Veranstaltungsprogramm

This study compares traditional research workflows with AI-enhanced methodologies in water engineering, focusing on how Large Language Models (LLMs) can expedite literature reviews, idea generation, and manuscript drafting. Building on historical examples like calculator adoption, we illustrate both the efficiency gains and potential drawbacks—including biased outputs, “hallucinated” references, ethical concerns, and environmental impacts. Despite these risks, LLMs offer significant promise in automating repetitive tasks and providing creative insights. By maintaining active human oversight and transparent practices, researchers can harness AI’s capabilities to enhance, rather than undermine, the integrity and depth of academic work.

Machine learning is widely used in urban drainage networks, and especially opaque models like the random forest, neural networks, or their variations are frequently used. Opaque models employ thousands of model parameters to model complex relationships in data, which limits the comprehensibility of their results for humans. In this regard, the European “Artificial Intelligence Act” was published in July 2024, specifying a transparent implementation of ML, especially in high-risk areas like critical infrastructure. To address this issue, two trained opaque XGBoost models of previous work are extended with different techniques of post-hoc explainability to understand the model’s decisions. The analysis reveals that the most important input features for both XGBoost models represent the engineered classification into dry or wet flow. The next most important input features for the purely data-driven ML model are related to historical rainfall data. The complex relationship between water flow and historical rainfall data is captured, and high quality data and processing is required to prevent the model from learning incorrect correlations. In contrast, the physics-leveraged ML model includes the output of a hydrodynamic model as an important input feature in the model’s decision, decreasing the influence of historical rainfall data on the model’s decision.

Accurate water level prediction in existing urban drainage systems (UDSs) is critical for reliable forecasting of pluvial flooding impacts and reduction of flood-related damages in cities. Conventional physically based urban drainage modelling approaches are constrained by the need for extensive hydro-meteorological, drainage network and surface terrain data and high computational demands. In this research, more computationally efficient Machine Learning based Feedforward Neural Network (FFNN), multi-head Convolutional Neural Network (CNN) and 1D-CNN models were developed and applied to simulate water levels at a bridge crossing downstream of an existing UDS in Kampala City. The study results suggested that the multi-head CNN Deep Learning model resulted in more superior predictive performance (NSE, RMSE and MAE of 0.564, 0.208, and 0.091) when compared to the physically based PCSWMM model (NSE, RMSE, and MAE of 0.505, 0.221 and 0.098). Furthermore, the SHapley Additive exPlanations (SHAP) approach was applied to explain the underlying processes in the developed ML models and to determine the most influential model parameters. The research demonstrates that explainable Deep Learning models can reliably simulate water levels in UDSs, and provide a robust basis for development of real-time pluvial flood early warning systems in data-scarce cities.

Commonly, the usable budet for operation of the urban drainage network is calculated at the end of the year based on the billed drinking water consumption at costumer sites. To estimate the available budget, the network operator uses a simple forecast of the annual water consumption using the average of the last four years. To further improve this process, different machine-learning based forecasting model were developed with the aim to quarterly predict the annual water consumption by integrating also actual weather and system states measurements with higher temporal resolution. As the results show, Support Vector Machine achieved the highest accuracy over the quarterly forecasting time points, followed by Linear Regression. Therefore, Linear Regression with Bayes Statistics was selected for the machine-learning based forecasting model, providing the network operator also an uncertainty assessment of the forecasting value.

Sewer flow forecasting is critical for managing the performance of sewer networks and their treatment plants. While simulators have been used in modelling the sewer flow for years, emulators recently have gained attention in making predictions with a higher computational speed and feasibility. In this research, a framework is proposed based on multi-input single-output Gaussian Processes for predicting sewer flow using time and rainfall as inputs. The predictions are presented as Gaussian distributions, showing the confidence levels. The results of the GPR on the data of a sewer system in this study demonstrated a robust performance of the model with 93.6% coverage of the predictions in the 95% credible interval and 89.5 L/s of RMSE.

High-fidelity hydrodynamic models, such as the Storm Water Management Model (SWMM), provide accurate simulations of urban drainage systems but are computationally expensive. Graph Neural Networks (GNNs) have emerged as promising metamodels to approximate SWMM behaviour efficiently. However, the impact of GNN depth on predictive performance remains underexplored. This study investigates how increasing graph layer depth influences the accuracy of a GNN metamodel for a sloping urban drainage system.

Using an auto-regressive GNN framework, we evaluated multiple model depths, ranging from shallow to deep architectures, on a calibrated SWMM case study in Loenen, Netherlands. Results show that shallow GNNs struggle to capture transport-dominated hydraulic dynamics, leading to poor performance. In contrast, deeper models more accurately approximate SWMM’s hydrodynamic responses, achieving RMSE reductions from 10 cm to 5 cm. Contrary to conventional GNN literature, which suggests diminishing returns with depth due to oversmoothing, our findings indicate improvement with metamodels of up to 12 layers. However, deeper architectures impose higher computational costs.

These findings emphasize the importance of depth as a key hyperparameter in GNN-based metamodeling for urban drainage applications. Future work should focus on optimizing computational efficiency through network skeletonization, improving the explainability of hyperparameters and attention weights, and hardware acceleration.