Research On Application Of Sequence Data Machine Learning Method In Rock Physics Parameter Prediction

The prediction of petrophysical parameters has always been a hot spot in oil and gas exploration.With the deepening of oil and gas exploration work,the development of lithological oil and gas reservoirs has gradually become the focus of current energy exploration,and lithological oil and gas reservoirs often require more precise petrophysical parameters,which puts forward higher requirements for parameter prediction models.Traditional petrophysical parameter prediction models often contain more empirical values and empirical formulas,and do not have strong generalization.By introducing machine learning methods,combined with deep learning models and attention mechanisms,the sequence characteristics of logging parameters can be effectively captured,and a high-precision sequence parameter prediction model can be constructed.The main research contents of the thesis are as follows:Firstly,the reservoir characteristics of the study area are analyzed.By analyzing the regional structural characteristics,sedimentary characteristics,reservoir characteristics and logging response characteristics of the reservoir,the actual geological meaning is provided for the petrophysical parameter prediction model.Then introduced the commonly used models in deep learning models and constructed lithology recognition models based on machine learning and deep learning in carbonate rocks and clastic rocks.And through multiple classification evaluation indicators,analysis methods differ in the recognition effect of different reservoirs.Secondly,analyze the role played by conventional machine learning in shear wave regression,and design a number of plans to further explore the role of logging parameter neighborhood characteristics and lithological characteristics.The basic ideas of a variety of machine learning methods are introduced and applied to shear wave recognition tasks based on logging parameters.In the analysis of the results of transverse wave regression,multiple evaluation indicators are introduced to analyze the performance of different machine learning methods from multiple angles.On the one hand,the log domain information is used to enhance the sequence feature expression of logging parameters,and on the other hand,the lithology sequence is introduced to constrain the shear wave regression results.By constructing a variety of feature blending schemes based on logging neighborhood information and lithology information,the specific role of different features in shear wave regression is analyzed.Finally,a deep learning shear wave regression model is constructed and two attention mechanisms are embedded to analyze the adaptability between the attention mechanism and the model.First,build a deep learning model based on convolution,and analyze the role of different improved convolution modules in shear wave regression by replacing the convolution module.Then channel attention and spatial attention are introduced to different positions of the network model to discuss the adaptability of the attention mechanism in combination with the deep learning model,and the different performance of attention in different positions.In this thesis,conventional machine learning and deep learning methods are introduced in the prediction of petrophysical parameters to realize the accurate prediction of lithology sequence and shear wave sequence.On this basis,the matching problem of attention mechanism and deep learning model is analyzed.The refined interpretation provides a foundation and new ideas for sequence data prediction.