Visualize Complexity, Discover Solutions, Shatter Limits

Case Study: Colorado Coal Gas

Client: 

Coal Bed Methane Producer, Colorado   
Project:

Predicting Methane Gas Reserves
Issues:

A critical problem for production companies is to defensibly calculate their reserves. The result strongly impacts a company’s market value and ability to attract investment capital. The calculation must be repeated time and again as properties are bought, drilled, produced, and sold. Engineers commonly use reservoir simulators to help with the calculation. Reservoir simulators use equations from physics to predict how much gas a well will produce over its lifetime. Typical inputs to the simulator are variables such as depth, permeability, porosity, and estimated gas-in-place. While reservoir simulators are important, they do have a few problems. For a specific well, the actual bottom-hole process physics can be different from that described in the simulator. Often the values of critical inputs such as permeability are poorly known. Both problems can significantly impact the accuracy of the simulator. Simulators can be laborious to set up and run, discouraging routine use.  
ADMi's Role:  An engineering team was tasked with predicting reserves in a coal-bed methane field of several hundred wells. Lengthy run times precluded simulating every well. A study was established to determine if data mining could produce a predictive model that would be accurate and fast enough to calculate the combined reserves of all of the wells. Simulations were run on 60 study wells and the predicted gas produced was compiled into a database along early stage actual gas (GAS-H) and water (WATER-H) production histories for each well. Correlation analysis indicated the estimated gas-in-place (GIP) was by far the most important simulator input. Analyses of GAS-H and WATER-H showed well-to-well similarity, but gas and water production rates varied by 1000%. The predicted variable, the cumulative gas produced (CGP), varied similarly. Correlation analysis indicated that the only variables needed to accurately predict CGP were GIP, GAS-H, and WATER-H. ADMi used a machine learning method called an “artificial neural network” (ANN), configured using “phase space reconstruction” from Chaos Theory, to create a predictive model of the wells

Results:   

Using just three variables, GIP, GAS-H, and WATER-H, the model’s predictive accuracy exceeded 90%. It could calculate CGP for all 800 of the field’s wells in just a few seconds. ANN-based models are compact programs that easily integrate with commercial databases, and could execute automatically whenever a gas production company updates its databases to provide near real-time estimates of the reserves for a single field or all production operations.  

 

Home |  Products  |  Services  |  Training  |  Technologies  |  News  |  Links  |  Contact

Send mail to info@advdataming.com with questions or comments about this web site.
Copyright © 2003 Advanced Data Mining, LLC