Prediction of well production from unconventional reservoirs is a complex problem even with considerable amounts of data especially due to uncertainties and incomplete understanding of physics. Data analytic techniques (DAT) with machine learning algorithms are an effective approach to enhance solution reliability for robust forward recovery forecasting from unconventional resources. However, there are still some difficulties in selecting and building the best DAT models, and in using them effectively for decision making. The objective of this study is to explore the application of DAT and Monte-Carlo simulation for forecasting and enhancing oil production of a horizontal well that has been hydraulically fractured in a tight reservoir. To do this, a database was first generated from 495 simulations of a tight oil reservoir, where the oil production in the first year depends on 16 variables, including reservoir characteristics and well design parameters. Afterward, using the random forest algorithm, the most influential parameters were determined. Considering the optimum hyperparameters for each algorithm, the best algorithm, which was identified through a comparative study, was then integrated with Monte-Carlo simulation to determine the quality of the production well. The results showed that oil production was mainly affected by well length, reservoir permeability, and number of fracture stages. The results also indicated that a neural network model with two hidden layers performed better than the other algorithms in predicting oil production (lower mean absolute error and standard deviation). Finally, the probabilistic analysis revealed that the completion design parameters were within the appropriate range.

即使有大量数据，非常规油藏的油井产量预测也是一个复杂的问题，特别是由于不确定性和对物理学的不完全理解。具有机器学习算法的数据分析技术 (DAT) 是增强解决方案可靠性的有效方法，可从非常规资源进行稳健的前向恢复预测。然而，选择和构建最佳 DAT 模型以及有效地利用它们进行决策仍然存在一些困难。本研究的目的是探索 DAT 和蒙特卡罗模拟在致密油藏水力压裂水平井石油产量预测和提高中的应用。为此，首先根据致密油储层的 495 次模拟生成数据库，其中第一年的石油产量取决于 16 个变量，包括储层特征和井设计参数。然后，使用随机森林算法确定最有影响力的参数。考虑每种算法的最佳超参数，通过比较研究确定最佳算法，然后将其与蒙特卡罗模拟相结合，以确定生产井的质量。结果表明，石油产量主要受井长、储层渗透率和裂缝级数影响。结果还表明，具有两个隐藏层的神经网络模型在预测石油产量方面比其他算法表现更好（平均绝对误差和标准偏差较低）。最后，概率分析表明完井设计参数在合适的范围内。