How to Get Optimal Production Utilizing Machine Learning...PI System data as well as external data...
Transcript of How to Get Optimal Production Utilizing Machine Learning...PI System data as well as external data...
How to Get Optimal Production
Utilizing Machine Learning
Optimizing Productivity Case Study 4
Artificial Lift Anomaly Detection
1
#PIWorld ©2019 OSIsoft, LLC
Outline
2
Examine the Ability of Machine Learning(ML) to predict
Anomalies on Artificial Lift Equipment to Pre-EmptFailure
and Eliminate unplanned Downtime
A Cross Section Analysis Including Plunger, ESP, and PCP pumps
• A look at how equipment failures can be predicted using MLintelligence
• Apply ML to detect scale, hydrates, and corrosion build up on your artificial lift
• Understanding the ability of ML to predictanomalies
Conclusion
Analytics Types
Applied Machine Learning
We use these processes and technologies because they work at solving real world industrial problems.
Planning & Forecasting
RankingKnowledgeBuilding
#PIWorld
But…• Machine learning needs fuel, that fuel
is data, often a large amounts of data.
• Machine learning workloads
involve
• training models using largescale
historical data sets.
• using these model to make newpredictions.
• ML requires secure, robust & scalable data, data in context.
#PIWorld
Predictive and Monitoring
Data in Context; PI AF Template Used for Well Plunger Analytics
#PIWorld ©2019 OSIsoft, LLC
Company faced highly expensive repair costs from electrical submersible pump (ESP)
failures so they wanted to reduce costs through the detection of failure before these pumps breakdow
Tom Hosea, Consultant
n.
• Using predictive insights, moved
from a reactive to predictive
maintenance mode
• A $5K opportunity cost/day/well
that scaled - 10,000 wells (with
a 10% failure rate & average
of 9 days of downtime), the
result is $45 million/yr
ESP Reliability Improvement Program Saves$45M/yr – PI & Azure Machine Learning
Company lacked visibility in predicting failures in electrical submersible pumps leading to high repair costs.
• $100M+/yr in repair of Electrical Submersible Pump (ESP)
• 3,800 ESPs with 20% failure rate per year
• Develop a PI System based application to predict ESP failure
• Setup PI Notifications to generate exception-based reporting
• Created and operationalized a proprietary predictive model that captured 60% of actual failures 60 days in advance
#PIWorld ©2019 OSIsoft, LLC
Production Improvement Opportunity
No. of ESP’s
• ~ 3800 installed ESPs
Gas Production
• ~ 6.6 MM BFPD
• ~ 250 M BOPD Gross
Failures
• Failure rate ~ 0.209
• ~ 800 failures/year
US$ 45 Million/Year in failure cost,
exclusive of lost production
#PIWorld ©2019 OSIsoft, LLC
ESP Data Object Model Structure
#PIWorld ©2019 OSIsoft, LLC
Identification of Valid Business Rule
#PIWorld ©2019 OSIsoft, LLC
#PIWorld ©2019 OSIsoft, LLC
With PI AF Asset Templates, Create QuickDisplays
• Can drill down to see what went wrong for each pump using context relative templated displays for scalability and ease of use.
• End user can create their own displays
#PIWorld ©2019 OSIsoft, LLC
Prescriptive
YPF – iUp intelligent Upstream
#PIWorld ©2019 OSIsoft, LLC
iUp Web
Well overview (PI VISION)
#PIWorld ©2019 OSIsoft, LLC
Well Torque (PI WebAPI) ESP Curves (PI WebAPI) Goodman (PI WebAPI)
ESRI Maps (PI WebAPI)
iUp Web
#PIWorld ©2019 OSIsoft, LLC
Widgets
• Torque
• Goodman
• Pump Curves
• Dyna Cards
iUp Web - Beam Pump Torque
• Minimum Net
Torque
• Minimum
CLF
#PIWorld ©2019 OSIsoft, LLC
iUp Web – Beam Pump Goodman
• Metallurgical
stress
• Service Factor
#PIWorld ©2019 OSIsoft, LLC
iUp Web – ESP pump curves
• Real-time data
• Operating
window
#PIWorld ©2019 OSIsoft, LLC
iUp Web: Dyna card diagnostic
#PIWorld ©2019 OSIsoft, LLC
RESULTS
Business Challenge: standardize data countrywide
17K wells in 220 oil fields
CHALLENGE SOLUTION
Collect, transform anddisplay data across the country in a uniform way.
Provide SCADA with template capability and give users awebsite with precise information
Everybody gets the same information, the same way and performs the sameanalysis.
• PI AF
• PI WebAPI
• SQL andOLEDB
• 700+ users
• Engineers moreproductive
• Increase uptime
• Field data collection
• Asset modelling
• Presentation layer
18#PIWorld ©2019 OSIsoft, LLC
Descriptive and Machine Learning
Predictive Maintenance on ESPs
#PIWorld ©2019 OSIsoft, LLC
Approach to Build a Model
#PIWorld ©2019 OSIsoft, LLC
Data Driven Predictions
#PIWorld ©2019 OSIsoft, LLC
• Linked PI data in PI AF with maintenance data in WellView to analyze job cost, runtime and
downtime – identified poorest performing pumps by OEM
• Used text analysis to identify reasons for failures.
