Application of NARX Neural Networks in the Prediction of Hydraulic Fracturing Treating Pressure

Abstract

AbstractThe treating pressure of hydraulic fracturing is an important parameter to evaluate for the real-time optimization of pumping schedule. However, the physics-based evaluation methods such as the analytical and numerical fracture models have limitations due to their simplification and not capturing all the geology and operation behaviors. In this study, machine learning was applied to predict the real time treating pressure during hydraulic fracturing. Several neural network models were compared and the NARX neural network model was used. Over 100 hydraulic fracturing stages were selected from several wells completed in Jimusaer field. Models were trained with the data of the previous 32 fracturing stages in a well, where several performance metrics were used to compare different models and parameters were optimized. The data of the first several minutes in a new stage were then used to predict the remaining pressure in the new stage. Results indicate that treating pressure can be predicted with an acceptable accuracy by the NARX neural network model, whereas results were more acceptable when pumping rate and proppant concentration were both used as the inputs rather than pumping rate being the only one, implying that the proppant concentration is a critical factor affecting the treating pressure especially in the sand carrying period. Since the comprehensive prediction accuracy is over 90%, NARX neural network model can be an alternative way to predict the real time treating pressure and optimize the pumping schedule.KeywordsNARX neural networkHydraulic fracturingTreating pressure prediction