Accurate traffic flow forecasting is essential forintelligent transportation systems and smart cityapplications. While spatio-temporal graph neuralnetworks (STGNNs) have achieved strong empiricalresults, they often lack physical consistencyand interpretability. Physics-informed neuralnetworks (PINNs), exemplified by models likePINT, have shown promise in embedding domainknowledge into temporal models, but typicallyignore spatial structure. In this work, we proposePI-STGNN, a novel framework that integratesphysics-informed constraints into both spatial andtemporal dimensions of STGNNs for traffic flowprediction. Building upon insights from STDENand PINT, our model enforces spatial flow conservationlaws and temporal periodicity throughdedicated physics-based loss terms. We evaluatePI-STGNN on the LibCity Traffic Dataset, demonstratingits potential to improve generalization,robustness, and physical interpretability in trafficforecasting tasks. Our approach aims to bridgethe gap between data-driven and physics-awarelearning in real-world transportation systems.