Study On Deep Learning For Seismic Prediction Of Fractures In Deep Carbonate Reservoir Based On Optimized Samples

Deep fractured carbonate oil and gas reservoirs,as one type of important subtle reservoirs,have great exploration potential and gradually become the hotspot of hydrocarbon exploration and development.Improving the accuracy of seismic fracture prediction and accurately describing the spatial distribution of fractures is the prerequisite for oil and gas development in deep carbonate reservoirs.As a hotspot in the field of artificial intelligence,deep learning has achieved remarkable application in seismic reservoir prediction in recent years.The important conditions that support deep learning applications are high quality and,complete sample sets and learning network structure that adapts to the characteristics of sample sets.However,deep seismic data suffer from low signal-to-noise ratio,limited frequency band and weak response characteristics of fractures.In particular,the number of deep drilling is limited,the target samples in well are insufficient,which makes it difficult to establish a high quality and complete sample set.Meanwhile,deep learning is prone to overfitting and poor generalization in fracture prediction of deep carbonate reservoirs.Therefore,constructing an optimized sample set with seismic data and logging data,designing a suitable network structure,and carrying out the exploratory research on fracture prediction in deep learning,have great significance for solving the difficulty of fracture prediction in deep carbonate reservoirs.This paper first uses planar structural gradient anisotropic diffusion filter and adaptive regularization-constrained compressed sensing to effectively suppress random noise and enhance the low-frequency weak signal.Based on this,multi-scale wavelet decomposition is applied to increase the number and scale of seismic samples,which lays the high quality samples foundation for deep learning fracture prediction.Based on the Backus average theory,the well target samples that can effectively characterize the fracture anisotropy are then constructed.Through the reliability analysis of the fracture-sensitive seismic attributes,the seismic attributes that characterize the “facies” of the fracture spatial distribution are quantitatively optimized.Interpolation constrained by the optimized attribute is used to increase the number of target samples,which provides a spatially complete target sample for deep learning fracture prediction.Based on the multi-scale and spatially complete sample set,applicability of the deep belief network in the identification of fracture feature parameters is discussed,which proves that the constructed optimized sample set can effectively solve the problem of overfitting.Furthermore,an optimal deep belief network structure that adapts to the identification of fracture parameters is designed.Finally,in order to further improve the accuracy of the deep belief network fracture prediction,an accumulating training strategy is proposed,and fracture parameter inversion of deep carbonate reservoirs is implemented.Model tests and the case studies of fractured carbonate reservoirs of Area S show that,(1)the planar structural gradient anisotropic diffusion filter and the low-frequency compensation based on adaptive regularization-constrained compressed sensing are conducive to the enhancement of weak signals.Moreover,multi-scale seismic samples processed by wavelet transform based on weak signals enhancement can improve the accuracy of fracture prediction using deep belief network.(2)The attribute constrained interpolation method can be used to extrapolate the interpolation of the target sample with finite wellpoint fracture characteristics according to the optimized attribute.Compared with solely seismic input samples and borehole fracture parameters,spatially complete fracture target samples contribute to higher accuracy of fracture prediction for deep belief network.(3)The optimal deep belief network structure and the cumulative learning training can maximize the accuracy of fracture prediction,contributing to 87.5% of coincidence rate between the results of actual drilling.The predicted distribution is in line with the qualitatively described fracture zones using the optimal seismic attributes.The deep learning technology series driven by well-seismic data proposed provide a theoretical and practical basis for the prediction of fractures in deep fractured carbonate reservoirs.