This paper provides the technical details of developing the automated stage-wise KPIs report generator, which is to be implemented as a module in the Real-Time Completion (RTC) analytics system. The generator is constructed with three models, two of which use Machine Learning (ML), that detect the stage start and end, identify the ball pumpdown and seat operation, and segments of a single stage of time series data into operationally similar sections. These tasks are performed based on the reliably available measurements of slurry rate and wellhead pressure, which enable the real-time automated stage-wise KPI analysis, and also lay the foundation for further advanced analysis regarding operational decision making.

The stage start and end detection is treated as a classification task. The slurry rate and wellhead pressure along with their first and second order derivatives are extracted by a fixed-length sliding window and structured as matrices, which resemble the data structure of ML inputs. A Convolutional Neural Network (CNN) is trained for the classification, and each data point is classified as either within a stage's pumping time or otherwise as the data is received, with minimal latency. Data of eight wells with 648 total stages were labeled for the stage start and end detection model. The five-fold cross-validation technique was used to evaluate the performance of the model, and a 15-second-window was used to extract data from the time series data. The model achieved an accuracy of ∼99.7% with a tolerance of 25 seconds in all blind tests, meaning the predicted start or end of the stage was fewer than 25 seconds before or after the actual flag.

After the start of stage classification is made, the ball pumpdown/seat recognition tasks take place. The ball pumpdown/seat event detection is a two-step strategy. The first step is to detect if there is a ball pumpdown/seat event in a stage, and the second step is to locate the end of that event. The first step is accomplished via a deep learning model with U-Net architecture, which detects the ball pumpdown and seat pattern within the slurry rate and pressure time series data. 179 samples are used to train the U-Net model. The transfer learning technology is used, as the dataset is small, and the U-Net is materialized with the pre-trained ResNet-34. The blind test's F1 score for the U-Net models is 0.97, which indicates excellent performance on the prediction. The second step can be achieved by a rule-based selection given the information from the first step.

The third model analyzes a single stage and splits the stage into differently categorized segments. The model takes a three-step strategy. First, the stage data is smoothed by the sequential application of three different filters. The smoothed data is used in the second step, which detects points of interest and categorizes the segments in between. Segments are categorized by the slope of a linear fit or the mean first order derivative along the segment. Finally, a rule-based method is applied to agglomerate segments, which leads to a more interpretable categorization.

The solution presented by this paper, provides real-time automated interpretations of hydraulic fracture events, enabling auto-generation of KPI reports, dispelling the need for manual labeling and eliminating human bias and errors. It fills the manual task gaps in the RTC workflow/data pipeline and paves the way for a fully automated RTC system.

You can access this article if you purchase or spend a download.