Dynamic Neural Network Model to Predict the Rate of Penetration Prior to Drilling
Operators and service companies are always interested to have a clear insight about the rate of penetration (ROP) since it will provide a good estimate for the time and the cost of the drilling operations. The aim of this work is to use recurrent neural networks (RNNs) to accurately predict ROP for any formation prior to drilling. A recurrent neural network was created to estimate the ROP based on some key drilling parameters and unconfined compressive strength (UCS). The data were divided into three sets; training, validation, and testing datasets. 70% of the data used for training, 15% for validation, and the rest for testing. The optimum number of hidden layers and the number of neurons in the hidden layer were obtained using trial and error by calculating the mean square of error (MSE) for each trail and then the lowest MSE was chosen. The final results showed that the supervised RNN has the ability to predict ROP prior to drilling for any formation assuming the key drilling parameters that are desired to be used to drill the formation were known as well as UCS of the formation. The network predicted ROP with an overall R2 of 0.94. This estimation is tangibly valuable for providing a robust image in regard the cost and time of the drilling operations.
H. H. Alkinani et al., "Dynamic Neural Network Model to Predict the Rate of Penetration Prior to Drilling," Proceedings of the 53rd U.S. Rock Mechanics/Geomechanics Symposium (2019, Brooklyn, NY), American Rock Mechanics Association (ARMA), Jun 2019.
53rd U.S. Rock Mechanics/Geomechanics Symposium (2019: Jun. 23-26, Brooklyn, NY)
Geosciences and Geological and Petroleum Engineering
Keywords and Phrases
Compressive strength; Cost benefit analysis; Cost estimating; Forecasting; Infill drilling; Rock mechanics, Drilling operation; Drilling parameters; Dynamic neural networks; Rate of penetration; Recurrent neural network (RNNs); Service companies; Trial and error; Unconfined compressive strength, Recurrent neural networks
Article - Conference proceedings
© 2019 American Rock Mechanics Association (ARMA), All rights reserved.
01 Jun 2019