Reducing uncertainty associated with CO2 injection and brine production in heterogeneous formations

The problem

One proposed way to reduce greenhouse gas emissions is to inject CO2 into deep rock formations filled with salty water (brine). Because brine is less compressible than CO2, injection is likely to raise brine pressure over large distances (Birkholzer et al., 2009). Pumping brine out through nearby production wells can limit this build-up (Buscheck et al., 2012). However, the low pressure around a production well also draws CO2 towards it. Once CO2 arrives there (breakthrough), it is pumped back up with the brine. This reduces the efficiency of storage.

How soon this happens depends on permeability, a measure of how easily fluid flows through rock. Permeability can differ by many factors of ten (orders of magnitude) and varies from place to place (heterogeneity). If an injection well and a production well both intersect the same feature of high permeability, such as a fault, CO2 may reach the production well more rapidly than expected. Our goal was to reduce the uncertainty in such predictions, using data that could plausibly be collected at a site.

What we did

We generated a synthetic map of permeability across a reservoir (a permeability field), called it the "true" field and treated it as a hypothetical storage site.

We placed one injection well at the centre of four production wells 1 km away (a five-spot pattern). Using FEHM, a code that simulates underground heat and fluid flow, we modelled 30 years of operation. The injection well was held 5 MPa (megapascals, a unit of pressure) above the starting pressure of 30 MPa, and the production wells 5 MPa below it. Each well was therefore offset by a sixth of the starting pressure (our calculation). We judged our prediction methods against the result for the true field.

We assumed that an operator knows only what could be measured: the permeability at each of the five wells, and interference tests between them. In an interference test, one well is pumped at a fixed rate and the pressure drop is recorded at a second, shut-in well. This drop indicates how well the rock connects the two wells.

The first method is Bayesian: it starts from a broad expectation (the prior) and narrows it with measurements. Our prior was wide. It allowed permeability to vary by "some orders of magnitude". The five well values gave a sample standard deviation of 0.73, less than the true value of 0.91. These correspond to factors of about 5 and 8 in permeability (our calculation). We used the well value of 0.73 to update the prior into a narrower distribution (the posterior). We built 324 permeability fields from the prior and 336 from the posterior, each matching the measured permeability at all five wells, and simulated storage in each one. The second method simulated the four interference tests in every posterior field. Fields whose pressure responses were closer to the true tests counted for more when we combined the predictions.

What we found

In the true field, a large part of the area between two of the production wells is more permeable than average. Fast pathways brought CO2 to these two wells within about 11 years. In the uniform model, CO2 reached none of the wells within 30 years.

The well measurements narrowed the 90% prediction ranges for CO2 retention, brine production and CO2 production by 40, 45 and 72%. The narrowing came mainly from excluding fields with the fastest or slowest flow. The true outcome fell within each range, although for brine production it lay close to the lower edge. For CO2 arrival times the improvement differed from well to well. Wells 1 and 2 had almost equal measured permeability and similar predicted arrival times. In the true field, CO2 reached well 1 early but did not reach well 2 within 30 years.

Weighting by interference tests narrowed the brine and CO2 production ranges by a further 38 and 69%. For CO2 retention the further reduction was 8%. This is probably because retention cannot be split into contributions from individual wells. After weighting, the ranges of predicted arrival time were narrower and included the true time at three of the four production wells. At well 3, the median prediction became 10 years, against a true value of 11.

Combining both steps, the brine production range is about 34% of its prior size and the CO2 production range about 9% (our calculation), as the widget shows.

Limitations

At well 2, which sampled a small permeable zone inside a larger region of low-permeability rock, weighting raised the median predicted arrival from 8 to 16 years, against a true value of more than 30. The methods need further refinement for cases like this. For all three flow outcomes, the true values tended to sit in the lower half of the prediction ranges. We attribute this to the method used to build the fields, which biases them towards the mean of the five well values. That mean is slightly above the true mean. The model also ignores buoyancy and measurement error. It assumes that permeability is correlated over at least 3 km: values at points closer together than this are not independent. Many of these assumptions, such as the long-range correlation, may be regarded as controversial. Where several models are plausible, the procedure would need to be repeated for each.

As new wells are drilled and more interference tests are run, the analysis can be updated to narrow the prediction ranges further.

Glossary

Brine is the salty water that fills the pores of a deep storage rock.

Breakthrough is the arrival of injected CO2 at a production well.

A five-spot pattern has one injection well at the centre of a square of four production wells.

Heterogeneity is the variation of a rock property, here permeability, from place to place.

An interference test pumps one well at a fixed rate and records the pressure drop at another, shut-in well.

MPa (megapascal) is a unit of pressure.

Permeability measures how easily fluid flows through rock.

The prior is the range of values we expected before seeing the data, and the posterior is that range updated with the data.

A prediction range is the band that contains 90% of the model results.

A production well pumps fluid out of the reservoir.

Retention is the CO2 injected minus the CO2 that comes back out with the brine.

The standard deviation of the logarithm of permeability measures how widely permeability varies, in powers of ten.