Reducing the computation time of well placement optimisation problems using self-adaptive metamodelling

In the proposed model management strategy, two surrogates are utilised. The first surrogate approximates the fitness function landscape, and the second one estimates the fidelity (accuracy) of the first surrogate over the search space. According to the estimated fidelity, the probability of using the EF is calculated for each individual, and then the algorithm stochastically decides to use the EF or AF. A heuristic fuzzy rule defines the range of probabilities in each evolution-cycle, based on the average fidelity of the second surrogate. The strategy was implemented on a genetic algorithm, with two neural networks, as the surrogates. The robustness of the proposed online-learning algorithm was analysed using a benchmarking analytical function and a semi-synthetic reservoir model, PUNQ-S3. The outcomes were compared with the results achieved by three algorithms, an unassisted algorithm, an offline-learning surrogate-assisted algorithm, and an online-learning surrogate-assisted algorithm with a random selection model management strategy. The comparison showed that the online-learning algorithm with the proposed strategy can outperform the other algorithms.