作者：高级编辑 Stephen Whitfield

非常规油气藏的水力压裂技术已发展成为一项工业化流程，每口井的压裂段数也比以往任何时候都多。为了提高作业效率，从而实现压裂规模化，该行业正越来越多地转向自动化。具体而言，该行业正在研究如何利用自动化来优化压裂速率和压裂段数，从而简化压裂设计流程。

Liberty Energy公司高级数据科学工程师Karn Agarwal表示，这种自动化技术可以帮助运营商更快地将油井投入生产，并提高每年的完井数量。然而，要将这一愿景变为现实，行业需要可靠的模型来预测和预估未来足够长的时间的处理压力。

在2026年2月5日于德克萨斯州伍德兰兹举行的SPE水力压裂技术大会上，Agarwal先生探讨了用于预测地面压裂压力的深度学习模型的开发和训练。该模型使用多种输入数据，包括以往压裂作业的时间序列数据和未来压裂作业的设计方案。它还可以整合阶段级完井细节，以及来自测井和钻井数据的岩石物性数据。

“我们希望将多个输入结合起来，并在这种神经网络架构中生成一个能够利用海量压裂处理数据进行训练的模型，”阿加瓦尔先生说。“这类似于大型语言模型，其本质是接收一系列单词或标记，并尝试预测序列中的下一个单词或标记。从这个意义上讲，压裂作业期间的地表压力响应，或者岩石的反应，就是这套设备的‘语言’，而我们预测的序列是一个时间序列。”

该模型完成后的最终目标是利用它开发一种可靠的实时压力预测工具，该工具可以捕捉到人类治疗主管可能不容易识别的细微差别和事件。

阿加瓦尔先生概述了该模型的几个潜在应用。首先，它可以用于提前检测处理压力上升或筛漏。这可以通过分类（根据一段处理数据和未来的设计预测是否会发生筛漏）或回归分析来实现。该模型还可以部署在实时应用中，根据需要进行多次预测，或者作为自动化工作流程的基础模型，用于优化泵速调度。

使用 1500 个阶段的数据进行训练

该模型于 2025 年使用两个未命名非常规油气盆地完成的 1500 个阶段的数据集进行训练和测试。在测试深度学习模型之前，研究人员运行了一个独立的传统机器学习模型，作为模型最低性能的基准。该模型使用与深度学习模型相同的数据，但没有采用能够自动从数据中学习模式的多层人工神经网络。

压力估计的训练均方根误差（RMSE），即各阶段预测值与实际值的偏差，为 390 psi。深度学习模型的 RMSE 为 280 psi，表明与基准模型相比，深度学习模型的预测不确定性更低。

这种相对较低的不确定性在测试过程中有所体现。例如，在某个测试阶段，浆料流速似乎是影响压力响应的主要因素。模型预测，在流速上升阶段，流速会突然下降，压力会上升；此外，模型还预测，在冲洗阶段（表面支撑剂浓度降至零时），由于静水压力的降低，表面处理压力会上升。这些仅通过分析测试数据集得出的预测结果与研究团队的预期相符。

“通过这项测试可以看出，该模型正在学习预测增压阶段处理压力的上升，而且预测得相当准确，”他说。“它还学会了预测冲洗阶段的压力上升，所有这些都是以数据驱动的方式完成的。我们没有告诉模型任何事情。看到模型能够预测这些影响，这非常令人鼓舞，”阿加瓦尔先生说。

该模型在测试中表现不佳的地方在于启动和停产之间的这段时期。阿加瓦尔先生表示，由于该时期浆液流量基本保持恒定，模型难以将预测结果控制在均方根误差（RMSE）以内。正因如此，模型难以从地层效应引起的意外处理压力变化中学习——他指出，这类变化在该时期并不常见。

“有时候，模型在这个时期会失去稳定性。如果由于井下效应导致压力发生任何突变——比如，压裂或井群堵塞，或者新的井群打开——模型就很难预测，因为它基于有限的输入范围，”他说。

测试中使用的数据集不包含任何被筛选掉的阶段，但阿加瓦尔先生指出，有些阶段的处理压力接近工作压力极限。

在一个示例阶段中，压力在整个阶段持续上升，并在阶段末期接近工作压力极限。在这种情况下，模型预测的趋势是正确的，即一旦达到设计速率，压力就会持续上升。然而，尽管模型能够显示压力持续上升的趋势，但却无法预测压力上升的幅度——阿加瓦尔先生表示，这可能是因为模型低估了该阶段所需的减摩剂浓度。未来对该模型的研究将探索如何提高模型在类似情况下的性能。

该模型未来的其他工作将侧重于整合更大的数据集和日志数据，以及确保对原始数据进行更好的质量控制。

更多信息，请参阅 SPE 230658，“利用深度学习实时预测压裂处理压力”。

By Stephen Whitfield, Senior Editor

Hydraulic fracturing in unconventional plays has scaled into an industrial process, with the number of stages being completed in each well higher than ever before. To boost the operational efficiencies that make scaling frac possible, the industry is now turning more and more to automation. Specifically, the industry is looking at how automation can be leveraged to optimize rate schedules and pumping stages, thereby streamlining the frac design process.

