No abstract

During hydraulic fracturing jobs, engineers must monitor the wellhead pressure and adjust the pumping schedule in real time to avoid screenout, optimize the proppant and fluid amounts, and minimize cost. In this paper, we use machine learning to predict wellhead pressure in real time during hydraulic fracturing. The new algorithm can assist engineers in monitoring and optimizing the pumping schedule. We explored several neural network models. For each hydraulic fracturing stage, we train a machine learning (ML) model with the data from the first several minutes and predict the wellhead pressure for the next several minutes; we then add the data for the next several minutes, train a second ML model and predict the pressure for the next couple of minutes; and so on. We used several performance metrics to compare different models and select the best model for deployment to the Cloud, where a real-time completions platform is developed and hosted. We selected more than 100 hydraulic fracturing stages from several wells completed in the Delaware Basin and tested several ML methods on the historical data. The wellhead pressure can be predicted with an acceptable accuracy by a slightly nonlinear machine learning model. We tested the ML model on the Cloud, where real-time streaming data such as slurry rate and proppant concentration are gathered. The computation is fast enough that a real-time wellhead pressure can be predicted.

No abstract

No abstract

This paper provides the technical details to develop a real-time deep learning model to detect and estimate the duration of downlinking sequences of Rotary Steerable Systems (RSS) based on a single measurement (standpipe pressure, SPP). Further analytics are derived based on the downlink recognition results together with other real-time log data (ROP, RPM, Torque, etc.) to drive directional drilling efficiency. Real-time RSS downlink recognition is treated as an image segmentation problem. The Deep Learning (DL) models were created using the dynamic U-Net concept and materialized with a pre-trained ResNet-34 as the underlying architecture. Transfer learning was used due to the limited number of training samples (≪ 100 downlinks per onshore well) to help with speed and accuracy. The SPP time series data was segmented based on stand of pipe drilled (one image per stand). This "image" was then fed into the model for downlink recognition. To further increase the accuracy, a second opinion mechanism was applied when the models were tested and deployed into the Real-Time Drilling (RTD) system. Using a dual model approach greatly reduced the number of false positives due to non-downlink pressure fluctuations causing "noise". The patterns of SPP and its rate of change (delta SPP) are quite different. They both have pros and cons for identifying the downlink, thus two independent models were built based on these two signals. The DL model A is trained based on the original SPP signal and the DL model B is trained based on delta SPP. A downlink is confirmed only when both models show positive results. Data of 10 onshore wells (2 rigs) drilled with RSS were segmented (8165 images in total) and labeled. There were 671 images with 795 downlinks and 7980 images without downlink. The five-fold cross-validation technique was used to identify the best model(s). The F1 score of blind test result was .991 (accuracy was ~99.82%, see Table 2). The relative error of duration estimation is 2.49%. The current rig fleet within the RTD system has a mix of drilling tool configurations - RSS and mud motors. To further validate the models’ robustness regarding drilling tools, additional tests were conducted using mud motor wells’ datasets from 21 rigs (25431 images without downlink). There were 3 false negatives from this extended test set, which resulted in a ~99.93% accuracy for the aggregated 31 wells dataset. These results suggest that the models are accurate, reliable and robust. The real-time DL solution presented in this paper enables operators to analyze RSS performance during and between downlinking events. This would allow drilling engineers and rig supervisors to make faster, more reliable data-driven decisions to optimize performance and directional control of the well path.