Published September 27, 2025 | Version v1

Forecasting Induced Seismicity Rates in Geothermal Fields with Temporal Fusion Transformer

  • 1. ROR icon Lawrence Berkeley National Laboratory

Contributors

Contact person:

  • 1. ROR icon Lawrence Berkeley National Laboratory

Description

This repository contains all the code, data preprocessing scripts, and training pipelines used to produce the results presented in the manuscript “Forecasting Induced Seismicity Rates in Geothermal Fields with a Temporal Fusion Transformer.”

The goal of this work is to improve the forecasting of induced seismicity rates in geothermal systems using deep learning-based method, which integrates historical seismicity and fuild-injection operation data. 

🔍 Overview

This codebase enables:

  • Construction of temporal forecasting datasets

  • Training and evaluation of TFT-based models

📦 Components

🛠️ Data Preparation

All raw data preprocessing scripts are available in the preparation/ directory:

  • raw_data_processing_Geysers929.ipynb: Preprocessing pipeline for Geysers dataset

  • raw_data_processing_UtahForge2024.ipynb: Preprocessing pipeline for Utah FORGE dataset

📈 Model Training

TFT-based training and validation code:

  • train_geysers929.ipynb: Training on the Geysers Prati-9/Prati-29 field

  • train_utahforge2024.ipynb: Training on the Utah FORGE 16A/16B dataset

Training can also be launched via terminal scripts:

 
bash run_geysers929.sh bash run_utahforge2024.sh

📂 Requirements

Install dependencies using the provided requirements.txt:

 
pip install -r requirements.txt

You may also recreate the environment using conda or venv if desired.

Files

QuakeCastNet-master.zip

Files (113.1 MB)

Name Size
md5:689ffbc9e6de09b047ca7d67ff5876cc
113.1 MB Preview Download