A Real-Time Fracturing Pump Pressure Prediction Method Based on Deep Learning

Abstract Real-time prediction of pump pressure during hydraulic fracturing operations enables dynamic adjustment of pumping schedules, prevention of screen-out, and optimization of proppant and fluid usage, thereby achieving safer and more efficient fracturing. During the fracturing design phase, physics-based mechanistic models fail to comprehensively describe the complex physical processes following fluid injection into formations, particularly due to challenges in acquiring critical parameters such as reservoir heterogeneity and near-wellbore tortuosity. Machine learning-based approaches offer distinct advantages over purely engineering models by simultaneously addressing spatiotemporal nonlinearity. In this study, feature engineering was employed to select key input parameters including surface pressure, discharge rate, and sand ratio. These were integrated with state-of-the-art temporal deep learning models to predict pump pressure. This combined approach not only allows the model to accurately account for intricate interactions among multiple critical factors but also effectively mitigates nonlinear relationships and stochastic noise inherent in field data. Experimental results demonstrate that the temporal deep learning models successfully predicted pump pressure variations during hydraulic fracturing, achieving a mean squared error (MSE) within 5% of measured values. This level of precision provides reliable indicators for downhole wellbore and reservoir conditions. The novel methodology described herein circumvents the limitations of current approaches that rely on monitoring complex physical processes, delivering a robust, accurate, and computationally efficient numerical solution throughout the pumping stage.