Research On Reservoir Identification Method Based On Convolutional Neural Networks

Accurate identification of oil and gas reservoirs plays an important role in oil and gas exploration.Based on the seismic prediction of deep reservoirs,the difficulty lies in the identification or information extraction of seismic response to pore fluid signals of the reservoir.The existing methods still have deficiencies,so the development of targeted research of new methods and technologies for reservoir identification is urgent.Deep learning is a state-of-the-art algorithm that can discover potential features inside the data.It has had many successful experiences in other fields,especially in computer vision.It has powerful feature learning ability and efficient expression of data.In recent years,it is gradually starting to emerge in the field of geophysics.Based on this,in order to fully exploit the reservoir information contained in seismic data and improve the utilization of that,this paper innovatively applies deep learning to reservoir identification problems.The analysis to high-level features extracted from deep learning shows that the DCAE(Deep Convolutional Neural Network)model proposed in this paper can be used to improve the detection and identification of oil and gas reservoirs.The main research contents of this article are as follows:Firstly,studying the theories of Convolutional neural networks and auto-encode networks and using the current excellent deep learning library TensorFlow,a deep convolutional auto-encode(DCAE)model is designed and written.The model is divided into three phases: the coding phase,the decoding phase,and the reconstruction phase.The coding phase and the decoding phase are symmetrical structures,including an input layer,an output layer,and a 17 hidden layers.The neurons in the hidden layers are similar to the traditional convolutional neural network.In the encoding phase,three sets of convolutional and pooling layers and coding full-connection layers are included.After the expression layer,the decoding phase includes the decoding full-connection layer and 3 sets of deconvolutional and unpooling layers,and finally the image is mapped to the neuron output by a non-linear function.By entering test data,the learned features are verified during the reconstruction phase.Secondly,in order to make the DCAE model obtain more accurate characteristics learning results,the operation mechanism of deep learning such as network hyperparameters,network structure,training sample space,and data distribution was analyzed and discussed.Through a large number of experiments,the optimal combination for the current reservoir identification task is obtained.DCAE judges the parameters by reconstruction results.The experiments have shown that when the model's learning rate is 0.0005,the training frequency is 50,000,the training volume is 20,and the convolution kernel is set to 3×3,the reconstruction result of the model is better,which can be thought that the model has fully learned to obtain the internal characteristics of the training data and comparedly instruct a certain directional meaning.These parameters pave the way for follow-up works.Thirdly,deep convolutional auto-encode is an unsupervised deep learning model,but the characteristic learning results lack directionality.In order to make the results of unsupervised learning easier to interpret,the correlation analysis of features is innovatively used to determine the deeper features what are more directional to the identification task of oil and gas reservoirs.Using seismic data,the internal features were learned through deep convolutional auto-encode learning,and reservoir descriptions of carbonate fracture-cavity-type oil and gas reservoirs in the Tahe workplace were conducted.The experimental comparison revealed that the characteristics from deep convolution layers with mid-high correlation coefficients can better indicate the distribution of the reservoir and achieve a higher recognition rate.The results show that the deep convolutional auto-encode network is feasible and effective for reservoir identification.It provides a new idea for reservoir identification and has novel academic and technical value.