Performance Analysis of Deep Learning Implementation in Operational Condition Forecasting of a Gas Transmission Pipeline Network

Abstract

Monitoring natural gas transmission in a pipeline network is important to maintain the supply and demand balance in natural gas transactions and distribution. Gas pressure, temperature, flowrate, and gas properties must be monitored during the transmission process. These variables, also known as operational conditions, need to be simulated carefully to understand the dynamics and behavior over time. Commonly used physical equations, such as thermodynamic or hydraulic equations, have limitations in simulating future trends because they need some known boundary conditions to be solved. In that case, data-driven method is needed, especially nowadays when data management is widely implemented. This paper implements a deep Recurrent Neural Network (RNN) to forecast the future behavior of gas pressure as an operational condition in a gas pipeline network receiving platform. Different types of recurrent cells are used, i.e., Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The model is trained in 8 years of a gas pipeline network operational data. Historical flowrate data in the end-nodes become the forecast input in addition to the past pressure data. The sensitivity of the model and learning parameters is experimented with and analyzed to understand the capacity of the RNN in the given task. Mean absolute error is set as the satisficing metric, whereas the training time is set as optimizing metrics. The obtained best model successfully forecasts the future pressure of one day ahead with only around 2% relative error.