An Efficient Approach for Automatic Parameter Inversion Based on Deep Learning

Abstract

Abstract Parameter evaluations are the first and primary tasks to understand the natural gas hydrate reservoirs. However, there still lacks some effective means for parameter evaluations in hydrate reservoirs. To improve this situation, this paper tries to combine the well testing with deep learning (DL) method for solving parameter inversion problems of natural gas hydrate wells. First, a radially-composite well testing model with dynamic interface is developed to represent the hydrate dissociation driven by depressurization. Then, by Laplace transform, the wellbore pressure is solved and adopted to train a one-dimensional convolutional neural network (1D CNN) and the optimal convolutional neural network (CNN) is obtained by minimizing mean square error. In the CNN, the wellbore pressure is used as input of the network after nondimensionalization, and the interpreted parameters are permeability, wellbore storage coefficient, skin factor and dissociation factor. Finally, the well testing and DL method is verified and applied in a field case. Results show that the sensitivity of the parameter on pressure transient behavior will affect the accuracy of parameter inversion. The 1D CNN is tested with synthetic data, which shows great practicality and high accuracy of curve matching. During the field application, when compared with manual match, the relative errors of wellbore storage coefficient and dissociation factor by the proposed method are 4.863% and 1.933%, respectively. The proposed well testing and DL method is proven to be suitable for problem inversion of natural gas hydrate wells, which may provide a new tool for engineers to understand the natural gas hydrate reservoirs.