Rock porosity is one of the critical parameters to characterize the reservoir.High-precision prediction of porosity is con-ducive to more detailed description of the location of highly porous and permeable reservoirs.Due to the strong heterogeneity and complex pore structure of the reservoir,there are many factors affecting the reservoir porosity,which brings difficulties to the ac-curate prediction of reservoir porosity.In recent years,the development of deep learning has provided a new idea for high-precision seismic porosity prediction.Accordingly,this paper presents a prestack seismic porosity prediction method based on the bidirection-al gated recurrent unit neural network and attention mechanism(BiGRU-Attention).In this method,BiGRU is used to realize the bidirectional propagation of information and the attention mechanism is added to amplify the key information.The P-wave velocity and density information obtained from the prestack simultaneous inversion is used as the input and the logging porosity value is used as the label to train and test the BiGRU-Attention network,so as to obtain an optimal model.This method establishes a com-plex mapping relationship between seismic elastic parameters(P-wave velocity and density)and porosity to achieve high-precision prediction of porosity.The actual data test results showed that compared with the conventional multiple linear regression(MLR),dense neural network(DNN)and gated recurrent unit neural network(GRU),the proposed method based on BiGRU-Attention had higher prediction accuracy in blind well testing.The root mean square error(RMSE)of prediction results and logging data was less than 0.0022,and the average absolute error(MAE)was less than 0.001 4.Application of this method to the seismic data of an actual 3D work area demonstrated that the predicted porosity profile matched well with the logging porosity value,indicating that the method has good practical value.
deep learningattention mechanismbidirectional gated recurrent unit neural network(BiGRU)porosity predictionreservoir parameter inversion
维普
万方数据
