A Graph Neural Network Surrogate Model for Critical Infrastructure Network Flow Optimisation

Flow optimisation models have been developed for complex and large critical infrastructure networks. However, their computational cost rises with the complexity of the topological and functional characteristics of the networks. This presents a significant challenge particularly when the probabilistic assessment of CI networks requires the simulation of a large number of scenarios of network state evolution and corresponding network flow optimization runs. To address this challenge, we propose a graph neural net-work surrogate modelling framework to accurately simulate the physical flow optimization model, achieving substantial computational savings. The surrogate model is trained on a high-fidelity dataset generated by the original optimization model, which solves the network optimal flow across network scenarios generated by Monte Carlo simulation. The proposed modelling framework is validated on a hypo-thetical natural gas transmission pipeline network, whose supply availability is estimated by Monte Carlo simulation. The surro-gate graph neural network model allows significant savings in computational time compared to the conventional optimization-based approach. Model robustness and generalisability are further validated under different operational scenarios.