• 400 problematic wells. Downhole Heat Exchangers
Diagnostic Analysis on ESPs
#PIWorld ©2019 OSIsoft, LLC
• Data scientists derived model features from PI data, static well and ESP attributes
• Features fed to a ensemble of standard ML models within Microsoft Azure Machine Learning Studio
• Data-driven predictive models capable of predicting ESP failure 60 days in the future
Predictive Analysis on ESPs
#PIWorld ©2019 OSIsoft, LLC
PI Data
Pump Vibration
Motor Temperature
Drive Frequency
Well Head Pressure
Flow Line Pressure
Pump Inlet Pressure
Pump Inlet
Temperature
Average 3 Phase
Current
Average 3 Phase
Voltage
Drive Output Voltage
Earth Leakage Current
Drive Output Current
Power Factor
#PIWorld ©2019 OSIsoft, LLC
Static Well Data
Oil Density
Well Test Basic
Sediments and Water
Well Test Gas Oil Ratio
ESP Metadata
Make
Manufacturer
Location
…
Data in the ESP Models
ESP Decision Tree
#PIWorld ©2019 OSIsoft, LLC
• On September 21, 2015 model predicted 6 wells would fail during next 60 days. As of November 5, 2015 (45 days
into prediction) 4 of the 6 wells were offline - 3 wells failed and 1 well had been taken down for maintenance. 60%+
success rate after first training
• Predictive insights were further analyzed in PI Vision and enabled targeted maintenance scheduling
• Predictions allowed outage for maintenance to be reduced from 30 to 21 days per pump across 1,100 wells avoiding
$millions in revenue loss
• Better decisions could be made around future ESP purchases and service contracts
Predictive Analysis on ESPs - Results
#PIWorld ©2019 OSIsoft, LLC
Overview: Transforming Data into Actionable Insight
Raw,
collected
dataPI System data as
well as external data
1. OSIsoft PI System on Microsoft Azure
Data-guided
decisions for
greater
operational
efficiency and
business profit
Advanced modeling
with machine
learning
e.g., Time to failure,
performance
forecasts
2. Prepare Data 3. Model Data 4. Visualize Insights
#PIWorld ©2019 OSIsoft, LLC 32
Flow Assurance and ML
Why has my well declined/died? …
Gas
Rat
e,e
3m
3/d
Wat
er
Rat
e,m
3
30
25
20
15
Gas Rate
Water Rate10
5
0
Date
#PIWorld ©2019 OSIsoft, LLC
• Liquid Loading?
• Backpressure?
• Scaling ?
• Hydrates?
• Corrosion?
• Reservoir?
… And how do I get it back on production?
Hydrate/Scale Prediction
#PIWorld ©2019 OSIsoft, LLC
Productivity Impairment – Scale…Predictive
(Calcium carbonate, calcium sulfate, barium sulfate mostly)
Shower example
Slotted Liner
#PIWorld ©2019 OSIsoft, LLC
Process Interruption & Operational IssuesCase Study: Reduce Process Interruptions & Optimize Assets
Process Pressure Control Fluctuations
Indicate Actual HydrateFormation
Actual Temp ~41 Deg
Hydrate formation Temp ~48 Deg F
24 hours
• Process
Temperature
Increase
• Control Stability
Restored
Background
•
•
•
•
Transports “Wellhead” Gas with High Concentration of
C3+ Hydrocarbons
Hydrates Can Form in “Heavy” or “Wet” Gas Applications
Hydrates Can Interrupt/Block Gas Flow in Piping & Equipment
Eliminate or Affect Hydrate Formation Line Via:
1. Introduction of Inhibitor/Methanol
2. Change Process to Avoid Hydrate Formation Area
Solution
• Leverage PI Analytics to PredictHydrate
FormationTemperature
• Use PI Notifications to Monitor Process/Hydrate
FormationTemperature
• Alert Key Personnel to PotentialHydrate
Formation
• Modify Process Accordingly to Avoid
Interruptions/Upsets
Results
#PIWorld ©2019 OSIsoft, LLC
• Modify Operations in Response to Notifications
• Reduce Operating Pressures to Avoid Hydrate
Formation
• Reduced Capacity Vs. CompleteOutage
• Minimize Dependency on Hydrate Inhibitor
• Reduce Costs &Consumption
• Eliminate Secondary Impacts of Inhibitor
PI System in AWS,Interfaces on Premises
IoT
Devices
IoT
Core
Redshift Kinesis
Quicksight
SageMaker Amazon
EMR
Kafka
Athena
S3
PI Integrator for Business Analytics
PI System
Primary network ingress
© Copyright 2019 OSIsoft, LLC© Copyright 2019 OSIsoft, LLC 37
Automating Prediction Retrieval
SageMaker Amazon API Gateway
Lambda
PI AFCustom Data
Reference
PI Vision
© Copyright 2019 OSIsoft, LLC© Copyright 2019 OSIsoft, LLC 3388
39
ConclusionsMachine learning needs large amounts of data