A Genetic Particle Swarm Optimization with Policy Gradient for Hydraulic Fracturing Optimization

Summary Particle swarm optimization (PSO) has been widely developed and applied as one of the important algorithms for optimization problems because of its few control parameters and low complexity. However, conventional PSO has the problem of premature convergence. Most of the control parameters of optimized PSO are adjusted manually, and the technology of adaptive adjustment based on reinforcement learning (RL) is still immature. Therefore, this study proposes a new genetic PSO (GAPSO) with policy gradient (PGGAPSO). The algorithm constructs a strategy neural network as an agent to dynamically output the action probability distribution and passes the action with the highest selection probability to the GAPSO section through the designed action selection function, thereby updating the parameters. The GAPSO section solves the optimization problem based on the new parameters obtained and uses the individual and global optimums obtained from the iterations as inputs to the policy neural network. In addition, an elite strategy is introduced to enhance the exploration capability of the PGGAPSO. Then, the performance of PGGAPSO was comprehensively evaluated using convergence analysis, statistical tests, Friedman rank, and runtime analysis. Finally, PGGAPSO is applied to the construction parameter optimization of hydraulic fracturing. The results show that due to the influence of the number of samples, the optimization performance of PGGAPSO is general in the early stage of iteration but shows excellent with continuous iteration. For unimodal and most of multimodal functions, PGGAPSO shows significant difference and excellence. Additional processes, such as elite strategy and policy gradient (PG), resulted in a 7.3% increase in runtime for PGGAPSO compared with GAPSO, but this additional computational loss is acceptable relative to the improved accuracy. The algorithm does not require an objective function and can be combined with the complex agent model of hydraulic fracturing to optimize the fracturing parameters. In the case of parameter optimization for hydraulic fracturing in the Jimsar sag, the optimized production was increased by more than 10% on average.