Feature-based time series clustering for efficient labelling of geothermal data

• Temperature differences and flow resistance time-series data were computed to characterise heat exchanger performance. • Thirteen time-series features were identified as characteristic descriptors of the system via self-supervised feature selection. • Nine clustering algorithms operating on the time-series features were optimised and evaluated with respect to their effectiveness in discriminating different operational regimes via unsupervised learning. • Temporal cluster labelling demonstrated the effectiveness of the proposed machine learning pipeline to identify different operational states of the heat exchanges before and after cleaning. Geothermal energy is rapidly emerging as a pivotal component of the global energy landscape, driven by technological advancements, policy support, and the need for sustainable, reliable power sources. There has been global interest in transforming the geothermal industry by applying Artificial Intelligence and Machine Learning methods across the value chain to optimise operations and improve efficiencies. However, a lack of annotations and labels has limited machine learning and artificial intelligence applications in the geothermal industry despite the volumes of data generated in facilities worldwide. In this study, an analytics framework is developed leveraging advances in feature-based time-series analytics to automate label generation for geothermal time-series data. The framework was applied to annotate the history of geothermal heat exchangers in the Wairakei binary power plant in New Zealand. For this application, temperature difference and flow resistance time-series data were computed from available sensor data. From these time series, thirteen (13) time series features were chosen out of 1588 candidates via self-supervised feature selection. The time-series features were used as input to nine (9) different clustering algorithms. Pareto-optimal hyperparameters were computed with respect to the Silhouette coefficient (SIL), the Calinski-Harabasz Index (CHI), and the inverse Davies-Bouldin Index (iDBI). The best performing algorithms achieved a SIL of 0.66 (MeanShift), a CHI of 1647 (k-Means), and an iDBI of 2.48 (Gaussian Mixtures), corresponding to a DBI of 0.40. The resulting optimised clustering models identified data groups representing periods before and after a cleaning activity on the heat exchangers. The ability to automatically annotate the performance history of such vital components in the geothermal energy production opens up new avenues for machine learning and artificial intelligence applications within the industry. Simplified schematic diagram of the Wairakei geothermal power plant (left) and visualisation of the unsupervised machine learning performance in three-dimensional hyperparameter space (right). Simplified schematic diagram of the Wairakei geothermal power plant (left) and visualisation of the unsupervised machine learning performance in three-dimensional hyperparameter space (right).