Last updated 9th June 1997by Austin Tate
SPIRIT: A knowledge-based oil well test interpretation system
SPIRIT is a knowledge based diagnostic system for a particular type of oil well test, which is used to determine how much oil exists in a reservoir, where it is and how it will flow to the well. The diagnosis draws on multiple sources of data and can handle uncertainty in the data. The prototype system was used by sponsors and a commercial product based on the system ideas has been released.
The motive behind the SPIRIT project was the perceived need by the project sponsors to improve the quality of well test interpretation as it is carried out within oil companies.
Existing conventional well test interpretation software assists reservoir engineers by removing some of the more mundane tasks as well as automating some of the manual interpretation techniques. However, the software generally provides no guidance on the most important stage in the analysis: that of selecting the most appropriate mathematical model to use to analyse the pressure test data. This task requires significant expertise and experience, beyond that of many reservoir engineers who use the software. PSTI wished to prototype a decision support system for reservoir engineers that encoded expert knowledge on well test interpretation.
PSTI (the major project funder) initiated the project as a means of exploring the requirements for the next generation of well test interpretation software. The aim was that petroleum software companies would then exploit the results of the project to improve their well test software, which would in turn benefit the petroleum industry as a whole. AIAI designed and developed the knowledge based component of the prototype software system.
SPIRIT is a knowledge-based diagnostic system for well test interpretation. SPIRIT makes use of several different types of information for interpreting a well test: pressure, seismic, petrophysical, geological and engineering.
SPIRIT includes a knowledge-based pattern matching module for the pressure data analysis. The pressure data is matched against the theoretical responses of all the known well test interpretation mathematical models. Each model has several "free'' parameters which govern the shapes of its theoretical responses. Each model therefore has several theoretical responses associated with it: each theoretical response corresponds to a "volume in parameter space'', i.e. a set of parameter ranges within which a change in parameter value makes no change to the shape of the curve. Pattern matching thus requires the ability to judge whether the actual test response is close enough in shape to any of the many possible theoretical responses. The difficulty of the model selection task is compounded by the non-uniqueness of type curves: two quite different models can have the same shape of theoretical response, depending on the parameter values. The output from the pattern match module is a ranked set of possible interpretations of the well test pressure data.
As stated previously an important aspect of SPIRIT was the requirement for managing uncertainty in the data. From speaking to experts it became clear that much of the uncertainty in data was expressed in non-numerical terms, and experts used symbolic terms to describe how likely a conclusion was based on a given set of data. based on this, a qualitative belief network approach was taken for encoding the experts' knowledge. The knowledge acquisition and encoding of the belief network was achieved using AIAI's own diagramming tool HARDY. Belief networks were shown to be an intuitive way of representing and reasoning with uncertainty. HARDY enabled the model of expertise to be captured in a graphical manner such that it was understandable to the expert, in a way which rules alone could not be. In addition the network was reviewed and modified interactively with the expert and the output was used directly in SPIRIT.
The developed belief network fuses information from the type curve matching and the other engineering and geological data sources. The result is that the level of belief in all the models known to SPIRIT are constantly kept up to date, given the data that have been input, and are then ranked and displayed. The user can select any of these models and can view why the model has the level of belief it does.
Last updated 9th June 1997
by Austin Tate