An Evolutionary Sequential Transfer Optimization Algorithm for Well Placement Optimization Based on Task Characteristics

Abstract

Summary Evolutionary transfer optimization (ETO) algorithms with the ability to learn from past tasks have made breakthroughs in more and more fields. When the experience embedded in the past optimization tasks is properly utilized, the search performance will be greatly improved compared to starting from scratch. Autoencoding evolutionary search (AEES) is an efficient ETO paradigm proposed in recent years. The solutions of each task are configured as input and output of a single-layer denoising autoencoder (DAE), and the across-problem mapping is established by minimizing the reconstruction error, which makes it possible to explicitly transfer the solutions across heterogeneous problems. However, despite the success of AEES, the population of the optimization task contains little information about the characteristics of the task and it is highly stochastic, especially in the early stages of searching. This restricts the effectiveness of the mapping constructed via AEES. On the other hand, most tasks do not save all candidate solutions in the search, which greatly limits the possibilities of traditional AEES applications; for example, well placement optimization (WPO) problems, which are a common engineering optimization problem in the oil industry. To overcome such limitations, a sequential ETO algorithm for WPO problems based on task characteristics and an autoencoder is developed in this paper. It uses the implicit relationship between reservoir characteristics and optimal well locations to learn from past tasks, and a mapping is calculated to transfer knowledge across tasks. The proposed algorithm aims to speed up the search for the optimal well locations and reduce the required time for WPO. The learned mapping is established by configuring the characteristics of past and current tasks as input and output of a single-layer DAE. The derived mapping holds a closed-form transformation matrix across heterogeneous tasks, and the optimal solution of the past task can be easily transferred to a dominant solution of the current task by matrix calculation, thus it will not bring much computational burden in the evolutionary search while improving search performance. Furthermore, according to the specific task, the construction scheme of the matrix of characteristics can be flexibly extended to achieve effective search enhancement. The comprehensive empirical studies of WPO and statistical analysis are carried out to verify the effectiveness.

Intelligent modeling with physics-informed machine learning for petroleum engineering problems

“…Kaydani et al (2011) integrated particle swarm optimization and neural networks to construct an efficient intelligent model to achieve the prediction of minimum miscibility pressure in carbon dioxide (CO 2 ) injection. Zhang's group combined optimization algorithms and machine learning for the optimization of well location (Qi et al, 2022), real-time production (Wang et al, 2022), and injection (Xue et al, 2022). A coupled architecture (Owoyele et al, 2022) combining machine learning and genetic algorithms was employed to solve the fluid flow prediction problem.…”