Abstract: The minimum miscibility pressure (MMP) is a critical parameter that determines whether a reservoir can be explored by miscible flooding. To accurately predict the MMP, the multi-layer perceptron (MLP) prediction model was optimized using an improved mayfly algorithm (IMA). The attention mechanism was used to extract the factors affecting MMP; the optimization capability of IMA was enhanced by incorporating chaotic Sobol sequences, nonlinear inertia weights, and reverse learning methods. These improvements can provide optimal weights and thresholds for the MLP, leading to the establishment of the IMA-AmMLP model for MMP prediction. The model was validated by the case study of a block in Jilin oilfield. The results demonstrate that the IMA-AmMLP model exhibit a higher degree of fitting between the predicted and actual values, with a mean absolute error (MAE) of 1.036 MPa, a mean absolute percentage error (MAPE) of 0.024, and a root mean square error (RMSE) of 0.835, and the values were all superior to those of the original model. This indicates that the IMA-AmMLP model can more accurately predict MMP, providing a valuable reference for the exploitation and management of reservoirs using CO₂ flooding in fields.

Key words: minimum miscibility pressure, multilayer perceptron, inertia weight, mayfly optimization algorithm, attention mechanism

摘要： 最小混相压力是衡量油藏能否达到混相驱的标准。为了对最小混相压力进行精准预测,运用改进蜉蝣算法(IMA)优化多层感知机(MLP)的预测模型。运用注意力机制实现对最小混相压力影响因素的提取;通过引入混沌Sobol序列、非线性惯性权重和反向学习的方法增强蜉蝣算法寻优能力,为多层感知机提供最优的权值和阈值,进而构建IMA-AmMLP最小混相压力预测模型;并以吉林油田实际区块为例,对使用效果进行了验证。验证结果表明,IMA-AmMLP模型的预测结果与实际值的拟合度更高,其平均绝对误差为1.036 MPa,平均绝对百分误差为0.024,均方根误差为0.835,均优于原始模型。研究结果表明,IMA-AmMLP模型能够更准确地预测最小混相压力,可以为运用CO₂驱开采油藏提供参考。

关键词: 最小混相压力, 多层感知机, 惯性权重, 蜉蝣优化算法, 注意力机制