Extreme learning machine for multivariate reservoir characterization

Reservoir characterization is one of the most important tasks in oil and gas field exploration and development. Different parameters reflect the relevant information of oil and gas fields from diversified aspects. We design a new reservoir characterization framework by introducing extreme learning machine (ELM) that is one of the state-of-the-art methods in machine learning. It is a single hidden layer feedforward neural (SLFN) network, while the input weight and the bias value of the hidden layer are randomly assigned and kept fixed for simplifying the calculation. Based on ELM, we achieve simultaneous prediction of multiple reservoir parameters (including lithofacies, porosity, shale content and saturation etc.) only through one training step. In order to combat overfitting when the number of hidden nodes is inappropriate or the training samples are inadequate, we extend the method by using biased dropout and dropconnect operations to regularize ELM. We describe the new method in detail and analyze its performance with varying input parameters. It is evaluated on well and seismic datasets by exploiting elastic attributes as training input. Compared with traditional SLFN-based method, ELM-based method uses less computational resources and costs less time on training without losing accuracy. The biased dropout and dropconnect operations further enhance the generalization ability.