Artificial neural network models for real-time prediction of the rheological properties of NaCl mud

Abstract

Drilling fluid’s rheology is a main parameter in optimization of drilling operations and reduction of the cost of drilling a well as it contributes to elimination of many drilling problems such as stuck pipe, hole cleaning, lost circulation, and other well control problems. Rheological properties of the drilling mud including plastic viscosity (PV), yield point (YP), flow behavior index (n), and apparent viscosity (AV) must be recorded in real time to calculate rig hydraulics. In rig sites, these values are recorded only twice a day because they require long time for measuring and cleaning the equipment. Artificial neural networks (ANN) can be used to get accurate models for estimation of the rheological properties of the mud in a real time based on the frequently measured parameters such as drilling mud density (MD), and Marsh funnel viscosity or time (FT). Application of ANN rewarded the objective of this paper to predict the rheological properties of the drilling mud by more frequently measured MD and FT. Using ANN has the advantage of retrieving the empirical correlations to calculate the rheological properties by whitening the black box which is another objective of this paper. Only field measurements of almost 814 datasets were divided into training and testing sets to validate the ANN empirical correlations. Levenberg-Marquardt backpropagation training algorithm (trainlm) and tangential sigmoidal (tansig) transferring function were used with all the empirical correlations in this paper. The optimized number of neurons for the ANN models was 38 for PV and 34 for n and YP, while it was only 24 for AV. The models were validated by almost 30% of the total datasets and resulted in very low average absolute percentage error (AAPE) which was 6.26%. The correlation coefficient (R) was more than 90% and comparing the AV model results with the previous AV models showed the superiority of the developed ANN-based empirical equation.