According to the requirements of Action Plan of Shandong Province for Improving the Guiding Ability of Graduate Tutors, Action Plan of Shandong Province for Constructing High-quality Education and Teaching Resources for Graduate Students and Notice of Shandong Provincial Department of Education on the Application of Tutor and Education and Teaching Resources Project in 2022, after unit application, qualification examination and expert review and other steps, In Shandong Province, 505 outstanding master's degree theses have been passed. Chen Guodong, a graduate student in 2018, won the excellent master's dissertation of Shandong Province in 2022 for his thesis Optimization Method of Production Allocation and Injection Allocation of Agent Reservoir Numerical Simulator, which was supervised by Professor Zhang Kai in the research group. Congratulations!
The main innovation of this paper is to construct the agent model using machine learning method, improve the low accuracy of the agent model when approximating the high-dimensional, strongly nonlinear complex system, and improve the optimization efficiency of the agent-assisted optimization algorithm when solving the optimization problems of large-scale expensive production. It includes the following aspects:
(1) Aiming at the optimization problem of water drive reservoir injection-production based on numerical simulation, the agent assisted optimization algorithm is proposed to accelerate convergence and reduce optimization calculation time. Due to the time consuming of single evaluation of reservoir numerical simulation, it is necessary to build a proxy model to approximate the input/output relationship of numerical simulation, and combine dynamic iterative sampling with evolutionary algorithm to gradually converge to the optimal solution under a small number of function evaluation times.
(2) Aiming at the formulation optimization problem of high-dimensional water drive reservoir development scheme, a proxy-assisted optimization algorithm embedded with dimensionality reduction algorithm was proposed. The Gaussian process was used as the proxy model to embed the Sammon mapping dimensionality reduction algorithm to improve the accuracy of the proxy model in high-dimensional optimization problems, and the auxiliary evolutionary algorithm was used to select the optimal solution for evaluation, which could effectively alleviate the dimensionality disaster problem.
(3) The accurate construction method of global plus local agent model is proposed to solve the problem of inaccurate construction of agent model for strong nonlinear problems. The global proxy model is constructed to smooth out the local optimal region and improve the global exploration performance of the algorithm. The accurate construction of the proxy model is achieved by constructing the local proxy model in a small local region and sampling the optimal point to accelerate the convergence process.
(4) The hierarchical proxy framework sampling optimization algorithm was proposed, and the radial basis function was used as the proxy model to select the optimal subgeneration individual. The uncertainty of each solution was quantified by Euclianian distance, and the subgeneration individual with the highest uncertainty was selected for evaluation. Two local proxy models were constructed respectively by using the solutions closest to the current optimal solution and the solutions with the highest fitness value. The accelerated convergence process is evaluated by sampling the two local agent model optima.
(5) In order to verify the performance of the algorithm, the proposed accurate construction method of the agent model and the layered agent framework sampling optimization algorithm are applied to the test function and the injection and production optimization example of water drive reservoir. Hierarchical agent framework sampling optimization algorithm is used to optimize and solve 25 problems of the international 20-100 dimension CEC2005 standard test function set. Compared with the five most advanced algorithms proposed recently, the global optimal solutions of 14 test problems are refreshed, and the rapid and efficient solution of 20-100 dimension optimization problems is realized.
Chen Guodong was the first author of Information Sciences, Applied Soft Computing, SPE Journal and Journal of Petroleum Science and Technology during his master's degree I have published four sci papers in Engineering journals, and one EI paper in the Journal of China University of Petroleum (Natural Science Edition) as the second author. Won the second prize of the 15th China Graduate Mathematical Contest in Modeling in Huawei Cup during my master's degree; Second Prize of SPE Oil Pool Digital Oilfield Design Competition; Second Prize of the 7th Shandong Science and Technology Innovation Competition; The second prize of the 17th Shandong Challenge Cup; Shandong Province Internet Plus Silver Award. He is currently studying for his PhD at the University of Hong Kong.
