Data for paper "Evaluation of Graph Neural Networks for Urban Drainage Metamodeling: Key Components and Transferability Analysis"

This dataset supports the article “Evaluation of Graph Neural Networks for Urban Drainage Metamodeling: Key Components and Transferability Analysis” (Garzón et al., 2025). It contains SWMM 5.1.015 simulation data and machine learning artifacts used to develop, train, and evaluate a Graph Neural Network based metamodel for two urban drainage systems: Loenen and Tuindorp. The data was produced as part of a PhD project at TU Delft to assess metamodel accuracy, speed up, data efficiency, and transferability under different conditions.<br>The folder "data" includes SWMM network files (.inp) and simulation outputs (lateral inflow, water levels, flows, and .out files) for training, validation, and testing. Testing included Dry Weather Flow conditions. Loenen includes complete training, validation, and testing sets. Tuindorp includes previously published training and validation data and newly added DWF testing data. All simulations were run using dynamic wave routing and rainfall events from the original case studies. No personal or sensitive data is included.<br>The folder "saved_objects" contains the machine learning components used in the study: pre trained model weights in PyTorch, feature normalizers used for consistent preprocessing, and pre processed time window datasets stored as PyTorch Geometric Data objects. These artifacts allow full reproduction of the metamodeling experiments and support transfer learning and further research.<br>The dataset enables researchers to replicate the methodology used in the paper, which combines SWMM hydrodynamic simulations with graph based deep learning models such as GINE and GAT trained on graph structured time series data. The workflow includes rainfall event simulation, feature normalization, graph construction, supervised learning, and evaluation on unseen inflow conditions including DWF baselines. <br>This work is supported by the TU Delft AI Labs programme.

本数据集配套学术论文《面向城市排水元建模的图神经网络评估：核心组件与可迁移性分析》（Garzón等，2025）。数据集包含SWMM（暴雨管理模型，Storm Water Management Model）5.1.015版本的模拟数据与机器学习工件，用于开发、训练并评估针对两个城市排水系统——洛嫩（Loenen）与图因多普（Tuindorp）的基于图神经网络的元模型。本数据集源自代尔夫特理工大学（TU Delft）的一项博士研究项目，旨在评估元模型的精度、加速比、数据效率以及不同工况下的可迁移性。
"data" 文件夹包含用于训练、验证与测试的SWMM网络文件（.inp格式）与模拟输出结果（侧向入流、水位、流量以及.out格式文件）。测试涵盖了非降雨工况（Dry Weather Flow，DWF）条件。洛嫩系统包含完整的训练、验证与测试集；图因多普系统则包含已发表的训练与验证数据，以及新增的非降雨工况测试数据。所有模拟均采用动态波演算方法，并使用原始案例研究中的降雨事件进行驱动。本数据集未包含任何个人或敏感数据。
"saved_objects" 文件夹包含本研究中使用的机器学习组件：PyTorch格式的预训练模型权重、用于统一预处理的特征归一化器，以及以PyTorch几何（PyTorch Geometric）数据对象形式存储的预处理时序窗口数据集。这些工件可完整复现元建模实验，并支持迁移学习与后续研究工作。
本数据集可支持研究人员复现论文中提出的研究方法：将SWMM水动力模拟与基于图的深度学习模型（如GINE与GAT）相结合，这些模型基于图结构时序数据进行训练。完整工作流涵盖降雨事件模拟、特征归一化、图构建、监督学习以及针对未见过的入流条件（包括非降雨工况基准）的评估。
本研究得到代尔夫特理工大学人工智能实验室（TU Delft AI Labs）项目的资助。