Research On Modeling And Approximate Dynamic Programming Of Alkali–surfactant–polymer Flooding For Enhanced Oil Recovery

At present,petroleum exploitation has entered the tertiary oil recovery stage with chemical displacement agent as the main means.Reservoirs at this stage often have the characteristics of high moisture content and low oil content,so how to further enhance oil recovery is an urgent problem to be solved in oil development.Alkali-Surfactant-Polymer(ASP)flooding is a tertiary oil recovery method which can effectively enhance oil recovery.However,ASP flooding has a long production cycle and high investment cost,so it is necessary to study ASP flooding technology scientifically in order to formulate a reasonable development plan.In this paper,a wavelet neural network(WNN)based on boundary value constraints(BVC)is proposed,then a data-driven identification model of ASP flooding is established.The model is solved by using approximate dynamic programming(ADP),the optimal injection strategy is finally obtained.For the typical distributed parameter system(DPS)of ASP flooding,this paper presents a spatial-temporal decomposition model of wavelet neural network based on boundary value constraints(BVC-WNN).Apply Karhunen-Loeve(KL)decomposition to decompose the water saturation of ASP flooding to obtain the spatial basis functions and the corresponding temporal coefficients.As for the terminal boundary constraints of moisture content,the topology of BVC-WNN is proposed,which can automatically satisfy boundary value constraints.The injection concentration of ASP flooding and temporal coefficients from the KL decomposition are used as input and output,then BVC-WNN can be used to identify the time domain model.Then,the spatial-temporal dimensionality reduction model of water saturation is reconstructed by combining the identified temporal coefficient and the dominant spatial basis function.In addition,the BVC-WNN is also used to identify spatial-temporal model between water saturation and the moisture content of production wells.Combining the two identification models,the spatial-temporal decomposition model between injection concentration and moisture content of production wells can be obtained.The simulation results show that the modeling accuracy of the spatial-temporal decomposition model is relatively high and can meet the requirements.As for the system of ASP flooding which has the continuous action space,an improved Actor-Critic algorithm is proposed,which has been applied to approximate the value function and control strategy of the ASP flooding respectively.Firstly,the net present value(NPV)maximization is taken as the value function,then the basis function construction method based on system characteristics is applied to construct the basis functions,and the adaptive basis function selection method is used to select suitable number of basis functions among the candidate basis functions;Furthermore,the action weighting method is adopted to restrict and approximate the optimal control action,in order to accelerate convergence,the eligibility trace is introduced.Finally,ASP flooding with four injection wells and nine production wells is used to test the effect of the proposed method.For the dynamic optimization problem of ASP flooding,this paper proposes a direct heuristic dynamic programming(DHDP)algorithm with initial weights optimization.Aiming at the problem that neural network is easy to be affected by random initial weights,this paper proposes an initial weights compositional correction method to optimize the initial weights of the network.Then the optimized weights are used as the initial weights of the action network and the critic network,and the gradient descent method is used to train the two networks.The simulation results show that this method can significantly improve the NPV of ASP flooding.