Hydraulic Fracturing Pressure Prediction with Deep Learning Approach

Abstract Maintaining proper treating pressure is critical in the execution of hydraulic fracturing. Inappropriate pressure management could lead to early bridging or premature screenout, which is detrimental to well completion efficiency and well productivity. In theory, treating pressure can be calculated; however, many factors make this task complex. In the field, pressure-based decisions predominantly rely on the experience of the engineer. In our work, a convolutional neural network (CNN) is used to predict treating pressure. In this way, human expertise and the deep learning method can synergize in the wellsite decision making and the fracturing job success rate can be improved. The workflow started with the comprehensive data preprocessing process in which the treating data from 631 stages in 271 wells in a tight gas sandstone field were collected and analyzed. Data were cleaned by an algorithm so that only meaningful main treatment data were preserved. The dataset contained wells with similar configuration and formations to remove the impact of outliers. To honor the physical process in the fracturing treatment, bottomhole treating pressure (BHP) and pressure derivatives were selected as additional features. A multilayer 1D CNN model was set up to train the dataset with the objective of predicting the treating pressure trend in a short time window, based on the input of a window of treating parameters. Surface pressure, pumping rate, proppant concentrations, and other hyperparameters were optimized based on the grid search method. A trained 1D CNN which split treating pressure with other inputs and honored historical treating pressure channel would be compared with a simple 1D CNN. K-fold validation methods were used to fully reflect the model's performance. It was possible to generate the pressure trend in a short time window in the future based on the input of historical treating pressure, pumping rate, proppant concentrations, and calculated BHP, etc. Both simple CNN and multihead CNN models could predict the treating pressure with reasonable accuracy. The deep learning approach can augment human decision making in managing treating pressure. This approach serves fast and reliable reference and helpful for staff training.