← Back to Sessions

Interpretable Deep Learning Framework for Forecasting Induced Seismicity in Geothermal Fields

Description: 

Induced seismicity presents a critical challenge in geothermal reservoir management. The occurrence of large seismic events poses public safety concerns. Forecasting induced seismicity provides valuable information for operators as well as better insights of the mechanisms of induced seismicity. Current physics-based approaches require heavy computation and detailed knowledge of the subsurface structure for accurate modeling of induced seismicity. Statistics-based approaches often provide better accuracy, but nonlinear relationship between injection parameters and seismicity provide difficulties.

In this study, we adopt a data-driven deep learning framework to forecast induced seismicity rates in geothermal fields such as Utah FORGE and the Geysers fields. Building on a modified Temporal Fusion Transformer (TFT) architecture, this approach integrates all the available geological and operational parameters, including historical seismicity and operational metadata, to predict future rates of induced seismicity. The model$B!G(Bs built-in attention mechanism identifies key contributing factors, enabling interpretable insights into the induced seismicity, without requiring prior assumptions about the importance of input features. Our approach forecasts the spatiotemporal distribution of seismicity rates by partitioning the study region into grids and associating input data with corresponding grid locations. The model processes static covariates (e.g., well locations), time-dependent past observations (e.g., historical seismicity and injection rates), and known future inputs (e.g., planned injection schedules). Outputs include probabilistic predictions across multiple quantiles, providing a estimation of uncertainty under various operational scenarios. This interpretable framework not only demonstrates superior predictive performance but also enhances understanding of the mechanisms underlying induced seismicity. It serves as a robust tool for risk mitigation and improved reservoir management strategies in geothermal operations.

Session: From Physics to Forecasts: Advancements and Future Directions of Induced Seismicity Research - I

Type: Oral

Date: 4/15/2025

Presentation Time: 08:00 AM (local time)

Presenting Author: Nori

Student Presenter: No

Invited Presentation: Yes

Poster Number:

Authors