6. PHASE I: ARTIFICIAL NEURAL NETWORK METHODOLOGY -
6.2. Artificial Neural Network Application in Petroleum Engineering
Artificial neural networks have been used since the 1980s. Neural networks have regained popularity as an important alternative analytical tool within the natural and social sciences (Wallace, 2008). Nowadays, they have been integrated into most fields and increasingly being used in various applications and studies in business, biology, medicine, engineering, physics, forecasting, etc. The key in using neural net works in petroleum engineering, or in any other discipline for that matter, is to observe, recognize, and define problems in a way that will be addressable by neural nets (Mohaghagh, 1995).
They were found to be very helpful and effective for solving petroleum engineering problems previously found to be difficult and complex using conventional methods. Examples of petroleum engineering projects that have benefited from the help of neural networks include reservoir characterization, zone identification, well testing, and drilling optimization (Altmis, 1996).
White, et. al., (1998) used several artificial neural networks to design and develop zone identification in a complex reservoir. In this study, several
ANNs were successfully designed and developed for zone identification in heterogeneous formations from geological well logs. Reservoir characterization plays a critical role in appraising the economic success of reservoir management and development methods. Nearly all reservoirs show some degree of heterogeneity, which invariable impact the production. As a result, the production performance of a complex reservoir cannot be realistically predicted without accurate reservoir description. The difficulty stems from the fact that sufficient data to accurately predict the distribution of the formation attributes are not usually available. One of the key issues in the description and characterization of heterogeneous formations is the distribution of various zones and their properties.
AL-Bulushi (2008) developed a methodology based on artificial neural network (ANN) models to predict water saturation in simple and complex reservoir formations using wire-line well logs and core data. ANN models were developed for simple sandstone reservoirs, where conventional wire-line logs were taken as input parameters. Input data was introduced into the ANN design using the operating data, which included both training and production data. The model was successfully tested on the Haradh sandstone formation (in Oman) yielding a prediction of water saturation with a correlation factor of 0.91 and a root mean square error of 2.5%.
Arehart (1990) has used a neural network to determine the grade (state of wear) of a drill bit while it is drilling. The network was applied using three-layer neural network and back-propagation as the learning algorithm. The input parameters to the neural network were rate of penetration ROP, weight on bit WOB, torque T, revolutions per minutes RPM, and hydraulic horsepower per square inch HIS. However it is important to note that Arehart’s system was trained with laboratory data collected using bits of known grades drilling through known lithologies. Also, the network was tested on synthetic formations of various bed thickness constructed from the test data.
Altmis (1996) used neural networks to predict the drilling parameters.
She used a set of data generated by an advanced, full size drilling rig simulator.
The parameters used to train the neural network were RPM, time, bit type, WOB, rotary torque, ROP, formation abrasiveness, formation drillability, bit
bearing wear, tooth wear, and SPM. Some of Altmis’s data was obtained from fields in the United State, but only RPM, time, bit type, WOB, rotary torque, ROP, and SPM parameters were included for prediction. It is important to note that Altmis used only three bits in her study.
Yilmaz, et al (2001) used the neural network to solve the optimum bit selection problem. Their model was developed using back propagation neural net by training the model using real rock bit data for several wells in a carbonate field. The input parameters used to train the neural network were sonic log, gamma ray log, depth, location, and rock bit data. In this study Yilmaz, et al used fractal geostatistics along with the artificial neural network in order to solve the optimum bit selection problem
Dashevskiy, et al (2003) used the neural network as a real-time drilling optimization tool based on MWD measurements for predicting control in drilling. Their model was developed using neural networks with at least a single hidden layer, aiming only at obtaining the optimum drilling parameters (WOB and RPM) to produce the optimum rate-of-penetration while drilling. The input parameters are surface and down-hole data given output quantitative advice for the driller on best weight on bit (WOB) and rotary speed (RPM). The Model uses True Vertical Depth (TVD) together with surface WOB, surface RPM (all averaged on one-minute intervals) as inputs to the neural network models to predict the ROP and down-hole diagnostics.
Note that True Vertical Depth (TVD) was used as a reference to determine formation properties at the corresponding depth from offset well data. Also, mud properties, flow rate and Bottom Hole Assembly (BHA)/bit were kept constant through the entire testing to minimize the number of factors affecting the drilling process.
Dashevskiy, et al carried out further investigations to predict formation properties. Other neural network models were tested to evaluate the formation properties at the bit using WOB, RPM, ROP values and down-hole diagnostics as inputs. These attempts did not yield very good results.
Fonseca, et al (2006) used the neural network, based on the Auto-Regressive with Extra Input Signals Neural Network, or ARX model, to approach the ROP modeling problem. The architecture used is a Feed-Forward
network, trained with the Levenberg-Marquardt algorithm. The parameters used to train the neural network were TVD, RPM, WOB, and two past values of the ROP are given as inputs ROP (t-1), ROP (t-2). Note that in this study Fonseca, et al used a real oil offshore field data set, which consisted of information from seven wells drilled with an equal-diameter bit. The correlation coefficient achieved with the methodology ranged from R=0.888 to R=0.9887 for the testing sets.