and is considered for operational use in this study. Comparisons between the DL model and physically-based models indicate that: 1) When using observed meteorology forcing from ERA5, the DL model demonstrates a notable performance compared to Wflow-Rhine simulation using the same forcing data. 2) When utilizing SEAS5 for forecasting, the DL model demonstrates skill over Wflow-Rhine in predicting discharge levels during the dry season up to 10 days ahead, as well as for discharges between 950 and 2200 m3/s across the entire forecast horizon. However, for discharges between 700 and 950 m3/s with longer LTs beyond 20 days, Wflow-Rhine shows skill over the DL model. 3) While FEWS-Rhine successfully forecasts drought events in 2018 throughout the forecast horizon, it tends to produce more Type I errors (false positives). The DL model, forecasting with SEAS5, accurately predicts drought events in 2018 for LTs up to 30 days and generally has higher precision values. Despite using different forcing datasets, the DL model can predict the timing and trend of past drought events, indicating its potential in capturing streamflow patterns.
This study contributes to operational water management in the Netherlands by employing the LSTM deep learning approach in an operational framework for drought forecasting. By leveraging historical observation data and forecasted meteorology forcing data, these models achieve skillful performances for streamflow drought forecasts. Future research could focus on further enhancing model performance, exploring the applicability of the LSTM-based models in other river basins, and validating the results in real operational settings.

Hydrological models are commonly used to predict future streamflow. However, the assumption of stationary model parameters obtained through calibration on past conditions may not accurately represent non-stationarity in hydrological system characteristics. Evidence suggests that vegetation adapts its root zone storage capacity in response to changing climate, emphasizing the need to account for non-stationarity in hydrological models. This study examines the effects of long-term climate variability on root zone storage capacity and its consequences on hydrological response. By using the method of Fu (1981), we determine whether the root zone storage capacity has significantly changed and evaluate the sensitivity of hydrological model predictions to these changes. We explain the changes in root zone storage capacity and evaporation using various climate indicators. For two large sample datasets CAMELS GB and CAMELS USA, we confirm that the Fu method can be used with the same omega parameter when transitioning from one decade to the next, indicating small differences in root zone storage capacity. However, for hydrometeorological data from the Meuse basin in Northwest Europe, we observe a trend where the actual evaporation is smaller than expected. This suggests that caution should be exercised when applying the Fu-method. In four different scenarios, we have implemented historical changes in evaporation as altered root zone storage capacity in the wflow flextopo model. The scenarios based on the evaporation trends observed in the Meuse data, result in less evaporation during the summer months (May, June, July) within a range of 0 to -22%, and an increase in streamflow during the autumn months (September, October, November) between 0 and +48%. On the other hand, the scenarios based on evaporation trends observed in all combined data (Meuse, CAMELS GB, CAMELS USA) result in changes around zero of evaporation during summer months between -7% and +5%, and of streamflow in autumn months between -11% and +10%. This study represents a step towards a more reliable and robust estimation of root zone storage capacity in hydrological modelling, thereby enhancing our ability to predict future streamflow.

The Saigon River, coursing through Ho Chi Minh City, is a vital yet alarmingly polluted waterway. It ranks among the top 50 rivers globally contributing to plastic pollution. This study delves into the complex mechanisms governing the transport of floating plastic within a tidal system, such as the Saigon River.
Our methodology adopts a multifaceted approach, combining visual observations, on-site measurements, and a comparison with existing literature data and methodologies. The factors influencing plastic transport in the studied river stretch are several and complex, ranging from the seasonal fluctuations in rainfall and tidal patterns to the role played by water hyacinths, acting as effective catchments for plastic litter, thus shaping the trajectory of these materials. Furthermore, we investigate the different types of plastic that flow on the river surface.
At the end of our research, we develop an early-stage conceptual model. This model serves as a framework that could help understanding plastic transport within the Saigon River and emphasizing the interplay of numerous influencing factors.
Our findings underscore the necessity for comprehensive investigations into plastic transport in the Saigon River. By addressing these knowledge gaps, we can develop more effective strategies to mitigate plastic pollution.

