Predicting oilfield performance is extremely challenging because of the large number of variables that can influence and control it. Traditional methods such as decline-curve analysis have been commonly used but have been shown to have significant shortcomings. In recent years, advances in machine learning (ML) have provided a new suite of tools to tackle complex multivariant problems such as understanding oil-reservoir performance and predicating the final recovery factor. In this study, the application of a random-forest algorithm to train three predictive models and investigate the influence of the various input variables was investigated.

To train the algorithm, a database was built that includes information on 32 variables from 93 reservoirs from the Norwegian Continental Shelf. These variables control or potentially influence field performance and include factors that are a function of geology, subsurface conditions, fluids, and the engineering decisions taken in field development. In addition to these controlling parameters, data were also recorded for the fields that record performance. These included information on the estimated recovery factor and production rates.

Eighty percent of the data were input into the random-forest algorithm to train the models, whereas 20% were retained to blind test the subsequent models. Model accuracy was measured by comparing actual and predicted observations for each prediction metric using an R² score, mean square error, and root mean square error. The production-rate model had a mean square error of 0.004, whereas the mean square error for recovery factor was 0.024. Estimates of average monthly depletion rate have a mean square error of 0.0104. Predictor importance estimates indicate that geology/depth-dependent variables such as stratigraphic heterogeneity, reservoir depth of burial, average porosity, and diagenetic impact are among the variables with high importance in predicting recovery factor. When predicting reservoir-oil rate, the most important variables are related to field size, such as cumulative oil produced, number of wells, oil in place (OIP), and bulk rock volume. In this study, we provide data-driven insight into understanding the relationship between subsurface and engineering conditions of reservoir producibility; we also provide a tool for predicating reservoir performance within a basin or region.

预测油田性能非常具有挑战性，因为可能会影响和控制油田的变量很多。传统方法（例如下降曲线分析）已被普遍使用，但已显示出明显的缺点。近年来，机器学习（ML）的进步提供了一套新的工具来解决复杂的多变量问题，例如了解油藏性能和预测最终采收率。在这项研究中，应用随机森林算法训练三个预测模型并研究各种输入变量的影响。

为了训练该算法，建立了一个数据库，其中包含来自挪威大陆架93个水库的32个变量的信息。这些变量控制或潜在地影响油田的性能，并且包括与地质，地下条件，流体和油田开发中采取的工程决策有关的因素。除了这些控制参数外，还记录了记录性能的字段的数据。其中包括有关估计的采收率和生产率的信息。

80％的数据输入到随机森林算法中以训练模型，而保留20％的数据以盲目测试后续模型。通过使用R ²比较每个预测指标的实际和预测观测值来测量模型准确性得分，均方误差和均方根误差。生产率模型的均方误差为0.004，而回收率的均方误差为0.024。平均每月消耗率的估计值的均方误差为0.0104。预测重要性的估计表明，与地质/深度相关的变量，例如地层非均质性，埋藏的储层深度，平均孔隙度和成岩作用是预测恢复因子的重要变量。在预测油藏含油率时，最重要的变量与油田规模有关，例如累计产油量，井数，到位油量（OIP）和大块岩石体积。在这项研究中，我们提供了数据驱动的见解，以了解地下与储层可生产性的工程条件之间的关系；