One and a half kilometres below ground level close to the small, town of Stenlille on Zealand, natural gas is stored in a water-filled layer of sandstone. Every year, large quantities of data have been collected from the total of 20 wells. How-ever, the behaviour of the gas is not easy to predict quantitatively because each operating year is different and as the geological conditions under which the gas is stored are, to a great extent, unknown.

Algorithm can find structures

The idea therefore arose to examine the possibility of constructing an algorithm or forecasting model which simulates withdrawals from and injections into the storage facility and which takes into account dependencies and interactions between the wells in the sandstone layer where the gas is located.

“Our objectives are improved operational reliability, better performance in general and the possibility of lower costs and higher sales potential,” says Martin Haarhoff, engineer in Gas Storage Denmark. 

Through a digital algorithm, historical data may, in fact, potentially be obtained from the total filling and operation of the storage facility. The algorithm can be used to find structures in the data, which can, in turn, be used to predict more accurately how the gas will behave in given scenarios. The model ignores unknown physical conditions in the geological underground and thus examines scenarios for withdrawals from the gas storage facility and the risks that a given with-drawal strategy entails. 

Balancing act in model construction

The pretotype process has lead to a new pretotype process as it was initially not possible either to confirm or completely reject the potential of a digital model of withdrawals from the gas storage facility.

“We’re now going through hypothetical experiments in which we test the algorithms in whether they can predict the things we already know with certainty. It’s a balancing act in which we prepare data for the models, while ensuring that we don’t give them data related to the answers we want the model to predict,” says Martin Hartvig. 

Through the pretotype processes, it has already now been ascertained that it is possible to construct models to predict reservoir pressure by looking at operating patterns. It is also the assessment that machine learning can optimise the operation of the gas storage facility, which is completely essential to Gas Storage Denmark’s business.

The gas storage facility in Stenlille.