Water distribution networks (WDNs) provide drinking water to urban and rural consumers through a network of pipes that transport water from reservoirs to junctions. Water utilities rely on tools such as EPANET to simulate and analyse the performance of water distribution networks (WDNs). EPANET solves the flow continuity and headloss equations through pressurised piped networks under steady-state conditions. The EPANET solver applies the iterative Global Gradient Algorithm (GGA) of until convergence to determine unknown flows and pressures at the same time. Despite the widespread success of GGA, the speed of the algorithm may hinder several applications that require many simulations, especially for large networks.

To overcome these issues, researchers often resort to surrogate models, also known as metamodels, to significantly reduce the simulating times while approximating the behaviour of hydrodynamic models. Machine learning methods are increasingly being used as surrogate models for WDN analysis. Among these, the multi-layer perceptron (MLP) is the preferred metamodeling choice. MLPs provide a fast alternative to hydrodynamic models, but may lack the desired level of accuracy for complex case studies and applications. This is due to the lack of domain knowledge. Domain knowledge in WDNs can be divided into the topology of the network and the physical equations that govern the system. Recently, researchers have expanded MLPs to include domain knowledge in the form of the topology of the system. These new machine learning-based surrogates are called GNNs. However, GNNs present the problem that they lack speed compared to MLPs and still do not include information on the physical equations of the system.

For this reason, we propose applying algorithm unrolling to the GGA to include information on the physical equations that govern WDNs. Algorithm unrolling (AU) is a promising technique within the field of model-based deep learning that provides an alternative to traditional data-driven methods for approximating the output of iterative algorithms. This approach involves deconstructing the iterative algorithm used to solve the optimisation problem of interest into blocks that can be represented as layers in a neural network. Our architecture identifies two steps in the GGA and designs layers of a deep neural network following the steps of the GGA. In addition, the architecture includes information of the system features, those that remain constant throughout the steps of the GGA. We do so by adding residual connections to the variables. We evaluate the proposed metamodel by predicting the heads of five WDNs with varying characteristics and present an ablation study to understand the effect of the different components of our architecture. Our results show that our metamodel improved the accuracies with respect to MLPs and GNNs while being up to 2000 times faster than EPANET. We believe that the increase in accuracy with respect to the state-of-the-art baselines and the increase in speed with respect to EPANET justify the use of this metamodel for applications in the field of WDNs that rely on many simulations of EPANET.

In this research the possibilities of the application of machine learning models at ‘Hoogheemraadschap van Delfland’ are studied. A random forest (RF) and an LSTM model are used for the prediction of the sum of the discharge in the next 2, 8 and 12 hours from the polders to the boezem canals. This research has showed the potential of machine learning models for the prediction of discharge for the considered pumping stations in the case area. This case area is clustered in the Sobek RR model as node 49. The RF and LSTM model are compared to the current Sobek RR model, the machine learning model of Delfland (ReRengAI) and a naïve model by calculating the root mean squared error (RMSE) for the last year of the dataset. For the prediction of the 2 hourly sum of Node 49 the RF model performs the best. Additionally, the performance of the RF model for the 12 hourly sum is satisfactory with a RMSE of 11,071 m3, though using a deep learning model (LSTM) the performance improved to a value of 10,181 m3 for the RMSE. Machine Learning models are known as black-box models and are hard to explain and interpret, which makes the practical implementation of these new models, despite good model results, challenging. Technical recommendations for implementation ML models are improving the quality and availability of the data, increasing the interpretability and explainability of the model, combining multiple objectives in the new model or combining a ML model with a physical model. Organizational recommendations are improving the knowledge about these models within the organization, studying the advantages of these models in comparison to the current model and involving different departments of the water authority in the development of these new models.

