Detection of abnormal operation in geothermal binary plant feed pumps using time-series analytics

Abstract

In this study, we investigate potential pump cavitation events in circulating feed pumps deployed in a geothermal binary power plant using feature-based time series analytics and machine learning. The feed pumps have been in operation for over a decade but have been experiencing episodes wherein the pumps were drawing in lower power even while pump speeds were maintained at the same levels.A workflow combining time-series analytics and first-principle thermodynamics was applied to characterise potential drivers of the abnormal events in the feed pumps. An approach based on systematic time-series feature engineering and semi-supervised machine learning was used to fully annotate the input data and identify twenty (20) undocumented occurrences of the target event. Time-series features extracted from the fully labelled dataset were then used to train forecasting models to identify potential drivers of the target events. Forecasting models trained using pump data outperformed those trained on times-series data from other power plant components (Matthew’s correlation coefficient score of 0.490 vs 0.255). This result indicates that the abnormal events are primarily driven by operating conditions local to the feed pump system. This finding was confirmed by Bayesian A/B testing, showing that the target events were more likely to occur by 22% when four (4) feed pumps were running compared to when only three (3) pumps were in operation. Furthermore, thermodynamic analysis of the state of the working fluid across the feed pumps shows irregularities at the pump inlet, where indications of partial vaporisation of the working fluid have been identified.