Artificial Intelligence Model for Predicting Formation Damage in Oil and Gas Wells

Abstract

An artificial neural network (ANN) was developed to predict skin, a formation damage parameter in oil and gas drilling, well completion and production operations. Four performance metrics: goodness of fit (R2), mean square error (MSE), root mean square error (RMSE), average absolute percentage relative error (AAPRE), was used to check the performance of the developed model. The results obtained indicate that the model had an overall MSE of 355.343, RMSE of 18.850, AAPRE of 4.090 and an R2 of 0.9978. All the predictions agreed with the measured result. The generalization capacity of the developed ANN model was assessed using 500 randomly generated datasets that were not part of the model training process. The results obtained indicate that the developed model predicted 97% of these new datasets with an MSE of 375.021, RMSE of 19.370, AAPRE of 6.090 and R2 of 0.9731, while Standing (1970) equation resulted in R2of −0.807, MSE of 9.34×1016, AAPRE of 3.10×106 and RMSE of 4.10×105. The relative importance analysis of the model input parameters showed that the flow rates (q), permeability (k), porosity (φ) and pressure drop (Δp) had a significant impact on the skin (S) values estimated from the downhole. Thus, the developed model if embedded in a downhole (sensing) tool that capture these basic or required reservoir parameters: pressure, flowrate, permeability, viscosity, and thickness, would eliminate the diagnostic approach of estimating skin factor in the petroleum industry.

Interpretation of Neural Network Models – New Insights of Estimating Static Bottom-Hole Pressures of Gas Wells in Nigerian Petroleum Provinces

“…In Table 7, the R 2 and R values show that the developed neural-based models predicted integral values were close to the actual values provided by Poettmann [14]. This statement is because the R 2 and R values for the various neural-based models were close to 1, which statistically means a good fit between the predicted and actual values [29,31]. Again, the good fits of the developed neural-based models are further observed in Fig.…”

“…The overall R-values for the 2-5-1 topology indicated that the developed network could predict Poettmann's integral values with 99.870% and 99.92% certainty with the maximum-minimum and clip normalization methods. According to Sircar et al [30] and Effiong et al [31], the mathematical representation of the neural network computations parameters (i.e., inputs, weights, biases, and output) in vector form and transfer or activation functions (i.e., tansig and purelin) are related, as expressed in Equation 5;…”

Machine Learning-Based Models for Basic Sediment & Water and Sand-Cut Prediction in Matured Niger Delta Fields

“…The limitation alluded to most neural network models in the literature is the availability of the essential details of the models for their reproducibility. One such detail is the weights and biases of the neural network [22]. This study's weights and biases of the BS&W network are depicted in Table 4.…”