Modeling the relationship between rainfall and runoff is a longstanding challenge in hydrology and is crucial for informed water management decisions. Recently, Deep Learning models, particularly Long short-term memory (LSTM), have shown promising results in simulating this relationship. The Transformer, a newly proposed deep learning architecture, has also demonstrated the ability to outperform LSTM in machine translation, text classification, etc. However, there has been limited research on applying Transformers for rainfall-runoff modeling.

The research examined the performance of using Transformer architecture, including its time series forecasting variants, to develop rainfall-runoff models using the CAMELS (US) data set. These models were compared to the LSTM regional rainfall-runoff models, with a particular focus on snow-driven basins as the attention mechanism in Transformer is believed to allow it to attend to the earlier precipitation events in the meteorological forcing. Additionally, the Transformer's potential as a global rainfall-runoff model was also tested using the global Caravan data to determine if it could learn and generalize a wide range of rainfall-runoff behaviors, allowing it to potentially be applied in ungauged basins.

The results suggest that while Transformer and its variants may not be able to fully replace LSTM for rainfall-runoff modeling, the variant called Reformer has shown promise for daily discharge forecasting in snow-driven basins, particularly in terms of peak flow and low flow prediction. However, using the global Caravan data for building a global rainfall-runoff model was not successful due to uncertainty in the forcing data, particularly precipitation. The code for Transformer-based rainfall-runoff modeling is available publicly at https://github.com/Numpy-Panda/neuralhydrology_Transformer.

Plastic pollution and accumulation in the riverine environment is of increasing concern. While most research focuses on microplastic contamination, the dynamics of macrolitter remain largely unknown. Large scale riverbank monitoring initiatives in the Netherlands reveal that macrolitter hotspots occur at several locations. Unfortunately, current knowledge on how these hotspots emerge and how this is influenced by hydrology and meteorology remains limited. As most studies are based on data from seasonal monitoring activities, short term variability remains unknown. This study is the first attempt to monitor and analyse the variability of riverbank macrolitter within a single location for over a period of three months. Behaviour of individual items is tracked and macrolitter exchange between water and riverbank is studied with regards to hydrology and wind. Finally, a conceptual model on riverbank macrolitter dynamics is presented in favour of supporting future research design.
A remote groyne field in the Waal has been monitored 21 times within the period of November 2021 until January 2022. The location of macrolitter items was recorded using Real Time Kinematic positioning. This allowed for analysing spatial patterns throughout time. Additionally, photographs of items were made in order to categorise the items without removing them from the riverbank. The river OSPAR protocol was used for item categorisation.The data shows that macrolitter primarily accumulates in the floodmarks. Rising water pushed items higher on the riverbank. Wind had a limited effect on item mobilisation as most items are wet and sandy. Analysis of item exchange between riverbank and water revealed that macrolitter deposition was observed at a relatively constant rate with minor deviations. Item uptake was heavily dependant on changes in water level. Uptake was initiated when the water level rises (dH > 0). The rate of uptake was higher with a larger water level increase. However, correlation was not statistically significant as riverbank morphology, substrate and vegetation may also influence uptake. After three months and two moderate discharge peaks, almost all items found on day 1 (estimated 99.6%) had been taken up. This indicates that under normal hydrologic conditions, the retention time of items within groyne fields is defined by the timing and magnitude of moderate water level fluctuations (assuming no accumulation under water within the groyne fields).
A conceptual model of riverbank macrolitter dynamics under natural processes is presented. Macrolitter can be stored in three domains: water, sediment and riverbank surface. Exchange within these domains occurs in four directions: in/out of sediment (storage/mobilisation) and in/out of water (uptake/depositions). Exchange is promoted by an interplay between item attributes, environmental processes and riverbank morphology. Future research on the interaction between these variables is needed in order to fully understand macrolitter dynamics.

