Geothermal Reservoir Evaluation Based On Integrated Well Testing and Resistivity Data Using Artificial Intelligence

Abstract

In tectonic regions, the flow of fluid within the subsurface is primarily influenced by subsurface temperature, pressure, porosity and permeability. It is therefore necessary to characterize these properties which help in estimating the ultimate productivity of a geothermal reservoir. These properties are used in volumetric calculation of fluids in the reservoir, calculation of fluid saturations and clustering of the reservoir in terms of aquifers, water confining stratum, hydro- thermal zones, lithological horizons, faults and fracture zones. Some of the methods commonly used for estimating these reservoir properties are time consuming and costly. A decisive method for the reservoir estimation is therefore desirable. Geophysical methods including seismology, gravity, magnetics and resistivity have been put in use for geothermal resource mapping at the Olkaria geothermal field for decades. By applying all the necessary geophysical study techniques and data integrated during interpretation, deeper wells producing up to 30 MWe have been drilled. However, despite all the advancements, geophysical integration of multiple datasets is mostly achieved through manual visualization. Machine vision of Artificial Intelligence is therefore desired. Reservoir temperature distribution and the electrical conductivity of rocks mainly depend on permeability, porosity and fluid chemistry. Machine Learning is needed to establish correlation between the temperature distribution and the electrical conductivity of rocks. This research focused on the integration of Olkaria Domes geothermal well testing and geophysical electromagnetic resistivity data. The aim was to establish an alternative estimation method for reservoir temperature through Machine Learning and application of machine vision perceptions and better visualization of images. To achieve these, Data Driven Discovery Predictive Model and multiple image stacking technique using Pivotal Focus Algorithms were built using Python programming language on Anaconda framework. The open-source web-based application Jupyter Notebook for coding and visualization was used. Different Regression models such as Polynomial Regression, Decision Tree Regression, Adaptive Booster Regression, Support Vector Regression and Random Forest Regression were attempted. The performances of the models were compared using R² (R-squared) and Mean Absolute Error (MAE). Based on the performance score, best performing model was suggested to predict subsurface temperature from resistivity. From the well recovery results, the Olkaria Domes reservoir can be classified as a convective heat flow system. Two main heat sources were inferred: One to the Northwest and the other to the Eastern side of the field. The two heat sources are separated by a NE-SW trending fault that is believed to control the fluid flow with natural recharge to the reservoir coming from the SW direction. The reservoir had two major feed zones at depths of (900-1300) m a s l and (250-0) m a s l. Step rate injection results indicated that both injectivity index and transmissivity are higher in the north east regions of the reservoir and decreases towards southwest. The resistivity structure of Olkaria domes at selected depths revealed three main resistivity regions; one low resistivity to the Northwest, the second low resistivity was observed to the Eastern side of the field. The two low resistive regions are separated by a third NE-SW trending high resistive region. Resistivity decreases with depth up to a depth of 500 m a s l then it increases with depth. From the resistivity cross sections, the results reveal three main resistivity zones. The first zone was characterized by a narrow layer of higher resistivity near the surface, likely to represent unaltered region. Underlying this layer was another broader layer of high conductivity that was interpreted to be due to high conductive hydrothermally altered mineralogy such as zeolites and smectites. A relatively higher resistive zone follows whose resistivity may be due to the formation of high temperature mineralogy at depth such as epidote. The image stacking of temperature and resistivity narrowed down Olkaria dome geothermal field into three main regions of interest. The first is located on the Northwest side of the field; the second is on the eastern side of the study area trending north-south. These were high temperature and high conductivity structures regions within the field. The third region is located in the southwest where both temperatures and conductivity were low. This could be the recharge zones where cold fluid is entering the reservoir. Decision Tree Regression (DTR) Machine Learning model was able to learn the trend of resistivity change and the predicted temperature graph matched well with the actual temperature graph. Training the model using the DTR algorithm approach provided superior outputs with the R2 of 0.81 and MAE of 29.8. DTR being the best algorithm based on the regression model employed was tested then with new data and the R2 was 0.835 with MAE at 21.7.