ROP in horizontal shale wells : field measurements, model comparisons, and statistical learning predictions
MetadataShow full item record
Rate of Penetration (ROP) is one of the most important indicators of drilling efficiency available to drillers and engineers. Optimizing the ROP on a well allows the operator to decrease the amount of time spent drilling, which reduces cost. Further reductions in cost can come from utilizing and accurate performance model to understand whether a trip to the surface for a new bit is necessary, or if a bit trip would just increase Non Productive Time (NPT) without significantly benefitting performance. Clearly, understanding the factors that affect ROP is an essential part of drilling a successful well. Models for ROP have been developed over the academic history of Petroleum Engineering. One of the first models was the model developed by Bingham (1964), which offered a simple formula relating the RPM, Weight on Bit (WOB), and the diameter of the bit to a calculated value of ROP. Further work has continued in ROP modeling by Bourgoyne and Young (1974), who created a much more detailed ROP model including eight input parameters, Hareland and Rampersad (1994), who developed a drag‐bit specific model, and Motahhari et al. (2010), who developed a model specific to wells drilled with a positive displacement motor (PDM) and a polycrystalline diamond compact (PDC) bit. These models have a varying number of input parameters, and each rely on the tuning of between three and eight empirical coefficients in order to optimize them to the well which is being studied. This study applies these traditional ROP models to data collected while drilling modern horizontal shale wells. These wells were drilled with a rotary steerable system, as well as a downhole PDM, and PDC bit. The traditional models were first fit to the drilling data by using the full range of the horizontal section of the well to optimize the empirical coefficients. This method resulted in the traditional models acting largely like a moving average of the drilling performance over the horizontal region. Then, the empirical coefficients were optimized based on 50 ft sections of the horizontal region, which produced a much tighter fit between the calculated and actual ROPs. However this fitting methodology was found to be erroneous, since it was generating a forced overfit of the model to the actual data. Finally, the Wider Windows Statistical Learning Model was applied to the drilling data. This produced the best fits of any of the models which were considered, and was the only one of the models which followed the high‐frequency changes in the actual ROP data. As a result, this was the only one of the models which could be considered accurate for not only the estimation, but also the prediction and optimization of ROP in horizontal shale wells.