Water utilities face many challenges, including pipe bursts that cause significant non-revenue water losses. Detecting those bursts early is important for the water sector in its path to achieve sustainable water resource management. This study presents a scalable data-driven methodology for burst detection in water distribution systems that is based on Long Short-Term Memory (LSTM)-based neural networks (NNs) and includes two stages: prediction and classification. Time-series of hydraulic (flow and pressure) signals are fed to the LSTM, whereas domain (time) features of the next time step are fed independently to regular neurons. These two streams of information are then concatenated to predict the values of the hydraulic features of the next time step. The model is trained on normal conditions only, so that when fed with data corresponding to a burst, the predictions will mismatch the observations. Comparison of the predictions to the observations is quantified though an error function, which is then used for classification. Specifically, a variable error threshold that corresponds to a pre-defined extreme percentile of the error distribution is used to discern bursts from normal conditions. The methodology is corroborated on two different types of bursts: (a) real bursts in district metered areas (DMAs) in the United Kingdom and (b) simulated fire hydrant leak tests in the same DMAs. For the real bursts, sensitivity analysis of the algorithm is performed to assess how data resolution and error threshold affect the performance. The flexibility of the method is studied for the simulated fire hydrant leaks, where additional information streams from new sensors are incorporated in the model by means of applying transfer learning and fine-tuning. The results obtained demonstrate that this scalable LSTM-based methodology works reasonably well in real-life settings and can successfully identify burst events, both real and simulated, even in DMAs with a small number of installed sensors. Furthermore, it is assessed how the flexibility of the LSTM neurons is pivotal for burst detection when utilizing a varying number of sensors.

Fluvial flooding poses a major threat to mankind and annually leads to major economic losses with many casualties worldwide. The consequences of this can be mitigated when accurate and rapid predictions are available when the water will arrive at which location. Current numerical simulations take a significant amount of time due to their computational demand, which is not affordable during a calamity where each second can make the difference between life and death. Literature has shown promising results in fluvial flood forecasting by using a deep learning model, which generally takes only a fraction of the computation time compared to conventional models. However, these models have thus far only been trained on static landscapes with changing boundary conditions, which prevents the model’s application elsewhere.

This research explores therefore the possibilities of creating and training a generic non-location bound deep learning model which can predict the spatial distribution of fluvial flood arrival times per grid cell. The architecture of the created deep learning model consists of five parallel encoder-decoders, which takes the elevation, slopes, elevated elements, land roughness, and initial water levels into consideration, depending on the dataset. The model is trained, validated and tested on four unique datasets, which consists of 30,000 flooding samples. The degree of complexity within a sample increases with each dataset number.

The average error for the four consecutive test datasets were 0.91, 1.41, 1.25, and 1.76 hours per cell. The differences in the predicted and the groundtruth are relatively small although the deviation tends to become larger at the end of the simulation, in regions with a strong gradient in arrival time, and in hilly and complex landscapes.

In addition, the model has been tested on various benchmark landscapes to examine specific flow phenomena, as well as a realistic test scenario in the Dutch dike ring 48. The model shows satisfactory performances for landscapes other than those present in the dataset, untill it encounters a feature on which the model was not trained for, such as undershots or irregularly shaped waterways.

This research has shown the potential of deep learning in predicting fluvial flood arrival times on unseen before landscapes. Recommendations for further studies include the use of an active dike breach and a variable flood location.

In the past decade, the application of Neural Networks (NNs) has received increasing interest due to the growth in computing power. In the field of computational mechanics, this has led to numerous publications presenting surrogate models to assist or replace conventional simulation methods. A subset of these networks, referred to as Graph Neural Networks (GNNs) impose the graph-like structure of many physical problems as a relational inductive bias. Several time-stepper implementations of these GNNs are reported to be able to simulate the dynamic behaviour of various physical objects. Within this work, it is investigated whether such GNN-based surrogate models can be applied to simulate the dynamic behaviour of lattice structures.
Upon inference of such a GNN surrogate model, the computational time required for studying lattice behaviour could be considerably reduced, thus advancing research into lattice structures as metamaterials. In addition, many large-scale structures also form a composition of beams, which could be modelled with a similar GNN.
To this end the following research question was defined: ”To what extent can GNNs be applied to simulate the dynamic behaviour of lattice structures using time-stepper methods?”. To answer this question, several GNN architectures were constructed and subsequently analyzed.

