Transcript ISPA for Auto - 2015 Rice Oil & Gas HPC Workshop

IBM GBS March 2015
Predictive Analytics at Work:
Oil and Gas Exploration using Watson and Data Streaming
© 2014 IBM Corporation
Price of Gas What and Why*
Numerical models use facts
e.g. traditional supply & demand
2
•
Tell most consumers or in fact most analysts the price of oil would go from
$110 to $50 but not justify Why it would not be trusted
•
The first shale oil extraction patent was granted in 16841
•
But it was only written about and not included in the models until recently
•
What if the models could learn beyond classic numerical machine learning
•
If the models could understand the text to know that shale oil is going to
have a critical causal relationship on the price of gasoline
•
That would change the What is predicted but more importantly explain with
the Why
1 http://en.wikipedia.org/wiki/Shale_oil_extraction
2 http://www.nasdaq.com/markets/crude-oil-brent.aspx?timeframe=1y
© 2014 IBM Corporation
Numerical elements to computer predicting gas prices
1. Proven reserves
2. Refining capacity for gasoline
versus diesel
1. High Frequency data on
planned pipeline capacity
2. Streaming data on current
demand
Connectors
Ratio of the price of gas to substitutes
•
Price inflection points on supply
1. It becomes economic to uncap hard
2.
3.
3
to recover oil wells
Profitable for shale oil extraction
Heavy oil / oil sands are affordable
(e,g. natural gas, alternative energy,
diesel, etc.)
Watson
© 2014 IBM Corporation
However you cannot easily get the required rest of the model
elements because people only written in text:
Weather
impacting
demand
Slow down in
Growth
Geography (BRIC)
countries GDP
4
Political problems
Venezuela
If OPEC decides
to restrict supply
• Find new reserves
• Improve reserve
Alaskan
• Improve recovery rates
intervention in
• Improve operation
efficiency
drilling or
• Extend
resource life
pipelines
build
out
management
• Improve oil & gas recovery
• Manage and optimize
Discussion on
completions
•well
Chemical
and thermal
recovery design
(bringing
new oil
• In situ combustion
online)
optimization
injections
© 2014 IBM Corporation
Extraction can find, structure and apply content from text:
•
Weather (the average winter temp in NE down by 2)
•
Political (Strikes by PDVSA oil workers paralyze oil production … has doubled
its workforce since the strike of 2003 even though oil production has stagnated
at well below pre-strike levels.)
•
Supply (OPEC Reference Basket slipped heavily from October’s record peak,
sliding $6.41 or 14%)
•
Geography (In 2007 China’s economy expanded by an eye-popping 14.2%
The IMF now reckons China will grow by just 7.8%)
Legend
5
•
Terms in red can be extracted and put into numerical models (e.g. temp down by 2)
•
Terms in green can be extracted and looked up in a reference domain model and then
applied to numerical models (e.g. OPEC = oil supplier, or NE = US East Coast)
•
Words in black are throw away words
© 2014 IBM Corporation
Everyone cares about the price of gasoline
Model with only numerical elements
Supply
Proven reserves * Refining capacity
---------------------------------------------Pipeline capacity
Current demand
Historical Demand for the period
------------------------------------------------Ratio of the price of gas to substitutes
6
© 2014 IBM Corporation
Everyone cares about the price of gasoline
Model with numerical & unstructured elements
Supply
Proven reserves * Refining capacity
---------------------------------------------- *
Pipeline capacity
OPEC & Venezuela factor * Shale Oil *
New Regulations * New completion rates
Current demand
Historical Demand for the period
------------------------------------------------- *
Weather * China demand * Environmentalism
Ratio of the price of gas to substitutes
7
© 2014 IBM Corporation
There is a lot of information, some of it streaming:
Streaming
Data
True vertical
depth
Fluids Data
Surface
Equipment
Data
BHA
Dynamics
Data
Lithological
Data
Structured
Data
Unstructured
Data
Data
Mapping
MWD / LWD
Data
Vibration
Data
Rock
Mechanics
Data
Mud weight
8
© 2014 IBM Corporation
There is a lot of unused unstructured information:
Streaming
Data
Regulations
Well plans
Manufacturer’s
FAQ
Shift End /
Morning
reports
Best
Practices
Structured
Data
Unstructured
Data
Driller's
Network
Data
Mapping
9
© 2014 IBM Corporation
Probability of mud circulation problems is now much more
complete
Model with numerical & unstructured elements
Classical numerical model for
predicting circulation problems
Written assessment of material in the
*
Material in the shakers, Drill plan notes
on fractures or have high permeability, etc.
10
© 2014 IBM Corporation
Where else can your firm Merge Numerical and Unstructured
Other parts of drilling
Plant maintenance
Exploration
Well planning
Environment
Health, Safety, Security and Environment
Major Capital Projects
Trading advisor
etc.
11
© 2014 IBM Corporation