PREDICTION AND PRESCRIPTION OF OPERATION UPSET IN H2S GAS SWEETENING UNIT: IMPLEMENTATION OF AN INNOVATIVE BIG DATA ANALYTICS PROCEDURE

摘要

This paper highlights the development and results of a machine-learning based end-to-end system for process upset and hazard events prediction in gas-sweetening procedures; this tool has been applied to production operations of an oil-and-gas field. High H2S concentration in the produced gas represents a serious issue due to its environmental impact, the impossibility to deliver acid gas to the distribution network and the asset deterioration. The proposed tool monitors the status of the equipment in near real-time. Whereby an alarm is raised, prescriptive information is provided to avoid, or mitigate, operational issues. This can be accomplished by using machine learning algorithms and data mining techniques in a Big Data Infrastructure. In the illustrated case, a complex data-lake was built by ingesting and aggregating in a Big Data Environment times-series data from field sensor network, maintenance reports and chemical analyses. A machine learning algorithm has been trained to identify faults in the gas-sweetening unit resulting in a high concentration of H2S in the processed gas. The development of the tool has been conducted in collaboration with site engineers and operators to identify the most relevant data sources describing the process and to validate the algorithm outputs. Several machine learning algorithms have been tested (Deep Learning, Random Forest, Gradient Boosting Trees) to improve model accuracy and clarify the interpretation of the phenomenon root causes. Finally, the tool is now fed with real-time data and predicts hazardous events in near real-time. The alerts raised by the system are stored and archived in the Big Data Environment for further analysis. Field operators and process engineers can therefore access those new insights, and the related data, using the tools already in use during the daily monitoring operations. Alongside, a dedicated visualization tool was designed to monitor the model performances and guarantee its life-cycle. The innovative characteristics of the tool lay in its ability to exploit the huge amount of field-data and to simulate complex phenomena through Big Data Analytics. It is now possible for the site operators to receive preventive warnings on relevant events, gather information on the possible root causes and on the recommended actions to prepare for the upcoming event. Ultimately, this framework allows to insure the constant flow of the gas into the distribution network, to avoid or mitigate halts in production and to guarantee asset integrity.