Study On Productivity Prediction Of Fractured Horizontal Wells In Tight Gas Reservoirs Based On Time Series Method

At present,the world ’s energy is gradually shifting to new energy sources such as natural gas.Tight sandstone gas is the most reliable resource and the best choice for mining unconventional natural gas resources under China ’s current economic and technological conditions.The traditional gas reservoir engineering method is not only complicated in calculation but also has many model constraints.Numerical simulation has limitations such as long modeling time,inaccurate description of fracture parameters and single seepage mechanism.Therefore,in the development of tight sandstone gas reservoirs,it is necessary to explore more accurate and efficient productivity prediction methods.In this paper,the historical productivity data and production system of 127 fractured horizontal wells in tight gas reservoirs in S block are used to carry out the following research.Firstly,this paper studies the productivity data of horizontal wells in S block,formulates the dynamic classification standard of gas wells,and divides the horizontal wells in S block into three categories according to the dynamic classification standard.Secondly,the time series method is used to study the productivity prediction of fractured horizontal wells in tight gas reservoirs,and the ARIMA productivity prediction model is constructed based on the traditional time series method.However,because the traditional time series method is only related to the single factor of time,the influence of production system and other factors on production capacity cannot be considered.Then,four capacity prediction models of BP,SSA-BP,LSTM and PSO-LSTM based on machine learning time series method are established.However,the four models are all driven by pure data,and the adaptability and interpretability of the models are not good.Thirdly,in order to overcome the poor adaptability of pure data-driven,the decline curve and neural network are combined to construct the decline curve and data-driven neural network model.The actual production data of three horizontal wells are used to verify the different decline curves and data-driven neural network models,and the MFF decline curve and datadriven neural network model is optimized.Finally,the ideal capacity data and actual production data simulated by CMG are used to verify the six models constructed in this paper,and the prediction effects of the six models are sorted.The prediction results show that for the ideal productivity data and the actual production data of the three types of wells,the accuracy of the MFF decline curve and the data joint drive model is the highest,and the RMSE is 0.001,0.058,0.078 and 0.001,respectively.The neural network model driven by decreasing curve and data is more ideal than the pure data-driven model.