A Surrogate Model-Based Optimization Approach for Geothermal Well-Doublet Placement Using a Regularized LSTM-CNN Model and Grey Wolf Optimizer

The placement of a well doublet plays a significant role in geothermal resource sustainable production. The normal well placement optimization method of numerical simulation-based faces a higher computational load with the increasing precision demand. This study proposes a surrogate model-based optimization approach that searches the economically optimal injection well location using the Grey Wolf Optimizer (GWO). The surrogate models trained by the novel Multi-layer Regularized Long Short-Term Memory–Convolution Neural Network concatenation model (MR LSTM-CNN) will relieve the computation load and save the simulation time during the simulation–optimization process. The results showed that surrogate models in a homogenous reservoir and heterogenous reservoir can predict the pressure–temperature evolution time series with the accuracy of 99.80% and 94.03%. Additionally, the optimization result fitted the real economic cost distribution in both reservoir situations. Further comparison figured out that the regularization and convolution process help the Long Short-Term Memory neural network (LSTM) perform better overall than random forest. And GWO owned faster search speed and higher optimization quality than a widely used Genetic Algorithm (GA). The surrogate model-based approach shows the good performance of MR LSTM-CNN and the feasibility in the well placement optimization of GWO, which provides a reliable reference for future study and engineering practice.