Abstract Decline curve analysis has been used as a reliable method to forecast conventional reservoir well production over the last decades. Recently, an increase in the demand for oil and gas has caused unconventional reservoirs to become a prominent source of energy. However, it is challenged if we still apply the decline curve analysis in unconventional reservoirs due to its limitations such as boundary dominated flow, constant operation condition, et al. Therefore, in this paper, two new methods are proposed using machine learning method to forecast well production in unconventional reservoirs, especially on the EOR pilot projects.
The first method is the Neural Network, which allows the analysis of large quantities of data to discover meaningful patterns and relationships. Both peak production rate and hydraulic fracture parameters are used to be the key factors. Lastly, Neural Network technology is applied to investigate the relationship between key factors and oil production rate. The second method uses the Time Series Analysis. Time Series Analysis is one of the most applied data science techniques in business and finance. Since the properties of unconventional reservoir make the production prediction more difficult, it is safe to say that Time Series Analysis can yield good results on the production rate forecast.
Field production data from over 1000 wells from different shale plays (Barnett, Bakken, Bone Springs, Eagle Ford oil, Eagle Ford gas, Fayetteville, Marcellus gas, Marcellus oil, Utica oil, and Woodford) is used to verify the feasibility of these two methods. The results indicate there is a good match between the available and predicated production data. The overall R values of Neural Network and Time Series Analysis are above 0.8, which demonstrates that Neural Network and Time Series Analysis are reliable to study the dataset and provide proper production prediction. Meanwhile, when dealing with the EOR production prediction, such as Huff-n-Puff, Time Series Analysis shows more accurate results than Neural Network.
This paper proposes a thorough analysis of the feasibility of machine learning in multiple unconventional reservoirs. Instead of repeatedly fitting the production data by decline curve analysis, it also provides a more robust way and meaning reference for the evaluation of the wells.