Application of Machine Learning for Production Forecasting for Unconventional Resources
- Cheng Zhan (Anadarko Petroleum Corporation) | Sathish Sankaran (Anadarko Petroleum Corporation) | Vincent LeMoine (Anadarko Petroleum Corporation) | Jeremy Graybill (Anadarko Petroleum Corporation) | Didi-Ooi Sher Mey (Anadarko Petroleum Corporation)
- Document ID
- Unconventional Resources Technology Conference
- SPE/AAPG/SEG Unconventional Resources Technology Conference, 22-24 July, Denver, Colorado, USA
- Publication Date
- Document Type
- Conference Paper
- 2019. Unconventional Resources Technology Conference
- 23 in the last 30 days
- 205 since 2007
- Show more detail
|SPE Member Price:||USD 9.50|
|SPE Non-Member Price:||USD 28.00|
Decline curve analysis is often the de facto method for large scale production rate forecasting based on empirical relationships. Often, we face a number of practical problems for reliably estimating rates due to operational changes in the field (changing chokes, common surface network), curtailed production, limited available historical data, complex flow behavior, well interference etc. Our main objective here is to develop a new methodology to make robust and accurate oil rate prediction based on limited initial production data. We will show that the resulting model is useful for production forecasting, business planning and decision making in response to the fast pace development for unconventional resource plays.
A machine learning approach based on LSTM (Long Short Term Memory) is used to tackle the production forecasting problem. Compared with the modified hyperbolic approach, where the problem has been reduced to a pre-defined equation and essentially determined by a global curvature structure, the LSTM model is more dynamic and has a better chance of capturing non-linear events. In time series prediction, one main difficulty is how to stabilize the solution, as the error can easily accumulate over time. Besides modifying the objective function that aims for long term accuracy or incorporating physics-based modeling, one effective way to make the algorithm more robust is through feature engineering. By leveraging historical data from other wells, the prediction has been improved significantly. We also build another model in the accumulated curve domain, and ensemble multiple models to reduce the variance.
Forecasting is highly challenging in many domains with complex multivariate correlation structures and nonlinear dynamics. We have utilized existing data and built two prediction models, one from the decline curve domain, the other from the accumulated curve. Based on the observation, the first model is slightly over-predicted, and the second one moderately under-predicted, and through integrating these two models, the final result is more promising. We have conducted hindcasting for more than 300 wells, and the mean difference between the predicted and actual accumulated production of the first 2 year is less than 0.2%, with the variance less than 5%.
Many empirical production decline models have been proposed in the literature, but most fail to capture the complexity of forecasting and reduce the problem to an over-simplified curve. Our data-driven procedure is a unique and novel approach, which is more dynamic and has a better chance of capturing non-linear events. This method can also be applied to conventional reservoirs.
|File Size||976 KB||Number of Pages||10|