Real-time analysis and forecasting of the microseismic cloud size: Physics-based models versus machine learning
The spatiotemporal distribution of hydraulic fracturing microseismicity is complicated and depends on various mechanical and diffusional parameters. Hydraulic fracture modeling can aid in understanding fracture propagation and microseismicity. Nevertheless, the complex spatial and temporal interaction of several processes occurring within and around the fracture represents a challenge for developing real-time tools for microseismic prediction. Two approaches were developed to forecast the microseismic cloud size in real-time. The first approach uses fracture propagation models to derive the cloud size directly from the microseismic observations. The second approach is based on a convolutional neural network (CNN) trained with the engineering parameters and past microseismic cloud size values. A rolling-forecasting strategy is employed to train consecutive CNN models in real-time to make predictions at a specified time lag. A data augmentation technique known as double noise injection is used to ensure that the amount of training examples available to the machine learning models at each time step is similar or larger than the number of free parameters. Results show that the CNN outperforms the quality of predictions of the physics-based models but with a reduced prediction capability. The physics-based approach can predict growth at any time but ignores the engineering parameters. In addition, the physics-based methods lead to real-time insights into the fracturing regime, revealing whether microseismicity is most likely generated due to a leak-off-dominated or a storage-dominated regime. The CNN model can forecast the cloud size only at a single future time lag while using the engineering parameters and past cloud growth as input. However, this approach does not provide a physical interpretation of the fracture propagation regime. The prediction accuracy of both methodologies varies depending on the microseismic behavior. We postulate that the CNN forecasts could be improved by including more physical constraints into the predictive model.