Page 117 - Intelligent Digital Oil And Gas Fields

Data Filtering and Conditioning 85

• if the last 3 daily averages are beyond the long-term average a single
standard deviation; and
• or if the last 5–7 daily averages on one side of the long-term average or
the other.
These rules are based on the statistical fact that less than 5% of normal process
variation should be beyond 3-day standard deviations. As the well behavior
does move over time, these calculations need to be monitored and may need
to be changed over time as the well changes. Furthermore, any time a large
variation occurs—for example, 10–15 times standard deviation—then the
data is likely bad.

3.2.3 Model-Based Validation Methods
The final type of the technology used to detect bad data is model-based
methods, which include two types: first principles or artificial intelligence
(AI). Any component that can be modeled with a first-principles technology
can be set up to have a value predicted from the model that is also measured
for comparison. Then you can look for deviations between the model-based
value and the actual reading from the instrument. Of course, deviations
could be due to either a model issue or a measurement issue. Further analysis
is necessary to determine the exact cause. AI can also be used to detect bad
data or process changes that may be due to bad data. Self-organized neural
networks or k-means clustering can be used as fault detectors. Furthermore,
they can be used with very large amounts of data. The AI methodology is to
download data and cleanse all bad data for the data set. Then train the cluster
on the good data. When the cluster algorithm is implemented in real time, it
issues an alarm if any data pattern is observed to be outside of the trained
cluster.

3.2.4 Data Replacement Techniques
If data validation routines detect bad or missing data, values may need to
be “created” to fill the gaps. For DOF, it is imperative to work with
continuous (and often high-frequency) data; most of the automated
engineering workflows (Chapter 5) work with continuous data, for
example, pressure and rates. In cases for which pressure or rates are miss-
ing or marked as bad data, the workflow should replace or populate this
unreadable data with alternative available source data. In our experience,
we believe there are several levels of complexity to replace or fill in data,
for example: