Abstract

Natural gas release from oil and gas facilities contributes significantly to the greenhouse effect and reduces the benefit of displacing heavy fossil fuels with natural gas. Real-time concentration spatiotemporal evolution forecasting of natural gas release is essential to predetermine atmospheric carbon trajectory and devise timely strategy to mitigate the expected impact on the environment. Deep learning approaches have recently been applied for spatiotemporal forecast tasks, but they still exhibit poor performance pertaining to uncertainty and boundary estimations. This study proposes an advanced Hybrid-Physics Guided-Variational Bayesian Spatial-Temporal neural network. Experimental study based on a benchmark experimental and simulation dataset was conducted. The results demonstrated that the additional uncertainty information estimated contributes to reducing the harmful ‘over confidence’ of the point-estimation models at the plume area. Also, the proposed normalized uncertainty and physical inconsistency constraint term ensured the accuracy at the plume boundary. By adopting the Monte Carlo sampling number m = 100, penalty factor λ = 0.1, and drop probability p = 0.1, the model achieves a real-time capacity of an inference time less than 1s and a competitive accuracy of R2 = 0.988. Overall, our proposed model could provide reliable support to maximize the environmental benefits of natural gas energy usage and contribute to the carbon peak and neutrality target.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call