Research On Optimization Methods For Reservoir Sandstone Thickness Prediction

Reservoir parameter prediction is the basis of reservoir evaluation,and sandstone prediction using seismic attributes is one of reservoir parameter prediction tasks.There are many different kinds of seismic attributes with different complexity,and the information they contained have different sensitivities to different prediction targets.In order to find the optimal seismic attribute combination,a feature selection algorithm in the computer field is introduced.For continuous feature variables and labels,consider using Gamma test as the evaluation standard for feature subsets.Gamma test can effectively measure the complex nonlinear relationship between inputs and outputs.It is a method based on neighbor sample analysis.In terms of search strategy,it is proposed to use the nested Monte Carlo tree search algorithm based on Gamma test(GNMCTS).Using the result of Gamma test as the reward function of the tree branch during simulation phases,and coupled the concept of Vratio with the tree depth to design a reasonable search stopping condition,which effectively overcomes exploration and exploitation dilemma.While maintaining accuracy,the improved version of the tree node selection strategy is used and the search process is accelerated by using multi-thread parallelism,which effectively improves computational efficiency.By analyzing the experimental results of UCI datasets and seismic datasets,and comparing with other recently proposed feature selection algorithms,it is verified that the proposed algorithm can select reasonable number of features from final UCI feature combination and seismic attribute combination and maitain important information in original feature space.After data preprocessing,the conditional generative adversarial network is used as regression model to realize the prediction of effective sandstone thickness.Conditional generative adversarial networks can effectively learn the conditional probability distribution of input and output without any assumptions about the type of distribution.Conditional adversarial generative networks can generate high fidelity samples and are widely used in computer vision,but there are few studies on regression problems.According to the characteristics of regression problem,it is proposed to apply the continuous conditional generative adversarial network in the image domain to the regression modeling of tabular data,and rewrite the objective function of the network.In order to improve the prediction accuracy,it is proposed to use improved-CGAN which improve the training objective function of the network by combining the singular vincinal loss based on the determinant point process with the regularization of the generator.So that the generated results can be compromised between diversity and accuracy.Through testing on synthetic datasets,real world datasets and seismic datasets,the results of visualizing the generative distribution of the network and the real distribution show that the improved conditional adversarial generative network can learn a variety of complex distributions,and can compete with the classical probabilistic regression model.The generator and discriminator network architecture and hyperparameters of the conditional adversarial generative network architecture can be done by automatic search optimization algorithm without manual trial-and-error adjustment.Due to the high drilling cost and the limited number of data set labels,the number of neurons in each layer of the network structure,activation function,dropout rate of each layer,batch size,initial learning rate and the learning rate schedule can be configured within various combinations.Using the mean absolute error of different search algorithms on a validation set of regression data,each candidate network is evaluated and the final network architecture is selected for sandstone thickness prediction of the reservoir.