[First Hit]

2019 AAPG Annual Convention and Exhibition:

Datapages, Inc.Print this page

Application of Neural Networks and Machine Learning in Tiltmeter Analysis in Hydraulic Fracturing Diagnostics

Abstract

This paper presents a new machine-learning method of processing surface or downhole deformation data during hydraulic fracturing to obtain more reliable fracture diagnostic information such as frac length and azimuth in real time. The results are then compared to the current method, the inversion of fracture models, by using synthetic and actual data. Several types of neural networks are designed, based on the type of known data and the desired output. These types include feed-forward, time-Previous HitdelayNext Hit, and pattern recognition neural networks. The number of hidden layers varies between 5 and 20, and number of training samples varies from 10,000 to 50,000 samples. After training, the network was tested against known fractures, and the error of fracture properties was measured. Gaussian noise was added to the input deformation data to simulate real-world conditions. Four fracture systems are defined in this work: a) single vertical fracture, b) dual vertical fractures, c) vertical + horizontal fractures, and d) single horizontal fracture. Using a pattern recognition network, 448 of 500 test cases with all possible fracture systems were recognized correctly, which equates to only a 10% error. In the next step, the fracture azimuth was estimated, based on surface tilt data for a single vertical fracture with additional noise on tilt data; the resulting estimate had an error rate of less than 5%. Next, the fracture volume was evaluated from surface tilt-magnitudes with a very low error rate (less than 2%). Fracture half-length and width were also estimated from surface tilt-magnitudes; the error was greater than in the previous steps because the half-length and width are almost interchangeable from the tilt-magnitude point of view. With downhole tiltmeters, the evaluation of fracture height was off by less than 10% for the height and 1% for the TVD of the fracture. Finally, two methods are presented to evaluate the network to obtain the probability of results and the uncertainty of output values for each case. Considering the low error rates and uncertainty in general, neural networks can be used in several aspects of tilt analysis. No applicable method currently exists with which to evaluate fracture properties (azimuth, half-length, height, and proppant volume) from deformation data in real time. This new machine-learning method can be used to quickly perform the deformation data analysis, rather than the current, more manual, inversion fracture models that search for the Previous HitminimumTop error.