Prediction of Well-Logging Curves Based on MPSO/D-Optimized Deep Neural Networks

This paper addresses the limitations of traditional logging curve prediction methods in complex reservoirs, particularly their inadequate generalization ability and the challenges associated with high-dimensional nonlinear modeling. We propose a deep neural network (DNN) logging curve prediction method, which is based on a multi-objective particle swarm optimization algorithm with target space decomposition (MPSO/D). This method effectively balances prediction accuracy and model complexity through target space decomposition, dynamic neighborhood search, and a constraint adaptive adjustment mechanism. This approach surmounts the issue of traditional parameter tuning, which often falls into local optima. The global search capability of MPSO/D is well-suited to the high-dimensional noise environment of logging data, thereby significantly enhancing the robustness of DNN in heterogeneous reservoirs. We applied this method to predict the resistivity curves in the logging data of wells B1, B2, and B3 in the A block of the Songliao Basin in the Daqing Oilfield. We compared our prediction results with those obtained from four other improved algorithms. The experimental findings indicate that the mean squared error values derived from MPSO/D-DNN are markedly lower than those produced by other models. Furthermore, the prediction curve exhibits the highest degree of congruence with actual values, thereby substantiating the efficacy and practicality of this method.