Efficient estimation of natural gas leakage source terms using physical information and improved particle filtering

Abstract

Natural gas pipeline leaks can cause fires or explosions, making quick and accurate leak source identification critical for emergency response. This study develops a natural gas pipeline leakage source inversion model, where a Proper Orthogonal Decomposition-Physics-Informed Neural Network (POD-PINN) is integrated as the gas forward diffusion model. The inversion model combines an improved particle filtering algorithm, gas sensor data, and the POD-PINN, enabling rapid identification of leakage source terms. The gas source estimation results using POD-PINN and the Gaussian model as forward models were compared across different scenarios, and the impact of sensor errors on the inversion model was analyzed. Using POD-PINN as the forward model preserves accuracy while improving computational efficiency. The inclusion of a Gaussian kernel function and Markov Chain Monte Carlo (MCMC) method addresses degeneracy and impoverishment issues in standard particle filtering, preventing convergence to local optima. Results show that, across different scenarios, spatial position estimation errors are under 5%, and source strength errors are below 8%. When sensor measurement error is exceeds 0.5, the model cannot accurately estimate all source parameters. The proposed inversion model is subjected to convergence analysis, confirming its feasibility.