Operational Decoupling of Demand and Pressure in Gas Transmission Networks: A Two-Stage Machine Learning Approach

Accurate forecasting of gas demand and pressure is essential for the stable operation of natural gas transmission networks. This paper presents a two-stage machine learning framework applied to two years of operational data from six gas stations. In Stage I, daily gas demand is predicted using Lasso, Random Forest, and XGBoost; Lasso achieves the highest accuracy (R_x۰۰۱۰_۲ _x۰۰۰۲_۶۰ ۰.۹۳) across most stations. In Stage II, outlet pressure is forecast using predicted demand and historical pressure features. Surprisingly, feature importance analysis reveals that pressure prediction relies overwhelmingly (>۹۰%) on autoregressive pressure history, with negligible dependence (<۱%) on demand forecasts. This operational decoupling challenges conventional assumptions of demand-driven pressure control. Results also highlight significant station-level heterogeneity, underscoring the need for tailored modeling. The proposed framework offers a practical, data-driven approach for enhancing pressure stability in gas networks.

کلیدواژه ها:

Gas transmission ، pressure forecasting ، demand forecasting ، machine learning ، operational decoupling ، two-stage modeling