Transfer-learning Based Production Optimization For Reducing Reservoir Development Risk

With the development of oilfield digitization and intelligence,production optimization in closed-loop reservoir development theory aims to utilize the optimization techniques to formulate reservoir development plans to achieve optimal development state based on existing production conditions through the lifetime monitoring of oil and gas reservoir.Most of the oil reservoirs in our country are water flooding reservoirs,and the flow field can be optimized and adjusted to alleviate the contradiction between oil reservoir injection and production and improve development effects.It is noted that a specific class of oil reservoirs share the relatively similar geological conditions,heterogeneity,and injection-production contradictions.Therefore,how to accurately learn from the development rules of historical blocks and achieve more economical and effective development of oil reservoirs becomes an urgent problem remained to be solved.In this dissertation,based on the similarity of the reservoir task modalities,two types of algorithms termed homogenous and heterogeneous transfer methods were studied.Aiming at the homogenous problems,a self-adaptive multifactorial evolutionary algorithm based on the progressional optimization instance representation is proposed.The progressional representation of tasks is used to achieve more accurate characterization and description of tasks.Based on the representation models,the similarity between tasks can be measured online,and the transfer intensity can be adjusted in real time according to the similarity between tasks.By solving multiple different problems in a unified search space at the same time,the implicit exchange of genetic material between different tasks can be realized to achieve the solution transfer between different tasks and the overall performance of multitasking optimization can be improved greatly.This transfer method may fail when the problem modal difference becomes larger.With this in mind,an affine transformation enhanced multifactorial evolution is proposed for heterogeneous problems.Based on the progressional representation model of tasks,a closed-form solution of affine transformation for bridging the gap between two distinct problems is derived mathematically to enhance the transferability between heterogeneous problems.When the variable dimensions of some problems are very large,the evaluation may take lots of time,so that efficient calculation is not feasible under the existing computing resources.A multitasking production optimization based on endogenous and exogenous knowledge transfer is proposed to realize the knowledge transfer within tasks(intra-domain)and knowledge sharing between tasks(inter-domain)and thus improve the computational efficiency.In addition to using a single-objective multitasking benchmark suite to verify all the proposed methods,two sets of reservoir models are also employed to verify the effectiveness of multitasking reservoir production optimization.The results show that the knowledge transfer between tasks can improve the overall performance of the task.The mapping method for heterogeneous problems can improve the matching between source and target tasks,and reasonably increase the transfer intensity to realize the transfermation of solutions between distinct reservoirs.In addition,multitasking production optimization based on double-domain knowledge transfer can combine inter-task knowledge transfer and intra-task knowledge to reduce the number of candidate solution evaluations and shorten the time for candidate solution fitness evaluation so as to reduce the computational burden and improve the efficiency of the algorithm.