Research On Automatic Oil Reservoir History Matching Method Based On Neural Network

Numerical reservoir simulation can help oilfield developers develop production plans,and the numerical simulation’s accuracy depends on the geological model’s accuracy as well as the fluid physical parameters.Data obtained through logging data,core analysis,etc.are not sufficient to build an accurate geological model.Therefore,the reservoir history matching process is very important to improve the accuracy of numerical simulation results by correcting the geological model and fluid properties parameters to match the already measured observations.However,the traditional automatic history matching method is difficult to balance the matching accuracy and time cost.Moreover,due to the increasing complexity of geological models,the history matching process will become more time-consuming.To solve this problem,this paper proposes a neural network-based automatic reservoir history matching method,which can accomplish the numerical simulation task and reduce the time cost of numerical simulation while satisfying the matching accuracy.In this paper,a sequential Gaussian stochastic modeling approach is used to generate a collection of stochastic initial reservoir models in a reservoir numerical simulation software,which is used as a training sample for the generative neural network.The successfully trained generative network can generate the corresponding reservoir parameter models(permeability,etc.)from low-dimensional latent variables,and then input the parameter models generated by the network into the reservoir numerical simulator to obtain the corresponding production data(water production rate,gas oil ratio,and bottomhole pressure,etc.).After that,the low-dimensional latent variables are updated by iterative data assimilation with the observed data through an automatic history matching method.Finally,the updated low-dimensional latent variables are returned to the generative network of the successfully trained neural network to update the corresponding reservoir parameter model,and the numerical simulations are performed and solved iteratively,and when the convergence conditions are satisfied,then the final matching results are output.The final results show that the WGAN-ES-MDA reservoir automatic history matching algorithm established in this paper based on Wasserstein generative adversarial neural network(WGAN)and ensemble smoothing multiple data assimilation algorithm(ES-MDA)can downscale the high-dimensional reservoir parameter model,and compared with directly updating the entire high-dimensional reservoir properties,this workflow uses 99% less time to update the set of permeability models once than the ES-MDA reservoir automatic history matching algorithm.This workflow takes 99.7% less time than the ES-MDA reservoir automatic matching algorithm,so it is more computationally efficient and less costly.Moreover,the total mean absolute percentage error(MAPE)of its best solution matching result is 3.40%,which is lower than that of the ES-MDA reservoir automatic history matching method with a best solution matching result MAPE of 4.84% and the particle swarm optimization algorithm(PSO)reservoir automatic history matching method with a best solution matching result MAPE of 7.64%,so its matching effect is better.The research in this paper has important practical significance for improving the matching and prediction of reservoir history data,which provides strong support and guarantee for reservoir management and development.Also,this study provides new ideas and methods for subsequent research in the field of reservoir history matching.