Predicting temperature profiles in producing oil wells using artificial neural networks

A novel approach using artificial neural networks ANNs for predicting temperature profiles evaluated 27 wells in the Gulf of Mexico. Two artificial neural network models were developed that predict the temperature of the flowing fluid at any depth in flowing oil wells. Back propagation was used in training the networks. The networks were tested using measured temperature profiles from the 27 oil wells. Both neural network models successfully mapped the general temperatureprofile trends of naturally flowing oil wells. The highest accuracy was achieved with a mean absolute relative percentage error of 6.0 per cent. The accuracy of the proposed neural network models to predict the temperature profile is compared to that of existing correlations. Many correlations to predict temperature profiles of the wellbore fluid, for singlephase or multiphase flow, in producing oil wells have been developed using theoretical principles such as energy, mass and momentum balances coupled with regression analysis. The Neural Network 2 model exhibited significantly lower mean absolute relative percentage error than other correlations. Furthermore, in order to test the accuracy of the neural network models to that of Kirkpatricks correlation, a mathematical model was developed for Kirkpatricks flowing temperature gradient chart.