Karn Agarwal, Senior Staff Engineer – Data Science at Liberty Energy, said this type of automation can allow operators to bring wells into production faster and complete more wells per year. However, to turn this into reality, the industry needs reliable models that can forecast and predict treating pressures far enough into the future.

At the 2026 SPE Hydraulic Fracturing Technology Conference in The Woodlands, Texas, on 5 February, Mr Agarwal discussed the development and training of a deep learning model designed to forecast surface treating pressures. The model uses a variety of inputs, including time series data for past treatments and future treatment design. It can incorporate stage-level completion details, as well as rock property data from logs and drilling data.

“We want to combine multiple inputs and generate something that can be trained on very large amounts of frac treatment data in this neural network architecture,” Mr Agarwal said. “It’s analogous to large language models, which, in essence, take a sequence of words or tokens and try to predict the next word or token in the sequence. In that sense, surface pressure responses, or the reactions of the rock during the frac job, that’s the language of this equipment, and the sequence we’re predicting is a time sequence.”

The final goal with the model, when completed, is to use it to develop a reliable real-time pressure forecasting tool that could capture nuances and events that a human treatment supervisor may not readily recognize.

Mr Agarwal outlined several potential applications of the model. For one, it could be used to detect rising treating pressures or screenouts in advance. This could be done as either a classification (predicting if a screenout will occur given a window of treating data and a future design) or as a regression. The model could also be deployed in real-time applications, where it could make a prediction as often as needed, or it could serve as a foundational model for automation workflows optimizing pump rate schedules.

Training on 1,500 stages of data

The model was trained and tested in 2025 on a data set of 1,500 stages completed in two unnamed unconventional basins. Prior to testing the deep learning model, the researchers ran a separate, traditional machine learning model that served as the benchmark for minimum model performance. This model used the same data as the deep learning model, but it did not feature the multilayered artificial neural networks that automatically learn patterns from the data.

The training root mean squared error (RMSE) for the pressure estimations, or the deviation of the predicted values from the actual values in each stage, was 390 psi. For the deep learning model, the RMSE was 280 psi, indicating that the deep learning model would have a lower relative uncertainty in its predictions compared with the benchmark.

This lower relative uncertainty manifested at points during the testing. For example, in one test stage, the slurry rate appeared to be the dominant feature affecting the pressure response. The model predicted a sudden rate drop and a rise in pressure during the rate ramp-up period, and it determined a rise in surface treating pressures during a flush (when the surface proppant concentration drops to zero) due to a decrease in hydrostatic pressure. These predictions, which were made solely through analyzing the test data set, lined up with the expectations of the research team.

“You could see with this test that the model was learning to predict the rise in treating pressures during the ramp-up period, and it predicted it fairly well,” he said. “It also learned to predict the rise in pressure during the flush, and it was all done in a data-driven manner. We weren’t telling the model anything. That’s very encouraging to see the model predict these effects,” Mr Agarwal said.

Where the model struggled in testing was in the period between the ramp-up and the shut-in. Mr Agarwal said the model had trouble keeping its predictions within the RMSE because the slurry rate is mostly constant during this period. Because of that, the model had difficulty learning from unforeseen treating pressure changes caused by effects in the formation – these types of changes, he said, do not occur as frequently in this period.

“Sometimes the model lost its footing in this period. If we had any abrupt pressure changes due to a downhole effect – say, a frac or a cluster gets blocked, or a new cluster opens – the model will have a hard time predicting because it’s working off of a limited range of inputs,” he said.

The data set used in the test did not contain any stages that screened out, but Mr Agarwal noted that there were some stages where the treating pressures were close to working pressure limits.

In one example stage, pressures rose throughout the stage and got close to working pressure limits toward the end of the stage. In that instance, the model predicted the correct trend by consistently forecasting a rise in pressure once the design rate was achieved. However, while it could show the trend of continuously rising pressure increases, it could not predict the size of the pressure increase – Mr Agarwal said this was likely because the model underestimated the friction reducer concentrations needed in this stage. Future work on the model will explore how the model’s performance can be improved in this type of situation.

Other future work on the model will focus on incorporating larger data sets and log data, as well as ensuring better quality control of raw data. DC

For more information, please see SPE 230658, “Using Deep Learning to Forecast Fracture Treating Pressures in Real Time.”