S-Wave Velocity Prediction Based on Probabilistic Neural Network (PNN) and Deep Feed-Forward Neural Networks (DFNN)

Abstract

ABSTRACT: The accuracy of the shear wave travel time prediction affects the overall accuracy of the elastic parameters, and this will have an impact on the reservoir characterization. Rock elastic parameters are critical to alleviate the risks associated with drilling and wellbore stability. This paper establishes the shear wave prediction model based on two artificial intelligence methods (Probabilistic Neural Network (PNN) and Deep Feed-forward Neural Networks (DFNN)) using Wellington oil field data in south central Kansas. Neural networks are potentially superior to linear or multi-linear regressions, and they can be used to model nonlinear functions. These techniques will be used to enhance the high frequency resolution and prediction capability. DFNN uses multiple hidden layers and offers significant advantages in terms of control of training parameters. More data is required to train DFNNs because of the introduction of additional parameters such as weights, but it proves to be a reliable methodology to quantitatively predict the reservoir’s properties by optimizing the parameters such as the number of hidden layers, the nodes in the hidden layers, and the number of iterations required to solve the problem. To validate the neural network’s approach, a number of wells were hidden, and the trained data was used to predict the S-wave velocities in those wells. The validation correlation shows how well the process will work on a new well that has yet to be drilled. The developed model provides an economical and reliable approach for the determination of geomechanical parameters where S-wave travel times are not available. 1. INTRODUCTION In this study, Probabilistic neural networks (PNNs) were applied to develop an intelligent predictive model for prediction of the shear wave velocity information. PNNs, firstly introduced by McCulloch and Pitts [8], is a mathematical model of biological events used to imitate the capability of biological neural structures with the purpose of designing an intelligent information processing system. An adaptive neural network is a network structure consisting of large number of elemental units, called neurons, organized in input, hidden, and output layers. Any neuron in the network is characterized by some features such as input weights, a threshold, and an activation function. The adjusting weights connect the neurons in different layers, so that a particular input, according to a learning algorithm, leads to a specific target output [8, 9].