We can now make improvements on the assessment of the external coating condition of underground pipelines: in fact, after excavation and examination, only a small proportion of detected coating anomalies contain also a metallic fault or worse corrosion.

GRTgaz, operator and owner of the longest high-pressure natural gas grid in Europe (approximatively 32 000km (19,884 miles), has tested an analytical approach based on historical data to increase the success rate (from the understanding of the problem and the available data, to the elaboration of algorithms and the optimization of their parameter). Different data sources to extract new knowledge have been used, works have been prioritized and the excavations have been optimized in order to have fewer but more efficient interventions.

The development and validation of the predictive model is ensured by a multidisciplinary team consisting of data scientists, corrosion experts and project managers. Based on historical data from 2013 to 2017, the model was calibrated taking into account both the indications of greater probability of metal damage and those of lower probability. The output of this model gives an indication of likelihood of corrosion along the pipes which cannot be pigged.

The paper will cover the construction method (based on Random forest algorithms), the first results. We will see the enhancements of the model by injecting more data and modifying mathematical rule, and how it will be able to integrate a new decision support tool.


GRTgaz is a major gas grid operator (TSO). It operates a grid of approximately 32 000km (19884 miles) which is one of the longest pipeline system in Europe.

The company has a legal obligation to inspect and maintain all of its pipelines with a 10 years periodicity. There are about 2 700km (1678 miles) (inspected per year which leads to 1300 excavations representing an investment of 40 million € (approximately $ 45 million).

This content is only available via PDF.
You can access this article if you purchase or spend a download.