Seismic Reservoir Prediction Method Research Based On Generative Adversarial Neural Networks

The stochastic seismic inversion method based on multiple-point geostatistics is a key technical means for seismic prediction of reservoirs.In this method,multiple-point geostatistics can acquire geological pattern knowledge by learning from training images and,combined with the seismic inversion framework,can obtain reservoir prediction results with pattern features.However,multiple-point geostatistics have difficulty in reproducing complex geological patterns,resulting in simulation results that do not match geological understanding;moreover,the entire inversion process is slow,and the simulation results have strong uncertainty.This paper develops a 3D seismic inversion reservoir prediction method based on generative adversarial networks(GANs).This method uses GANs as the foundation to train and learn underground buried river reservoirs and obtain an underground buried river model generator.Based on generator and the Bayesian seismic inversion framework,combined with the adaptive Markov chain Monte Carlo(Mc MC)sampling method,a generator-based seismic inversion prediction program has been compiled to predict the buried river karst reservoir in the Tahe area under the constraints of well data,seismic data,and geological pattern knowledge.(1)A random generation algorithm based on karst processes for underground karst rivers was proposed,and a corresponding software program was developed using the Python programming language.Based on the formation mechanism and influencing factors of underground karst rivers,the scale parameters and probability parameters of karst river reservoirs in the study area were statistically analyzed to determine the degree of karst erosion in different directions.The software program,based on cellular automata,randomly generated 1000 underground karst river karst reservoir training images,which contained geological knowledge about the development,processes,and results of karst river caves.These images can serve as a source of geological knowledge for the learning of generative adversarial networks programs.(2)A multi-parameter generative adversarial networks evaluation system was established to select the optimal generative adversarial networks model generator for geological pattern features of underground karst river reservoirs.Seven categories of feature statistical parameters were proposed from the perspectives of similarity,diversity,and computational efficiency.The performance of four types of generative adversarial networks,including traditional generative adversarial networks,self-attention generative adversarial networks,progressive growing generative adversarial networks,and stylebased generative adversarial networks,were quantitatively compared.It was concluded that the style-based generative adversarial networks are the optimal model generator for fracture-type carbonate reservoirs.These seven categories of feature statistical parameters compared the similarity of mathematical distributions and structural features with the geological pattern features of training images.The diversity of generative adversarial networks-generated geological patterns was quantified based on the number of distinguishable categories,overcoming the limitations of conventional evaluation methods that did not consider the diversity of fracture reservoir geological patterns and multi-scale features of geological bodies.(3)A seismic reservoir prediction method based on model generator has been developed,along with a seismic reservoir prediction software with independent intellectual property rights.The method utilizes the model generator to provide prior geological knowledge(pattern features)and establishes a Bayesian seismic inversion framework.By incorporating well data and seismic data as observational data,it simulates the process of continuously revising prior geological understanding based on the observed data.By computing the solution of the posterior distribution(predictive model),it obtains reservoir prediction results that conform to both geological knowledge and observational data.In this process,Markov Chain Monte Carlo sampling method and parallel computing using multiple GPUs are employed to adaptively sample well points and observational data,ensuring a high conditional matching rate.Multiple prediction results are simulated,improving the efficiency of the inversion iteration and enabling the acquisition of multiple stable distribution geological models.The seismic reservoir prediction program,compiled in Python,is based on the hidden river karst model generator and is suitable for reservoir prediction work in carbonate reservoirs with hidden river karst backgrounds.It can revise geological models based on seismic data and gradually reduce the uncertainty of seismic reservoir prediction during the revision process,ultimately obtaining the reservoir prediction result with the highest probability.It is an intelligent seismic reservoir prediction software program driven by well-seismic data and constrained by geological prior knowledge.(4)In the research area of the single-layer underground river structure in the Tahe Oilfield,the seismic reservoir prediction program based on the model generator is applied to obtain the predicted reservoir model of the underground river karst.By inputting well data and seismic data,50,000 iterations are completed within 200 minutes,simulating 16low-uncertainty prediction models.The results show a high level of match between the prediction models and the seismic data,with a 100% conditional well matching rate.Compared to the multi-point geostatistical geological modeling method,the seismic reservoir prediction method based on the model generator exhibits better lateral continuity.Compared to the multi-point geostatistical seismic inversion method,the method based on the model generator produces prediction results with more stable and continuous geological patterns,higher conformity with seismic data,and significantly faster computation speed,resulting in lower model uncertainty.By utilizing the model generator trained with single-layer underground river training image knowledge,the seismic reservoir prediction method is applied to a research area with a dual-layer underground river structure in multiple wells.The results show that the prediction model exhibits a dual-layer underground river structure but lacks continuity,with a well matching rate of 90%.The prediction performance in this research area is inferior to that in the single-layer underground river research area,demonstrating that the geological knowledge contained in the training images is an important foundation for determining reservoir prediction effectiveness.