Dynamic production optimization based on transfer learning algorithms

Dynamic production optimization involves continuous cycle of model-predictive control initiated at specified times to maximize production net present value (NPV) throughout the expected life of the reservoir. Re-evaluating predictive models using traditional methods is computationally expensive. The existing methods for updating production controls, do not take the inter-cycle correlation into account. But such information is valuable for boosting the problem-solving efficiency on the current cycle by utilizing the experience or knowledge extracted from the previous developed cycles. Originally, transfer learning algorithms are difficult to be implemented into the oil and gas industry, due to its high computational cost and random or unavailable learning samples. Therefore, we propose a new transfer learning based optimization framework for dynamic production optimization problems. First, domain adaptation learning (DAL) is used to represent data between two inter-cycles, to decrease the dissimilarity between them. Second, extended boundary constraints (EBC) is a technique used to embed the optimization problem into the learning samples during DAL stage. EBC reduces the burden on computational facilities and makes the algorithm suitable for production optimization. Third, a transfer component analysis (TCA) method is used to simplify the data representation and also extract the data correlation. Then, the extracted correlation is used to produce an effective population for MOEAs. The developed framework is incorporated into three well-known evolutionary algorithms, nondominated sorting genetic algorithm II (NSGAII), multiobjective particle swarm optimization (MOPSO), and multiobjective evolutionary algorithm based on decomposition (MOEA/D) and one single objective optimizer, particle swarm optimization (PSO) for NPV maximization and robust optimization respectively. The proposed method is tested on a series of dynamic benchmark problems dynamic and a practical case based on a three channel reservoir model. Results showed that the proposed method reduces the number of simulation calls needed to reach optimum control options when using population-based evolutionary algorithms. Also, using the proposed technique, a higher NPV and better convergence speed in comparison to their original evolutionary algorithms is achieved.