基于条件生成式对抗网络的油藏单井产量预测模型

为克服机器学习方法在油藏单井产量预测中的过拟合问题, 提高油田生产中的产量预测精度, 提出一种基于条件生成式对抗网络(CGAN)的油藏单井产量预测模型。该模型使用长短期记忆、全连接等基础神经网络, 构建生成和判别网络模型。生成网络模型以产量影响因素为条件输入, 生成预测产量数据, 利用对数损失函数评价预测数据与真实数据之间的偏差, 通过条件生成式对抗网络的博弈训练, 并结合贝叶斯超参数优化算法, 优化模型结构, 综合提高模型的泛化能力。基于Eclipse数值模拟软件建立同一井网条件下不同地质和生产条件下的油藏单井产量数据库, 以地质与生产条件等产量影响因素作为模型的条件输入, 进行油藏单井产量预测。结果表明: 与全连接神经网络(FCNN)、随机森林(RF)以及长短期记忆神经网络(LSTM)模型的预测结果相比, CGAN模型在测试集上的平均绝对百分比误差分别提升了2.59%、0.81%以及1.72%, 并且过拟合比最小(1.027)。说明CGAN降低了机器学习产量预测模型的过拟合程度, 提高了模型的泛化能力与预测精度, 验证了所提算法的优越性, 对指导油田高效开发和保障我国能源战略安全具有重要意义。

A Single Well Production Forecasting Model of Reservoir Based on Conditional Generative Adversarial Net

To address overfitting problem of production forecast model in machine learning and improve accuracy of production forecast in actual oil field, a model for single well production forecasting of reservoir based on Conditional Generative Adversarial Net(CGAN) was proposed. The model hybridizes two types of basic neural network structures, i.e., long short-term memory network and fully connected network, to construct a generative model and a discriminant model. The generative model takes production influencing factors as conditions to generate the forecasting production data. It defines a logarithmic loss function as a residual between predicted and real data, to improve comprehensively the generalization ability of the model. Bayesian hyperparameter optimization algorithm was used to optimize the model structure through game training of CGAN. With numerical simulation software Eclipse, single well production database with same well pattern under different geological and production conditions was established to train CGAN, which can be used to predict rapidly single well production of reservoir by taking geological and production factors as condition input of the model. Experimental results show that compared with prediction results of models FCNN, RF, and LSTM, mean absolute percentage error of CGAN model on the test set is increased by 2.59%, 0.81%, and 1.72%. The overfitting ratio is the smallest(1.027). It indicates that CGAN reduces overfitting degree of the machine-learning-based production forecast model, and improves generalization ability and accuracy of the model as well. This verifies superiority of the algorithm, which is of great significance to guide efficient development of oilfields and ensures security of national energy strategy.