Abstract
Data-driven deep learning algorithms have shown good performance in the field of petroleum industry. However, some research has begun to be keen to incorporate physical laws into machine learning algorithms, so as to establish a "data + physical laws " dual-drive model, which can more effectively guide deep learning. In this study, reservoir geology, hydraulic fracturing, and dynamic production data were considered to establish a fracturing effect evaluation model for coalbed methane reservoirs. The combined network is designed to fully excavate the characteristics of dynamic and static data and solve the problem that the network ignores static data due to excessive dimensions of dynamic data. Furthermore, a neural network considering physical constraints was developed to better evaluate the fracturing effect by incorporating the initial conditions and expert experiences into the loss function. The deep learning-based fracturing effect evaluation model not only fits data driven methods including reservoir geology, hydraulic fracturing and dynamic production data, but also adheres to the guidance of physical constraints. The experimental results show that compared with the conventional machine learning methods, the fracturing effect evaluation model has better performance on the prediction of crack half-length and permeability after fracturing due to combined network and physical constraints, with the overall RMSE of 6.11 m and 0.533mD respectively. In addition, through the analysis of influencing factors, it can be obtained that reservoir geology and hydraulic fracturing parameters can contribute more than 90% to the prediction of fracture half-length. Moreover, reservoir geology, hydraulic fracturing and dynamic data all play an important role in the permeability after fracturing, among which dynamic data has the highest contribution rate, with more than 40%.
NSTL
Elsevier