The Method And Application Of Production Prediction For Strongly Heterogeneous Reservoirs Based On Artificial Intelligence

Strongly heterogeneous reservoirs with large reserve are widely found in major oil fields in the world.However,due to the complex geological conditions and extreme vertical and horizontal heterogeneity,the direction of water flow in this reservoir is not clear during the water flooding,and the production of wells varies greatly,which also leads to the fact that a complete and rapid production prediction method for strongly heterogeneous reservoirs has not been developed.Therefore,it is very necessary to quantitatively study the dynamic connectivity between injection and production wells and establish a production prediction method.In this study,artificial intelligence algorithms and numerical reservoir simulation methods are used to predict the production variation of production wells in strongly heterogeneous reservoirs from the perspective of data analysis.Firstly,we use multiple linear regression,least squares and constrained optimization by linear approximation methods to establish an inverse model of inter-well connectivity in strongly heterogeneous reservoirs considering the time lag and fading of injection water propagation,and use the conceptual numerical model data and actual field data to demonstrate the rationality of the model.Then,based on the inversion results of dynamic connectivity between injection and production wells,we develop single-well production prediction models(GCN-LSTM)and multi-well production prediction models(MGC-LSTM)for "multi-injection and one-production" and "multi-injection and multi-production" production scenarios using Graph Convolutional Network,Long Short-Term Memory network and Fully Connected Neural Networks,respectively,and design validation experiments to analyze the prediction performance of the models.Finally,based on the Latin hypercubic sampling and Monte Carlo methods,an uncertainty analysis method for production prediction is developed to analyze the uncertainty of the predicted production from the single-well production prediction model and the multi-well production prediction model,respectively.The study finds that:(1)the inversion results of the dynamic connectivity inversion model between injection and production wells match with the geological characteristics of the reservoir,which demonstrates the rationality of the method;(2)the single-well production prediction model(GCN-LSTM)using an artificial intelligence algorithm to capture both temporal and spatial correlations outperforms the production prediction models(LSTM,GCN)that only capture temporal or spatial correlations.(3)the model can reduce the error between the predicted production and the actual production by exploring the interactions between different wells from different perspectives and coupling multiple production influencing factors;(4)the uncertainty analysis method can give the predicted production under different probabilities and increase the reasonableness of the prediction results.