Conference article

Accurate estimation of ?ow in drilling operations at in?ow and out?ow positions can lead to increased safety, optimized production and improved cost ef?ciency. In this paper, Dynamic Arti?cial Neural Network (DANN) is used to estimate the ?ow rate of non-Newtonian drilling ?uids in an open channel Venturi-rig that may be used for estimating out?ow. Flow in the Venturi-rig is estimated using ultrasonic level measurements. Simulation study looks into fully connected Recurrent Neural Network (RNN) with three different learning algorithms: Back Propagation Through Time (BPTT), Real-Time Recurrent Learning (RTRL) and Extended Kalman Filter (EKF). The simulation results show that BPTT and EKF algorithms converge very quickly as compared to RTRL. However, RTRL gives more accurate results, is less complex and computationally fastest among these three algorithms. Hence, in the experimental study RTRL is chosen as the learning algorithm for implementing Dynamic Arti?cial Neural Network (DANN). DANN with RTRL learning algorithm is compared with Support Vector Regression (SVR) and static ANN models to assess their performance in estimating ?ow rates. The comparisons show that the proposed DANN is the most accurate model among three models as it uses previous inputs and outputs for the estimation of current output.

drilling operations, open channel venture ?ume, non-Newtonian ?uid, ?ow rate estimation, ultrasonic level measurements, recurrent neural network, real-time recurrent learning