Summary Hydraulic fracturing is an effective method for enhancing both the initial reservoir production and ultimate recovery. Nevertheless, the conductivity of proppant fractures is a pivotal factor in the optimization of fracture designs within the context of fracture modification. Experimental testing methods for proppant fracture conductivity are costly and time-consuming, and the physical model is excessively complex and incomplete to account for all the influencing factors, resulting in low computational efficiency. A backpropagation neural network (BPNN) model was constructed using the RAdam optimization algorithm to identify a more efficacious method for predicting the conductivity of proppant fractures. The model was used to predict the fracture conductivity of two data types pertaining to the experimental data on the conductivity of geothermal and volcanic reservoirs. The prediction model is enhanced for three key areas. First, an isolated forest algorithm is used to assess and discard anomalous data points. Second, the objective function is optimized by employing the RAdam optimization algorithm, which has the advantages of both Adam and stochastic gradient descent (SGD). This guarantees rapid convergence and prevents the initial training phase from converging to a locally optimal solution. Moreover, this approach enhances the stability of the model training process. Finally, the rectified linear unit (ReLU) activation function may result in issues related to neuronal activity, including the potential for its disappearance. This study addresses this problem by employing the Kaiming initialization method. The experiments used a series of evaluation metrics, including the mean square error and coefficient of determination, to assess the predictive performance of the two data sets in the two distinct models. The experimental results indicate that the BPNN with RAdam optimization is a more effective approach for data pertaining to volcanic and geothermal reservoirs. Moreover, the prediction of the geothermal reservoir data is more precise than that of the volcanic reservoir data. This model can be used for rapid predictions based on existing fracture conductivity data, which can better guide the design of fracture modifications and is of great importance.
Read full abstract