Application of Artificial Neural Network to Predict Downhole Conditions of Oil Wells Using Wellhead Data

Abstract An accurate prediction of downhole conditions using wellhead data is highly vital in petroleum production engineering applications. It is very useful in the management of the life cycle of a well. Although there are techniques in literature to predict downhole conditions using wellhead data; however, most of these techniques are empirical correlations derived from few datasets. The study presents a new robust and self-computing approach that was developed using Levenberg-Marquardt algorithm in MATLAB software. This model predicts the downhole conditions such as pressure drawdown (ΔP) and productivity index (PI) using wellhead data. The model was trained and validated using 3200 production data points from eleven (11) oil wells in the Niger Delta. The model predicted both the ΔP and PI with a training accuracy coefficient of determination (R^2 ) of 0.980 and the root mean square error (RMSE) of 13.560. This prediction closeness resulted in R^2 and RMSE values of 0.964 and 2.552 for ΔP and 0.964 and 2.552 for PI. The performance analysis of the newly developed ANN model against some of the input variables for ∆P were analyzed. It was observed that choke size less than 40 has a wide spread percentage error bandwidth, though majority of the predictions falls below the ±40 percentage error. For the effect of GOR on the performance of the newly developed model, the percentage error bandwidth is quite high, but majority of the predicted values are in the +20 and -40 range. The water cut data affect the performance of the model when it is less than 20. However, for water cut values greater than 20, the percentage error predictions are within the ±20 range. The results suggest that the ANN model presents a better choice for the prediction of downhole conditions with minimum error and acceptable accuracy.