Improving forecasts of injection-induced seismicity

Student thesis: Master's ThesisMaster of Science by Research (MScR)

Abstract

Hydraulic Fracturing Induced Seismicity (HFIS) becomes a public concern as it can occasionally cause damages and casualties. The Epidemic-Type Aftershock Sequence (ETAS) model, a popular statistical model for describing clustered seismicity including aftershock sequences, has been modified to forecast Hydraulic Fracturing Induced Seismicity (HFIS) by assuming a linear relationship between injection rate and the Poisson-distributed background rate (herein called the injection rate driven ETAS model). It showed some predictive skill in forecasting HFIS during the stimulations in 2018 to 2019 at Preston New Road (PNR), UK. However, the relationship between the induced seismicity and injection rate is complex and non-unique, and may not be well captured by the Poisson distribution, in which the rate parameter is linearly scaled with injection rate. This thesis seeks to find a better probabilistic description of the event count distribution with parameters coupled to injection rate. We analysed a catalogue with 38,383 events ranging from magnitudes MW -2.839 to 1.155 recorded during the stimulation at the PNR site, and found that the negative binomial distribution (NBD) describes induced seismicity during injection better than the geometric and Poisson distributions as it has better Akaike Information Criterion and χ2 goodness-of-fit test results, indicating substantially greater uncertainties between the seismicity and injection rate than commonly assumed in non-homogeneous Poisson process models. We then proposed the Injection rate driven Negative Binomial (INB) model, in which the injection induced seismicity is described by an injection rate driven NBD whereas the trailing events during an injection hiatus are modelled by the ETAS model. We found that both the in-sample and out-of-sample forecast version of the INB model performed better than their benchmark counterpart, which are the equivalence versions of injection rate driven ETAS model tested at the same site. We infer three reasons for the better performance by INB models: i) the simple linear scaling between injection rate and Poisson distributed background rate is inappropriate due to non-linear growth in mean and variance of the event counts during injection; ii) the NBD copes with the growing variations in the data better than the Poisson distribution, as the latter cannot cope with the overdispersed data; iii) a better description of the seismic response to injection leads to a more accurate ETAS parameters estimation. Hence, our new model can provide more robust forecasts of HFIS and thus more information for deciding operational plans and risk mitigation measures.
Date of Award10 Dec 2024
Original languageEnglish
Awarding Institution
  • University of Bristol
SupervisorJames M Wookey (Supervisor) & Max Werner (Supervisor)

Cite this

'