In this research, it was found that the complexity of lattice structures could not be modelled in such a way as to obtain reliable, generalisable and stable behaviour, using a time-stepper method with an architecture similar to that of Pfaff et al., 2020. It was found that due to the existence of three physically different coupled Degrees of Freedom (DOF) per node the behaviour was too complex to learn for the proposed surrogate.
At the time of writing, there is no publication presenting an effective surrogate model to simulate the dynamic behaviour of a Timoshenko beam using time-stepper methods. It is concluded that the need to capture both bending and shear behaviour using a Timoshenko beam formulation is the bottleneck for successfully modelling lattice structures.

This thesis investigates the impact of a phenology model on conceptual hydrologic model. In conventional conceptual hydrologic models the evapotranspiration is partitioned into evaporation and transpiration by a combination of the potential evaporation and the availability of water. This way the seasonal differences in vegetation dynamics are not taken into account. These vegetation dynamics represent the presence of transpiring leaves in summer, changing to an almost complete absence of transpiration in winter for some vegetation types. Including information on these vegetation dynamics could potentially improve the way streamflow and evaporation is modelled by conceptual hydrologic models. To investigate the impact of a phenology model on conceptual hydrologic models the FLEXmodel is adjusted with a phenology model based on the ’Kc-ETo’ approach of the Food and Agricultural Organisation (FAO) of the UN. With the conventional FLEX-model and the FAOadjusted FLEX-model, 28 catchments from the USA are simulated in terms of streamflow and transpiration. The 28 catchments differ in terms dominant vegetation type and climate indices. The conventional model and the adjusted model are calibrated with streamflow observations and transpiration data from NASA Global Land Data Assimilation System (GLDAS) both separately and together. The conventional FLEX-model and the FAO-adjusted FLEX-model are compared to each other in order to determine under what climate conditions and for which vegetation types the phenology adjustment improves the performance of the model in terms of the ability to simulate streamflow and transpiration and the predictive capacity of the model. The streamflow simulation of the FLEX-model does not improve by the FAO-adjustment. The FAO adjustment does cause an improvement in the mean seasonal sum of the discharge in the spring months, which is structurally underestimated by the conventional model due to an overestimation in spring of the transpiration caused by the lack of information on vegetation dynamics. However the transpiration simulation of the FLEX-model does improve drastically by the FAO-adjustment. The large overestimation of the transpiration in early spring by the conventional FLEX-model is solved by the FAO-adjustment, improving the NS- and NSlogefficiencies for all catchments, regardless of their vegetation type of climate conditions.

Leakage is the main source of water loss in water distribution networks (WDNs). Therefore, leak detection and localization technology is a major concern for water utilities to save water and meet the ever-growing water demand. This study presents two methodologies for leak localization in District Metered Areas (DMAs): (1) a model-based method using the hydraulic model, flow and pressure measurements, as well as leak flow; (2) a data-driven method that relies on graph-based interpolation. The performance of the model-based method is proved to be negatively affected by model errors and limited sensors. To solve these two problems, on the one hand, flows and residuals between observed and model-simulated data in the non-leak situation are used to develop a residual model to calculate offset values for model output correction. On the other hand, graph-based interpolation is introduced to create ‘virtual’ sensor measurements in the presence of a limited number of sensors. The data-driven method proposed in this work uses graph-based interpolation to estimate the head signals at the nodes without sensors and subsequently create pressure maps. Leak localization is achieved by comparing pressure maps in the non-leak and leak situations. In this process, this methodology does not require a well-calibrated model and leak flow information. Both two methodologies are tested on fire hydrant leak tests in DMAs in the United Kingdom. Results obtained by using the model-based method illustrate the positive impact of model output correction on localization results and the performance of this method under different conditions such as different times of the leak and different sizes of leak flow. The data-driven method performs fairly well in DMAs with a higher spatial density of sensors. Furthermore, the results of both two methods are compared to demonstrate the suitability of the methods in different cases.