Page 184 - Intelligent Digital Oil And Gas Fields

Components of Artificial Intelligence and Data Analytics 141

integration of reactive and proactive dynamic data from nearly 5000 well
intervention events in a 3-year period, focusing on sanded/seized failure
events.
Last but not least, we touch on the application of predictive modeling to
fault/failure analytics of ESPs, which are currently the fastest growing
artificial-lift pumping technology, deployed across about 15%–20% of oil
fields worldwide. However, ESP performance is often observed to decline
gradually and reach a point of service interruption because of factors such as
high gas volumes, high temperature, and corrosion. Numerous workflows
have been implemented to monitor ESP performance and suggest action in
case of a failure. However, most such workflows are reactive in nature,
where action is taken after the failure event. Recently, the E&P industry
has seen an emerging trend of deploying down-hole sensors for real-time
surveillance of parameters impacting ESP performance, with an opportunity
to predict and prevent ESP failures using data analytics. Such data-driven
models would, for example, use the following ESP performance data: his-
torical data with time series values for critical parameters, maintenance logs
and calibration data, and operational specifications of the ESPs. Applying
analytics to this type of available data provides the ability to rank the ESPs
for priority attention based on fault analysis and then recommend appropri-
ate maintenance (repair, rehab, or replace).
For example, Gupta et al. (2015) have proposed a three-stage workflow
based on statistical MVA to detect and diagnose impending problems with
ESP operation. In the first stage, the key operational variables (decision vari-
ables) affecting ESP performance are identified and evaluated. They devel-
oped a hybrid monitoring-intervention model based on a robust PCA,
which triggers an alarm if the operational attribute under surveillance
exceeds the normal operating range predicted by the model. The second
stage involves principles of diagnostic analytics (see Fig. 4.2) aiming at the
potential cause that led to the failure. To better understand the root cause
and take appropriate action, an importance/sensitivity model (see tornado
chart in Fig. 4.11) was built to assess the contribution of the various decision
variables toward failures and rank them according to these contributions.
The third stage of the proposed workflow involves elements of prescriptive
analytics (see Fig. 4.2) and suggesting preventive actions. Such data-driven
workflows enable building an ESP health monitoring plot (Fig. 4.19), which
visualizes the principal components obtained from the model output, cap-
turing observed variances within specific confidence interval limits. Trends
or patterns during normal operation are identified and correlated to either