Machine learning for creation of generalized lumped parameter tank models of low temperature geothermal reservoir systems

Lumped parameter tank models have gained renewed interest in recent years as an alternative tool for geothermal reservoir analysis and production planning. The models can be structured in various ways regarding the number of tanks, connections between the tanks and the parameters representing the physical properties of the geothermal system. It usually requires a time consuming and difficult process of trials and errors to manually decide the optimal configuration of a tank model. Inspired by recent development in the use of machine learning methods, we propose a method for automatically generating accurate and computationally feasible generalized tank models for isothermal, single phase, reservoirs. This is an extension of earlier work on complexity reduction of generalized tank models (Li et al., 2016). Here, a recursive “switch-back” method is constructed to maximize prediction accuracy of the model. It is also shown how the K-means clustering algorithm can be used to aggregate production wells in generalized tank models. One synthetic example and one field application from t Reykir geothermal fields in Iceland are used to illustrate the effectiveness of these methods.