Research Of Interpretable Machine Learning Reservoir Prediction Method Based On Big Data

Directly predicting reservoir properties from seismic data is one of the feasible means to simplify the links and reduce the cost in oil-gas exploration and development.Due to the complex nonlinear relationship between seismic data and reservoir properties,traditional reservoir prediction methods based on seismic and well-log data have limited resolution in reservoir characterization and hydrocarbon prediction.The artificial intelligence(AI)techniques represented by machine learning(ML)have the advantages to automatically extract data features and uncover the intrinsic high-dimensional nonlinear mapping between the input and the output data,which provides a novel and feasible approach for reservoir prediction.However,the ML methods have the“end-to-end” feature,which makes the intermediate process of reservoir prediction difficult to be controlled and to be understood.And the results of the methods may be sufferred from multiple risks including unreliability,instability,and uncontrollability.In addition,ML methods usually need sufficient samples and labels to train the network,which is unattainable in the application scenario of reservoir prediction.Poor physical interpretability and extremely insufficient labels are the two key scientific problems faced by the ML reservoir prediction methods.To partially solve the two problems,this thesis takes reservoir gas bearing prediction as an example to implement a comprehensive research on interpretable ML-based reservoir prediction using big data and multiple ML algorithms.To improve the interpretability of the ML methods,the thesis first outlines a seismic reservoir prediction method based on the conventional geophysical theory to explain the feasibility of seismic reservoir prediction.Then fix the framework of ML reservoir prediction methods.And the ML reservoir gas bearing prediction problem is solved as a supervised classification task.We adopt the local waveform characteristics as the training or test samples.And the gas-bearing curves obtained from the logging interpretation provide the labels of the samples.The setting strategy of the samples and labels is conducive to increase the connection with the traditional seismic reservoir prediction technology and enhance the interpretability of ML-based reservoir gas bearing prediction methods.The k-nearest neighbor(kNN)method,fully connected neural network(ANN),and convolutional neural network(CNN)are selected to solve the ML reservoir gas bearing prediction,respectively.The three ML methods are different in the interpretability but have the interconnected principles among each other.Through the mutual explanation of the three ML methods,the effectiveness and reliability of the ML gas bearing prediction method are improved.A new data-driven seismic attribute,the kNN gas bearing attribute,is proposed based on the kNN method.This new attribute has the advantages of simple operation,specific physical meaning,and good correspondence with the distribution of reservoirs.It can be used as a favorable reference for gas bearing reservoir prediction.Based on the ANN method,the ANN-kNN gas bearing prediction method is proposed.First,ANN is used for gas bearing prediction.The prediction process of ANN is monitored by the kNN method,and the intermediate features of ANN are used as new samples for the kNN method.In the proposed method,multiple prediction results are generated for the same target,and the final result of gas bearing prediction is obtained from integrated learning and comprehensive analysis.To address the problem of lack of sufficient training samples for AI methods,especially deep learning(DL)methods,the sample enhancement methods for geophysical data are investigated.This thesis takes CNN gas bearing prediction method as the research object,and the sample augmentation method via the full utilization of geophysical expert knowledge is tested.Based on the full use of unlabeled data(i.e.,seismic data without well),another sample augmentation method based on feature space similarity is proposed.The application results on real data illustrates that the sample enhancement strategies above can improved the effectiveness of CNN gas bearing prediction method to a certain extent.Finally,through the comprehensive use of all ML gas bearing prediction methods in this thesis,an interpretable ML reservoir gas bearing prediction method based on big data is formed.The ML gas bearing prediction method can be used as a strong reference and effective complement to the traditional seismic reservoir gas bearing prediction methods.However,in the research framework of this thesis,the ML gas bearing prediction method can only predict the distribution of gas bearing reservoirs qualitatively.Thus,it is impossible to get rid of the negative influence of limited single-trace seismic information.The improvement of the resolution comes from the calibration to local waveform characteristics by logging data.Providing more complete seismic and logging information and expert knowledge is the key to solve the problem of reservoir gas bearing prediction.Therefore,it is important to study the feature engineering and geophysical knowledge constraints for seismic and logging data in the future.