### Reservoir Simulation Goals
### Automatic Real-Time Probabilistic Global Predictive
Optimization and Control
"Real-Time" and "and Control" are longer-term
objectives. We now have
deterministic and probabilistic workflows with
automated iterative optimization for all of these components, except for characterization in a
geological model*. The 4 iterative optimization problems are almost
identical and differ only in the definition of the benefit function to be
minimized or maximized, and identity of the adjusted variables. Workflows
for characterization, upscaling, and history matching are essentially the
same. They minimize uncertainty by minimizing a mismatch function.
Predictive optimization maximizes a value function such as BOE or NPV.
We and
others** are working on integrating and automating the entire system in a
continuous workflow (shown) that automatically evolves with time and new data to
continuously generate better upscaled history matches, optimizations, and
predictions and achieve our ultimate goals. Automating the workflows
and and the system requires batch processing of all components. The
probabilistic workflows require probabilistic representation and treatment
of uncertain input data, from which probabilistic results can be robustly computed. Many have referred to this
automated system as the "Digital Oilfield", but most have so far limited
their definitions and applications to surveillance and remedial response
through direct or remote control, and as far as we know, no others have
included a robust framework or method for treatment of uncertainty.
Probabilistic
treatment is usually needed to answer questions in reservoir
modeling, because our heterogeneous reservoirs are mostly unknown systems.
Reservoir properties are observed or interpreted only at sparse locations
where wells are drilled, and in between wells those properties are populated
in a geological model using geostatistical methods, which means that there
are an infinite number of possible realizations that match the observed
initial data, and that may equally well or sufficiently match history.
When there are large numbers of uncertainties, as there usually are, no
individual realization (description) is ever correct, and only probabilistic
results are meaningful. The number of uncertain variables is generally
equal to several times the number of gridblocks and wells, and may include
boundary (operating) conditions that are a function of time. See
Uncertainty Quantification -
P10, P50, P90, SensorPx, and
SensorMatch. We provide some basic
probabilistic tools that use simple statistics and probability theory to
form the building blocks of these probabilistic workflows.
Probabilistic forecasting is currently limited to small
systems due to the large numbers of runs usually required to obtain a
statistically significant set and reliable probabilistic results (or it
requires large numbers of computers to make large numbers of
simultaneous serial runs). Optimizations however sometimes
only require a few runs, and can often be performed using prototype models
with no need for upscaling or history matching.
If the reservoir simulator could be replaced with some
much simpler and faster model, such as a proxy
model or a streamline model (or run simultaneously on very many computers), those probabilistic model size limitations
could be practically eliminated. Unfortunately, proxy methods become
very difficult and expensive with very large numbers of variables, and others make
assumptions that do not generally apply (eliminating them was the purpose
for which numerical reservoir simulators were invented). Solutions for
single cell or 1D, or single component or single phase flow, or Buckley Leverett 2 phase
immiscible flow, do not generally apply to multi-component multiphase flow
in real reservoirs,
and cannot be used to represent reservoir performance or to infer any
characterization from performance (see What is
Drainage Radius?). Our probabilistic tools and workflows can
handle an unlimited number of uncertain variables.
* Automated characterization requires a batch
probabilistic geological model capable of producing unlimited numbers of
equally-probable realizations of the uncertain, geostatistically-populated
properties. The characterization problem is equivalent to a history
matching problem at time 0 on the geological model scale.
** Iterative optimization and workflow integration
software is provided by third-parties. |