Event Recognition on Time Series Frac Data using Machine Learning – Part II
- Alberto Ramirez (Well Data Labs) | Jessica Iriarte (Well Data Labs)
- Document ID
- Society of Petroleum Engineers
- SPE Liquids-Rich Basins Conference - North America, 7-8 November, Odessa, Texas, USA
- Publication Date
- Document Type
- Conference Paper
- 2019. Society of Petroleum Engineers
- Hydraulic Fracturing, ISIP, Machine Learning
- 3 in the last 30 days
- 109 since 2007
- Show more detail
- View rights & permissions
|SPE Member Price:||USD 5.00|
|SPE Non-Member Price:||USD 28.00|
Hydraulic fracturing pumping data is recorded in the field at one-second intervals. Engineers spend hours identifying events such as Instantaneous Shut-in Pressure (ISIP) in the time-series data that is generated. The ISIP flag is placed at the end of the stage pumping time, immediately after shut-in and before the pressure starts to drop. This is estimated by placing a straight line on the early pressure decline and locating the point in time where the pressure rate is zero. Manual selection of this flag is time-consuming, prone to error, and inconsistent due to differing interpretation methods across the industry. The purpose of this study is to demonstrate an automated process to identify accurate and consistent ISIP events in a high-frequency time-series data set using machine learning algorithms.
This study is based on the analysis of metered high-frequency fracturing treatment data from wells landed in different formations across North America coupled with supervised machine learning algorithms. The pumping data includes treating pressure and slurry rate for 870 stages from the Wolfcamp, Bone Spring, Granite Wash, Barnett, Meramec, Niobrara, Codell, Bakken, Three Forks, Haynesville, Bossier, Caney, and Marcellus formations, for a total of over 7 million rows of data per channel. Eighty percent of the data is used to train the model, seven percent is used for validation, and the remaining thirteen percent forms the test set used for the final evaluation. To allow the algorithm to run leaner, the dataset was pre-processed using smoothing techniques, and the rate of change of the main data channels were added. The selected algorithm, an artificial neural network (ANN), was trained to recognize and isolate the necessary data from the treating plot that will be used to predict the ISIP. Once the data is isolated, a filter is used to extract the portion of the data to be used. A second machine learning algorithm, linear regression, is then applied to the portion of extracted data to predict the ISIP value when the slurry rate is equal to zero.
Classification techniques were used to generate an accurate suggestion of the reduced dataset needed to recognize the ISIP event in a high-frequency treating plot. The neural network achieved a classification accuracy (on the training and validation sets) of approximately 98 percent when isolating the target region. The subsequent ISIP predictions from the linear regression on the test set had an average accuracy of +/- 50 psi when compared to the manually picked values. Considering that the typical range for ISIP values is between 2,500 psi and 9,000 psi, 50 psi represents a 0.5% to 2% error. A limitation of this method is that it requires periodic re-training with new field data to improve the prediction robustness and to maintain high accuracy.
Automatically labeling relevant regions of high-frequency hydraulic fracturing treatment plots using classification techniques can lead to simple and effective procedures for identifying events of interest. Accurate flag selection makes processing large volumes of fracture treatment data viable and significantly reduces the time spent reviewing field data for quality control. The method will also allow rapid reprocessing of historical data. The benefits of using simple (and accurate) models include ease of deployment, ease of debugging, and extremely fast prediction and re-training (updating the model).
|File Size||1 MB||Number of Pages||11|
Bushaev, V. 2018. Adam - latest trends in deep learning optimization. Towards Data Science, 22 October 2018, https://towardsdatascience.com/adam-latest-trends-in-deep-learning-optimization-6be9a291375c (accessed 02 June 2019).
IHS-Fekete, 2014. Minifrac Pre-Closure Analysis, Instantaneous Shut-In Pressure. http://www.fekete.com/san/webhelp/welltest/webhelp/Content/HTML_Files/Analysis_Types/Minifrac_Test_Analyses/Minifrac-Pre-Closure_Analysis.htm#ISIP (accessed 02 June 2019).
Iriarte, J., Merritt, J., & Kreyche, B. (2017, March 27). Using Water Hammer Characteristics as a Fracture Treatment Diagnostic. Society of Petroleum Engineers. doi: 10.2118/185087-MS
Kholy, S. M., Mohamed, I. M., Loloi, M., Abou-Sayed, O., & Abou-Sayed, A. (2019, February 1). Development of an Empirical Equation to Predict Hydraulic-Fracture Closure Pressure from the Instantaneous Shut-In Pressure Using Subsurface Solids-Injection Data. Society of Petroleum Engineers. doi: 10.2118/187495-PA
Ramirez, A., & Iriarte, J. (2019, April 22). Event Recognition on Time Series Frac Data Using Machine Learning. Society of Petroleum Engineers. doi: 10.2118/195317-MS
Roussel, N. P. (2017, January 24). Analyzing ISIP Stage-by-Stage Escalation to Determine Fracture Height and Horizontal-Stress Anisotropy. Society of Petroleum Engineers. doi: 10.2118/184865-MS