Tip of the Week #99                   Tip Index

Go to the Prior Tip Risk, Uncertainty and Investment Decision-Making in the Upstream Oil and Gas Industry
Go to the Next Tip Extraordinary Popular Delusions and the Madness of Crowds
Go to Tip 051  Neural Networks
Return to MaxValue Home Page

Neural Network Control System for Steam Injection

"Control system optimizes EOR steam generator output" by George S. Cheng and Li-Qun Huo, CyboSoft, Oil & Gas Journal, Sept. 8, 2003, pp. 46-50.

This interesting article describes how a "model-free" adaptive control system helps improve safety and oil recovery in China.

The business and engineering problem is controlling the quality of steam that is injected into an oil reservoir for enhanced oil recovery (EOR).  Optimal "dryness" is 72%.  If too wet, their client, PetroChina, would recover less oil. If too dry, then energy is wasted.

Key process variables include water flow, gas pressure, steam temperature, steam pressure, exhaust air temperature, and furnace temperature.  Process control is made difficult by changes in the reservoir pressure, changes in the natural gas fuel pressure, non-linear process in the steam pressure loop, time variability of the steam temperature loop, time delays of the steam loop process.  These factors challenge a conventional control operation with manual and proportional-integral-derivative (PID) controllers (a 60-year old technology).

CyboSoft, General Cybernation Group Inc. developed new approach based around a neural network (NN; see Tip 51 for a discussion of NNs).  NNs are used in analysis for many-factored, non-linear regression.  This article is the first that I have seen where a NN is used for real-time control of a plant process.

The control objective is to minimize the error between the setpoint and process variable (% dryness). The neural network has one input layer (the process variables), one hidden layer, and an output layer consisting of only one neuron.

A traditional plant control system has almost no memory.  The neural network, by comparison, learns from experience what settings work best.  The system updates the network weighting factors with a learning algorithm.  The neural network outputs adjusts the gas flow, water flow, and air flow.

This is an excellent artificial intelligence application.  Rather than model and solve the entire process, this neural network handles a localized control challenge.  CyboSoft has a simple, elegant approach to what would otherwise be a complex problem.

—John Schuyler, September 2003.

Copyright 2003 by John R. Schuyler. All rights reserved. Permission to copy with reproduction of this notice.