Construction Of Data Driven Shale Gas Productivity Forecast Model Based On Deep Neural Network

In recent years,the exploration and mining of shale gas has become a global hot spot.China is also accelerating the development of shale gas,which is not only conducive to China's clean energy transition,but also an inevitable choice for the natural gas industry.Before the development of shale gas reservoirs,it is of great guiding significance for shale gas development to make good production capacity predictions.Conventional production capacity prediction methods are inadequate for extracting data features,and the selection of regression curve parameters is highly subjective,while deep neural networks can fully learn the characteristics of the data,making the predicted production capacity highly credible.Therefore,the deep neural network can be applied to shale gas production capacity prediction as a supplement to the production capacity prediction method.In this paper,based on the large amount of data generated by the shale gas reservoir numerical simulation software,CNN and LSTM are used to construct shale gas productivity prediction models,and the main work is carried out as follows:First,select the appropriate factors affecting shale gas reservoir productivity as feature data,use the Latin hypercube sampling method to obtain multiple sets of feature data,and use shale gas reservoir numerical simulation software to generate production data,preprocess the data to construct The data set required by the module.Secondly,based on shale gas reservoir feature data,CNN and RF regression algorithms were used to establish an initial production prediction model.The results show that the error predicted by CNN is 7% lower than that of the RF model.The method of using output data to guide engineering design is designed.According to the highest production data predicted by the CNN model,the corresponding geological parameters are fixed,and the engineering parameters are randomly generated using the shale gas reservoir numerical simulation software,and the corresponding production data are generated to find the engineering parameter combination that maximizes the initial production.Thirdly,the initial production capacity is added to the characteristic data,and the dynamic prediction models of shale gas production and cumulative gas production are established by using CNN and RF regression algorithms.Comparing the two,it is found that CNN has better prediction effect.Finally,based on the historical production data of shale gas reservoirs,an LSTM model was established to predict later production.Compared with RNN,it was found that the prediction error of LSTM decreased significantly.Since the production data generated by the shale gas reservoir numerical simulation software is more obvious,and the actual production data is more random,the production data is revised.Add errors at specific nodes,and use linear interpolation at non-nodes to generate new output data,which is more in line with the distribution of actual production data.LSTM is used to model the revised data to predict later production.