Predicting rate of penetration during drilling of deep geothermal well in Korea using artificial neural networks and real-time data collection

We present a feasibility assessment of predicting the rate of penetration (ROP) during wellbore drilling using data obtained as the drilling progresses within a given well. Drilling data from a 4.2 km-deep well at an enhanced geothermal system project are analyzed using an artificial neural network (ANN), with a total of nine input drilling parameters. We also assess the influence of the traditional input drilling parameters used for ROP prediction, and compare groups of these parameters to improve prediction. We then evaluate three different training data scenarios, and predict 5% of the coming section. The first data training scenario uses accumulative data, the second uses different amounts of data, and the third implements square root resampling. The results show that the total average values of mean percentage error for accumulated training, fixed sizes of 300, 200, 100 points, and data square root resampling are 21%, 24.5%, 29.2%, 20.5%, and 15.6%, respectively. In general, accumulative training captures the general ROP trend with depth, but its accuracy decreases in deeper sections. Data resampling returns the lowest mean percentage errors. We suggest that accumulative data and data resampling can be used during for ROP prediction to assist drilling management and decision making.