Basin-Scale Machine Learning Prediction for Injection-Induced Seismicity in Oklahoma
Résumé
Induced seismicity is a major concern for underground fluid injection, including injection of wastewater and CO2 and H2 for carbon neutrality. Machine learning (ML) appears to be promising for analyzing injection-induced seismic events (IISE) without explicitly solving complex multiphysics; however, applying ML for analyzing IISE in space and time has never been attempted. In this study, we apply ML to analyze IISE in Oklahoma at the basin scale. We applied a deep learning (DL)-Multilayer perceptron (MLP) model to correlate injection fluid volume and pressure with IISE at 53 areas of central and northwestern Oklahoma in different years. In the model, we implicitly considered the properties of geological formations (e.g., permeability, Young's modulus, and strength) and fault locations by using a novel "neighboring" approach. In this approach, seismic events at each location correspond not only to the injection and seismic history of its own but also to those of the immediate neighbors. We use the MLP model to predict the injection-induced seismicity rate (µ) and the total seismicity rate (λ) defined in an empirical model. We trained the MLP model by using 12-month injection volume and pressure, and historical values of µ and λ of each location and its immediate neighbors, along with the coordinates. We then used the test data to assess the accuracy of the prediction. A Mean Square Error (MSE) of 0.003 was obtained for µ and 0.033 for λ. We conclude that the new MLP model is promising for basin-scale IISE in space and time.
Domaines
Planète et Univers [physics]Origine | Fichiers produits par l'(les) auteur(s) |
---|