通过人工智能优化esp的操作参数来提高石油产量。第二阶段

In 2021, Vital Energy partnered with SoftServe to prove the feasibility and validate the business value of an ML-basedoptimization solution for ESP wells. The primary objective was to maximize well life and increase production rates by optimizingESP systems. We created a proof of concept (PoC) phase that included the ML model trained on small data samples anddashboard prototyping.Check out this PoC case study here. After the project’s first phase, we moved to the pilot phase to extend the existing functionality and unveil a completed modelthat can be applied to all wells currently in operation. BUSINESS CHALLENGE During the PoC phase, we establishedrelationships in well controls andidentified optimization opportunities. Ourfocus during the PoC was on optimizingESPs that had been running at a stablefrequency for a while and were notshowing signs of gas locking, critical scalebuildup, or pump malfunctioning. By addressing these optimizationopportunities, we demonstrated thepotential for even greater improvementsin efficiency and performance and provedthe technology's ability to safely boostproduction. incorporated into an equipment healthand data monitoring system, along with adeveloped data quality monitoring system. Additional data annotation andinterpretation solutions were presentedto help capture subject matter experts'knowledge and incorporate it into theML models. After the PoC, development efforts werefocused on scaling the solution to allwells running on ESPs. A top priority wascreating effective approaches for anomalydetection, such as gas interference at anearly stage, scale buildup, near-frackingactivities, and many others. To achievethis, multiple machine learning (ML)models and dashboards were created.Anomaly detection ML models were As a result, the ESP optimization solutionis now enhanced with additional modelsand dashboards, making it more valuableand creating a highly scalable solution thatcan work under different circumstancesand configurations. THE APPROACH The SoftServe team developed thedata processing and visualization parts,a set of automated python scriptsand dashboards with well productionoptimizations and model forecastsanalysis. The dashboards allow usersto interact with the developers and themodel itself with optimization feedbackand telemetry constraint forms. The SoftServe and Vital Energy teams collaborated on ESP wellproduction optimization, focusing on an ML model designed foraccurate production forecasting. The model takes ESP telemetrydata and well production rates for the days before prediction andpredicts well production for the following day with given ESP controlparameters. In addition, the ML model predicts ESP telemetry forthe next day. This approach allows the client to not only predictwater, gas, and oil outcomes for different control parameters to findthe most optimal ones but also describe the approximate state ofthe system (temperatures, pressures, power consumption, etc.). With the introduction of the analyticsportal, all existing and new dashboardsare available in one place. Access tothe data and insights, cooperation,and solution adoptions are even morestraightforward. Dashboard accessis restricted based on the user’s role.Users, user groups, and dashboardcards can be managed from the adminpanel. The analytics portal is easilyaccessible from any type of device:mobile, tablet, and desktop. It alsoallows for the quick deployment of newdashboards to production, making itpossible to test new machine learningmodels and routines right after theprototyping phase. Conceptually, the solution consists of two parts: The dashboardsvisualizing the results Automated python scripts thatrun with AWS step functions THE SOFTSERVE TEAM HELPED DEVELOPTWO MORE DASHBOARDS: The Data Health Monitoring dashboardallows the client to monitor the data completeness and prioritize wells with the highestnumber of NAs or noise for further inspection. It classifies wells according to four different health categories (no data, high risk, mediumrisk, and normal) based on the average health statistics thresholds. Main data health statistics includebut are not limited to: List of available datasources and theirassociated completenesspercentage. Number andpercentage missingcritical tags. Percentage of tagvalue duplicates, zeros,out of operational andhistorical statisticalranges. Percentageof anomalies andrelative completenessand data frequencychanges. The Gas Interference Detection dashboardmonitors wells and alarms about gas interferenceand gas lock events. It also enables the labelingof these and other events for further ML modelre-training. ESPs have gas lock and PID controllers thatactivate when there is too much gas in thesystem. Gas lock events lead to a reduction in oilproduction and faster equipment deterioration.The current gas inference detection algorithmprovides operators with daily forecasts of gasinterference events that provide early warningsigns about potent