A Multi-LASSO model to forecast induced seismicity at enhanced geothermal systems
Metadata only
Date
2024-05Type
- Journal Article
ETH Bibliography
yes
Altmetrics
Abstract
Geothermal energy is an attractive option for the renewable energy industry, and especially enhanced geothermal systems (EGS), which can be implemented virtually anywhere on Earth. However, one main challenge this technology faces is seismic risk. Risk can be mitigated using an adaptive traffic light system (ATLS) that forecasts seismicity and updates the fluid injection plan accordingly. Here, a Least Absolute Shrinkage and Selection Operator (LASSO) model is tested with operational variables derived from injected volumetric rate V˙(t) and well-head pressure P(t) to forecast induced seismicity rates λ in time bins τi (here of 1 h). LASSO is chosen for its light computational cost, integration of the null hypothesis (H0) of a linear relationship λ(t)∝V˙(t), and automatic variable selection. Using datasets from nine deep fluid injection operations at three different sites in Australia, France and Switzerland, a LASSO with seven features is first demonstrated to perform better than the null hypothesis H0. To extend the time horizon of the model beyond τi, a multi-LASSO (called m-LASSO) is built from a sequence of sub-LASSO models 1≤j≤m, each predicting the rate at a next time bin τi+j−1. This model is shown to perform better than H0 (in 89–100% of cases for the first bin depending on the metric), before tending to H0 on time horizons where past data becomes uninformative. On average, m-LASSO performs better than H0 for a time horizon of 14 h (for m=24), after which the null hypothesis forecast should be preferred for its stability. To make m-LASSO a general ATLS forecasting tool, a maximum magnitude (Mmax) predictor is also provided as an add-on to the model. LASSO provides a practical compromise between more complex physics-based models (black boxes) and simpler statistical models (H0), by encoding the complexity of flow rate, pressure, and seismicity coupling, while keeping any uninformative feature to a zero-weight. m-LASSO is thus proper for a transparent ATLS, which is crucial for better-informed decision-making. Show more
Publication status
publishedExternal links
Journal / series
Geoenergy Science and EngineeringVolume
Pages / Article No.
Publisher
ElsevierSubject
Machine learning; Geothermal energy; Underground stimulation; Induced seismicity forecastMore
Show all metadata
ETH Bibliography
yes
Altmetrics