The Reservoir Temperature Prediction Using Hydrogeochemical Indicators by Machine Learning: Western Anatolia (Turkey) Case

Abstract

Although geothermal fluids can be used with different purposes, it requires expensive detailed exploration studies for a region. These geothermal exploration studies mainly consist of; geology, geophysics and geochemistry disciplines to understand location, dimension, possible capacity and temperature of the source before a drilling operation. Geothermal drilling operations are also complex and quite expensive operations that the costs may change 1.8-2.9 million USD for 1500-2500 m in Western Anatolia (Turkey). Because of the high operational costs, exploration phase of the geothermal projects is of great importance to reduce project costs. Evaluation of existing earth sciences data, detailed geology studies, mapping and some geochemical studies such as using geothermometers can provide important information to reach geothermal reservoir in a geothermal field. Nowadays, developing technology may give a chance to predict geothermal reservoir temperatures in geothermal fields. Machine learning (ML) is a data analytics technology that teaches computers to do learn from ML algorithms use computational methods to learn information directly from data and these methods adaptively improve their performance as the number of samples or information available for learning increases. The technology addressed to many study areas such as; science, marketing, engineering. In this study, a deep learning classifier has been developed to classify predicted geothermal reservoir temperatures based on the hydrogeochemistry data. We defined three main categories for the geothermal reservoir temperatures to classify our data: Low (50-89 °C), medium (90-150 °C) and high ( more than 150 °C). To compare the prediction performance of our proposed deep learning classifier, two traditional classification approaches, k-nearest neighbors (KNN) and linear support vector machines (SVM), have been performed. and the results have been presented. The results have been obtained as categories and the models have been compared with their accuracy in this study. The performance comparison showed that our deep learning model achieved better prediction performance than traditional machine learning techniques.