Recognition And Prediction Methods Of High-quality Reservoirs Of Tight Sandstones In Shaximiao Formation In Chuanxi Depression

The development of tight sandstone reservoirs is seriously affected by the reservoir quality.Accurate recognization and prediction of high-quality reservoirs are important for efficient development of tight sandstone reservoirs.Although traditional reservoir characterization methods can provide a set of existing solutions,there are still many problems to be solved for complex tight sandstone reservoirs,such as low accuracy of well log interpretation,low resolution of seismic prediction.This thesis focuses on sandstone gas reservoirs in the Shaximiao Formation in the Chuanxi Depression,takes deep learning and machine learning as main techniques,and finally proposes a set of novel data-driven intelligent reservoir recognization and prediction methods.These methods help to improve the accuracy and efficiency of the recognization and prediction of high-quality reservoirsThe isolated channel tight sandstone of the Shaximiao Formation has strong heterogeneity in rock mineral composition,porosity,permeability and pore structure According to the controlling factors of reservoir quality,six types of petrophysical facies are summarized.Three rock types are obtained by data-driven clustering method using porosity,permeability and FZI.A comprehensive classification of tight sandstone reservoirs combines the petrophysical facies with the clustering rock types.The rock types ? and ? are treated as high-quality reservoirs.A novel intelligent unified well log interpretation model of the Shaximiao Formation is built based on deep learning Compared with other machine learning methods,this method improves the accuracy and efficiency of well log interpretation.The mean relative error of the predicted permeability of blind wells reduces from 1.16 to 0.53.Sensitivity analysis and type curve analysis methods,which are novel mechanism interpretation methods of deep learning,are proposed for well log interpretation models.Analyzing the deep learning permeability model Black Box,it is believed that deep learning model has learned the influence of pore structure on permeability from well logs.For seismic lithology and high-quality reservoir prediction,four different deep learning inversion methods are proposed:DNN,CNN,CWT-DNN and CWT-CNN.Combing continuous wavelet transform and convolutional neural networks,CWT-CNN model performs the best for the lithology and high-quality reservoir prediction of the Shaximiao Formation,especially for medium-thin channel sands.The mean relative error of the CWT-CNN predicted sand thickness larger than 5m,reduces from 0.66 to 0.34.The mean relative errors of the CWT-CNN predicted rock types ? and ? thickness larger than 5m,reduce from 0.75 and 0.72 to 0.34 and 0.15,respectively.CWT-CNN is a high-resolution intelligent seismic inversion methodThis paper proposes a set of novel data-driven intelligent reservoir characterization methods,which have been applied in the study of the Shaximiao Formation and achieved good results.It demonstrates the feasibility and potential of data-driven methods based on deep learning,which have broad application prospects in the study of oil and gas exploration and development.