Research On Approximate Modeling And Dynamic Programming Of Alkali–surfactant–polymer Flooding

With the old oilfields entering into the late period of development,oilfields are faced with the problems of high water-cut and low oil recovery.Updating technologies to enhance oil recovery has been the significant measures to stabilize oil production.Alkali–Surfactant– Polymer(ASP)flooding is an important tertiary oil recovery technology which can significantly enhance oil recovery,and a large number of field tests have achieved good results.However,the development process of ASP flooding is time-consuming and slow effect,and the price of ASP is high.Besides that,the mathematical model of ASP flooding is complex.In order to achieve maximization of economic benefit,the scientific and effective development programs should be made.In this paper,we propose a new method,which is multi-scale wavelet kernel extreme learning machine to improve the accuracy of reservoir models by optimizing relative permeability data.Then based on dynamic recurrent wavelet neural network,the spatialtemporal decomposition model of ASP flooding is put forward.Finally,we adopt approximate dynamic programming to obtain the optimal control strategy of ASP flooding.Aiming at the weakness of the original fruit fly optimization algorithm(FOA),using existing research results,an improved fruit fly optimization algorithm(IFOA)has been proposed.Self-adapting adjustment of the iteration step value and the mutation operation which is guided by a defined evolution speed factor have been introduced to the proposed algorithm.So IFOA has stronger local searching ability,as well as the global converging character,which can significantly improve search precision and convergence speed.Through the testing of the six standard test functions and optimal control problem of CSTR,the excellent performance of IFOA is proved.Oil-water relative permeability data is an important essential data for the reservoir numerical simulation.In view of the shortcomings of high complexity and low accuracy in computing the relative permeability data,a new kernel extreme learning machine is proposed,which combines the multi-scale analysis property of wavelet function and kernel extreme learning machine,and the related parameters of the algorithm are optimized by the IFOA.By using the reservoir core data for the MATLAB simulation,the simulation results show the proposed method offers an extremely high accuracy and certain application value.ASP flooding is a kind of distributed parameter system with infinite dimension feature.When solving mathematical model of ASP flooding,we must solve the high order partial differential equations with high computational complexity and low efficiency.Because of the above problems and characteristics of ASP flooding,K-L decomposition is used to decompose the grid water saturation of reservoir into spatial basis functions and temporal coefficients.Due to the dynamic features of reservoirs,we propose a new dynamic recurrent wavelet neural network(DRWNN)which trained by the IFOA and gradient descent method to improve the ability of the dynamic modeling.The injection concentrations of ASP flooding and temporal coefficients from the K-L decomposition are used as input and output,then DRWNN can be used to identify the time domain model.By integrating the time domain model with the spatial basis functions,the dynamic model between grid water saturation and injection concentrations can be built.In addition,we also adopt DRWNN to identify the relationship between the grid water saturation and moisture content of the production wells.After combining the above two models,we can get spatial-temporal decomposition model of ASP flooding.By reservoir numerical simulation software CMG to get the identification data of ASP flooding model with four injection wells and nine production wells,the ASP flooding model can be get by using the above method.The testing results indicate the accuracy of the built model.For the dynamic optimization problem of ASP flooding,aiming to maximize the net present value,approximate dynamic programming is applied in this paper.Based on the built spatial-temporal decomposition model of ASP flooding,we use grid water saturation to construct basis functions and adopt Actor-Critic algorithm to approximate value function and control policy by using linear functions.The optimal control policy can be determined iteratively.The simulation result shows that the approximate dynamic programming can obviously improve the net present value of ASP flooding.