The Algorithm That Might End Pipeline Leaks

In 2012, a natural gas pipeline outside of Charleston, West Virginia exploded, leveling a number of houses and business, turning a patch of rural Appalachia into literal hell. At the time, a state police sergeant described a neighboring interstate highway as a molten sea of blacktop, with the affected buildings being basically unrecognizable.

The cause: pipeline corrosion, rust.

If we're to have pipelines crisscrossing the landscape— some 1.7 million miles in the US—they should at least be safe. Fossil fuels combust, they burn and explode; that's why we're piping them around the country in the first place. As the West Virginia accident demonstrated, poor monitoring leads to very bad things.

A new study in the journalIndustrial & Engineering Chemistry Research offers the possibility of cheap, continuous monitoring using hydraulic models. Just with software, even small leaks might be detected, albeit indirectly.

Software detection exists currently, but the authors, Gary Valtinson and Miguel Bagajewicz, note that so far such systems have failed to accurately model pipeline pressure drops, with error rates up to 21 percent.

The new method implements what's known as the generalized likelihood-ratio (GLR). This is basically a way of taking the likelihood of one hydraulic model (the normal case) and putting it in terms of a ratio against another model (the leaky one). What are the chances that these data points might occur in "normal" circumstances vs. some other ones?

The researchers ran an economic analysis of their model, comparing costs for false positives and missed detections for a variety of modeling techniques. Using GLR techniques would have saved millions, according to the study. Their model managed to push error rates down to 3 percent.

Of course, the gas company has to do something about that leak once detected. When the WV pipeline exploded, melting a freeway and obliterating peoples' homes, the pipeline's control center somehow was still left unaware. No warning sounded, proving that leak detection is only as good as the environment in which it's implemented.