A visual analytics approach to anomaly detection in hydrocarbon reservoir time series data*

Detecting anomalies in time series data of hydrocarbon reservoir production is crucially important. Anomalies can result for different reasons: gross errors, system availability, human intervention, or abrupt changes in the series. They must be identified due to their potential to alter the series correlation, influence data-driven forecast, and affect classification results. We have developed a visual analytics approach based on an interactive visualization of time series data involving machine learning approaches for anomaly identification. Our methods rely upon a z-score normalization technique along with isolation forests. The methods leverage the prior probability of anomalies from a time-window, do not require labeled training data with normal and abnormal conditions, and incorporate specialist knowledge in the exploration process. We apply, evaluate, and discuss the methods' capability using a benchmark data set (UNISIM-II-M-CO) and real field data in three visual exploration setups. The ground-truth annotations were done by human specialists and considered different interventions in the reservoir. Our methods detect approximately 95% of the human intervention anomalies, and about 82%-89% detection rate for other anomalies identified during data exploration.