Shale Oil And Gas Productivity Prediction And Fracturing Optimization Based On Deep Learning

The porosity and permeability of shale oil and gas reservoirs are extremely low,and fracturing is required to obtain benefits.However,the cost of fracturing is generally high,and it is particularly important to realize the efficient development of shale oil and gas reservoirs.Optimizing fracturing parameters with net present value objective function is an important means to reduce cost and increase efficiency.The calculation of net present value is inseparable from the productivity prediction.Using numerical simulation to predict productivity is time-consuming and inefficient,while neural networks have the characteristics of high efficiency and high precision,which can replace numerical simulation for productivity prediction.In this paper,different neural network models are established for different problems to predict productivity.For horizontal wells that will be fractured,the convolutional neural network is used to learn the relationship between the geological parameters,fracturing parameters,etc.(static data)of existing wells and initial production,production performance,so as to predict the productivity of the wells.For fracturing horizontal wells with a short production history,convolutional neural network and long short-term memory neural network are combined to establish a convolutional coding long short-term memory neural network model(CNN-LSTM),and a fusion layer is customized in the model to fuse static data with production data(dynamic data)to predict its future dynamic production sequence,and realizes production sequence-to-production sequence(Seq2Seq)prediction.Secondly,based on the production dynamic prediction of the convolutional neural network,this paper uses four algorithms of PSO,GA,DE and SA to optimize the fracturing parameters.Finally,the actual shale oil production data is used to verify the accuracy of the productivity prediction model and compare the improvement of the NPV after particle swarm optimization.The above methods are used for productivity prediction and fracturing parameter optimization.The results show that:(1)Convolutional neural network has excellent performance for initial production prediction.The prediction error of more than 60% of the samples is less than 2%,and the prediction error of more than 99% of the samples is less than 20%;(2)Convolutional neural network has excellent performance for production dynamic prediction,more than 58% of samples have prediction errors less than 3%,more than 96% of samples have prediction errors of less than 10%,and all sample errors are within 20%;(3)The CNN-LSTM model has better performance for the subsequent production prediction of horizontal wells with short production cycles.The prediction error of more than 93% of the samples is less than 3%,and the error of all samples is within 10%;5% and 10% are added to the data.% and 20% of three different levels of error,the results all show that after adding static data to CNN-LSTM,the prediction effect is higher than that of the LSTM model without static data;(4)Fracturing parameter optimization Among them,particle swarm performs the best,it takes less time,converges quickly,and has a high net present value;(5)The analysis of actual data shows that CNN and CNN-LSTM are reliable in prediction results,and after particle swarm optimization,the corresponding net present value increased by 18.6%.