Abstract
The main goal of hydraulic fracture treatments is to achieve a target geometry and conductivity within given operational and budget constraints. Evaluating past frac designs can help one to understand completion performance by highlighting and summarizing relevant engineering considerations. For example, to simulate fracture growth, the model requires a pumping schedule as an input. This study describes an automated procedure that identifies the different steps of a pumping schedule and generates several pertinent statistics based on the hydraulic fracturing time-series data collected in the field. Essentially, it returns an actual or "as pumped" fracture treatment schedule, which may differ from the designed one.
The dataset analyzed in this study includes the slurry rates and proppant concentrations for 577 stages from all major North American basins. The algorithms were calibrated using 112 stages and tested on the remaining 465 stages. The procedure first identifies the start and end times of various sections of the pump schedule (pad, acid pad, slurry, sweeps, and flush) by smoothing, normalizing, and then "rounding" the proppant concentration signal. The procedure then isolates sustained intervals with positive proppant concentration to identify acid pads and proppant stages. The remaining time intervals are associated with pads, sweeps, and flushes. Various statistics such as volumes, durations, and averages are computed for each interval.
Each slurry interval (proppant ramp) is then further broken down into its proppant steps. This is accomplished using quantization ideas from digital signal processing. Quantizing the signal maps the observed proppant concentrations in each step (noisy signal) to a representative value for that step. If the stage is pumped as planned, then these representative values should be close to the designed concentrations. The authors have two techniques to accomplish this: the first is based on clustering, and the second uses a piecewise constant regression based on recursive partitioning (decision-tree regression). The representative values are used to identify the proppant steps, and once the steps are identified, the process generates statistics for each step.
Hydraulic fracturing time-series data is an ideal target for analysis using signal processing techniques. The process correctly recognizes the start and end times of the various pumping steps: 95% of the picks (target events) identified by the procedure are within three seconds of the manual picks. This automation significantly reduces the time required for picks, allows fast auditing of existing picks, and provides an efficient method for analyzing historic pumping schedule data that is available only as PDF files. This is the first paper that describes a computational method to automatically extract a pumping schedule from hydraulic fracturing time-series data. The method uses techniques from digital signal processing and is accurate, robust, transparent, and fast.
