Model bidirectional GRU with bayesian optimization for dry gas pressure forecasting in transmission pipeline networks

Abstract

Operational robustness and infrastructure safety require that stable dry gas pressure in transmission pipeline networks be maintained. Due to the highly nonlinear and time-varying behaviors of the dry gas pressure, conventional methods cannot predict its tendencies accurately. Deep learning models have recently gained good performance for time-series forecasting, however combination of Bidirectional Gated Recurrent Unit (BiGRU) and Bayesian Optimization based approach for dry gas pressure forecasting remains underexplored. Based on these, this study develops Bayesian optimization-enhanced forecasting framework for BiGRU to enhance the prediction performance. The framework is based on a Gaussian Process surrogate model and an acquisition function — in this case Expected Improvement, which guides the search for optimal hyperparameter configurations. We used operational SCADA time-series data consisting of hourly measurements (pressure, temperature, flowrate and gas composition) collected from a dry gas transmission pipeline across two annual periods. Pre-processing of data consists of dealing with missing data, Min–Max normalization and sliding window conversion. Here a multistep forecasting scenario for the next 20 hours was used, and thismultistep prediction was evaluated calculating the MAE, RMSE, and R². The optimized BiGRU performed the MAE of 11.7729 psi, RMSE of 14.7827 psi and R² of 0.9389, which enhanced the baseline model by 13.67%, 12.90%, and 1.60 percentage points respectively. These results indicate that Bayesian Optimization improves the forecasting performance of BiGRU and, at the same time, decreases the manual hyperparameter tuning efforts.