Field-theory inspired physics-informed graph neural network for reliable traffic flow prediction under urban flooding
提出一种物理信息图神经网络模型，结合交通流物理约束与图结构，在降雨到洪涝的数据稀缺场景下仍能可靠预测城市交通流，为应急决策提供可解释结果。
The increasing frequency and severity of extreme weather events, particularly rainfall-induced urban flooding, pose significant challenges to urban transportation systems, affecting public safety, economic productivity, and overall quality of urban life. Traditional traffic prediction methods, while effective under normal conditions, struggle to maintain reliability during flood events. Recent advancements in artificial intelligence, despite showing promise, face critical limitations in handling data scarcity during flooding conditions, generalising to unprecedented scenarios, and providing interpretable predictions for emergency decision-making. This research proposes a novel Physics-Informed Graph Neural Network (PINN-GNN) model for predicting urban traffic flows specifically under rainfall conditions ranging from light precipitation to heavy rainfall events, which can lead to urban flooding impacts. Our approach innovatively integrates complex physical constraints with advanced data-driven learning techniques through three key components: a dynamic graph-based representation of urban transportation systems that captures essential spatial relationships, a hybrid architecture combining GNN with physics-informed neural networks that ensures predictions adhere to fundamental traffic flow principles, and a field effect module that successfully captures long-range dependencies and flood-induced failure propagation patterns. The most significant contribution of our work lies in the model's ability to maintain reliable predictions in data-scarce flood conditions while providing interpretable results. Through comprehensive testing across multiple cities with diverse characteristics, our model demonstrates consistent performance improvements over existing methods under various rainfall conditions and urban scales. The effectiveness of PINN-GNN stems from its unique combination of physical principles with data-driven learning, enabling robust predictions even in unprecedented scenarios where historical data is limited. This novel algorithm offers valuable support for urban planners and emergency responders by providing reliable and interpretable traffic predictions during urban flood events, contributing significantly to the development of more resilient cities and more effective emergency response strategies.