MACHINE LEARNING INVESTIGATION OF INJECTION-SEISMICITY IN ROTOKAWA GEOTHERMAL FIELD

Abstract

Understanding the injection-seismicity relationship in geothermal reservoirs can provide insight into reservoir connectedness. One challenge is that, in real fields, fault and reservoir complexity make it difficult to apply simple analytical models to understand the data. Here, we use a machine learning technique called time-series feature engineering to study relationships between aspects of fluid injection and microearthquakes in Rotokawa geothermal field, New Zealand. We took four years of injection data between 2012 and 2016 and sliced it into smaller sub-windows. For each window, the average seismicity in a look-back period was computed, and then binary label of 1 was assigned if it exceeded a threshold. Automatic time series feature extraction from the raw and transformed injection data in each window was performed using Python package tsfresh. Significant features of the data were identified on the basis of distribution discrepancy between the two labels. The results show that the injection rate at some wells is a predictor of longterm (fortnightly) earthquake rates. At other wells, there is a poor correlation between injection rate and seismicity. We have been unable to find any link between rapid changes in injection rate and seismicity spikes, as suggested by some theoretical models.