A New Artificial Neural Network Model to Predict Cutting Transport Efficiency in Deviated and Horizontal Oil Wells

Abstract

Inadequate hole cleaning impacts drilling operations. For example, poor wellbore cleaning can result in a number of drilling issues, including low drilling rate (i.e., low ROP), early bit wear, and in extreme situations, a well loss from a stuck pipe. Numerous studies have been carried out to comprehend how to reduce transport efficiency and offer potential remedies for the issue. They frequently provide empirical correlations that are based on data from experiments. Many engineering fields have recently made use of artificial intelligence and machine learning. Consequently, oil and gas companies started frequently using artificial neural networks (ANN) to forecast a number of crucial metrics. The purpose of this study is to use artificial intelligence approaches to forecast hole-cleaning efficiency. Two layers, TANSIG and LOGSIG transfer functions, and several training functions were used to construct feed-forward backpropagation ANN models. 1620 experimental data recordings from the investigations served as the basis for the investigation. Cutting density and pressure losses are included in the input parameters. Additionally, the model input consisted of drilling characteristics such as the drill pipe rotating speed (RPM), flow rate (GPM), pipe, and hole inclination angle. The best-performing model was chosen using sensitivity studies of 2, 4, 6, and 10 neurons for each transfer function (LOGSIG and TANSIG). With a correlation coefficient (R) greater than 0.9, the results showed that the constructed model correctly calculates the cutting transport efficiency (TE) in the wellbore. The findings demonstrated that as the number of neurons increases, the model's accuracy in terms of R for training and testing increases as well. The expected and real TE utilizing the TANSIG transfer function, GDM versus GD learning function, and four different training functions indicate that using the GDM adaption learning function generally outperforms the GD function. For instance, the correlation coefficient (R) for 10 neurons using the GDM function is 97.45 compared to 97.08 for the GD function. Results also indicate that the LOGSIG transfer function a bit overperforms the results estimated by the TANSIG function at two and four neurons. However, at higher numbers of neurons, the TANSIG function performs better. Therefore, we could suggest that using the TANSIG transfer function with the GDM for future prediction of TE is practically valid.