Model-free estimation of fracture aperture with neural networks |
| |
Authors: | Marek Kacewicz |
| |
Institution: | (1) ARCO Exploration and Production Technology, 2300 W. Plano Parkway, 75075 Plano, Texas |
| |
Abstract: | The feedforward backpropagation technique provides a model-free estimation with neural networks. The algorithm was used to estimate fracture aperture of natural fractures in three dimensional space. A three-layer neural network with at least 5 nodes in a hidden layer was trained on a data set consisting of formation imaging microscanner logs (FMS)from horizontal boreholes. Sensitivity studies were performed to account for the rate of learning convergence, convergence to local error minima, etc. Among the factors contributing mostly to the overall good or bad performance of the network, the following are worth mentioning: number of data points, data spacing, and data variability. It is shown that a smoothing operation applied to aperture data along the wellbore often helps to reduce disorientation of the network and to switch from oscillations or chaotic jumps to convergence. |
| |
Keywords: | naturally fractured reservoirs fracture networks fracture apertures neural networks feedforward backpropagation |
本文献已被 SpringerLink 等数据库收录! |
|