ESP shutdown prediction using hybrid ... - spe … · for ESP shutdown prediction on a number of...

Copyright 2018 Baker Hughes, a GE company, LLC (“BHGE”). All rights reserved. The information contained in this document is company confidential and proprietary property of BHGE and its affiliates. It is to be used only for the benefit of BHGE and may not be distributed, transmitted, reproduced, altered, or used for any purpose without the express written consent of BHGE.

ESP shutdown prediction using hybrid support vector machine and improved parallel particle swarm optimizationMoradeyo Adesanwo, Oladele Bello, Lajith Murali, Ajish Kb, William Milne, BHGE

EuALF 2018 European Artificial Lift Forum13-14 June 2018, AECC, Aberdeen, UK

June 27, 2018

Copyright 2018 Baker Hughes, a GE company, LLC (“BHGE”). All rights reserved.

Outline

June 27, 2018 2

• Introduction and current challenges

• Current challenges in why predict ESP shutdown?

• Objectives

• Methodology

• Field examples - results and discussion

• Conclusions


Background and motivation for ESP-based well operations management

June 27, 2018 3

• Big data technologies are now being used to address many industrial problems & challenges related to high volume, high velocity, high variety

• Digitization using advanced sensors in intelligent oilfield solutions has resulted in the generation of huge amount of data that can be used to enhance the performance, operational health surveillance & management of ESP

• Need to use powerful analytics to estimate inexpensively and accurately equipment performance degradation, equipment failures, or detrimental interactions among subsystems

• Enhanced monitorability to reduce downtime and improve hydrocarbon recovery

Facility Performance Operational

Casual Data

- Reliability Data- Equipment Design Data - Well Design Data - Reservoir Data - PVY Data

- Cross Flow - Water Cut - Gas Cut- Prod. Data - Gas Coning - Water BT Time - Dynamic Reservoir Properties

Casual Data Casual Data - Motor Data- Pump Data - Controller Data- Set-Points - Maintenance Data- Equipment Condition

Benefits

Advanced Analytics

Intelligent Insights

Higher Production Rate

Lower Prod. Loss

Explicit View of Performance Alignment with Production KPI

Improved ProductionCasual analysis to identify process bottlenecksControl variability in the process

Value

Heterogeneity of data from dispatch control systemsWell management systems and SCADA

Variety Velocity Volume

Historical data amounting to several terabytes for analysis

Heterogeneity of data from dispatch control systemsWell management systems and SCADA


Current challenges in smart ESP-based well operations management

June 27, 2018 4

• Overwhelming downhole & surface sensor data has become too large and complex to be effectively processed by traditional approaches

• Acquiring, cleansing, processing & analyzing data has to be done in shorter time frames

• Downhole big data analytics requires a range of technologies (big data platforms – Spark; data warehouses and databases; data ingestion, pipeline, integration; virtualization and containerization of workloads; servers, storage and networking infrastructure; analytics model tuning, visualization and reporting applications;

• Current upscaling of online ESP surveillance practices, are largely determined by the size of the training data and the limited computational infrastructure for storing, managing, analyzing and visualizing big datasets

Kafka Cluster

Spark Streaming

Spark Engine

Cassandra

ESP Real-Time Analytics

ESP Offline/ Batch

Analytics

Dis

trib

ute

d D

ata

Acq

uis

itio

n C

lust

er

(co

nn

ect

ing

to

De

vice

s)

ESP Data Source

ESP Data Management

ESP Data & Analytics Service

Cassandra


Objectives

June 27, 2018 5

• Develop a general purpose system architecture for predicting ESP shut-down focusing on low-cost cloud –based computing platform

• Explore the feasibility of using real-time data collected by downhole sensors as the only source of information about the normal and abnormal operating conditions of a ESP system

• Investigate the potential of using data-driven artificial intelligence and statistical techniques for efficiently extracting and capturing information about the expected patterns of the signals coming from the ESP sensors, controller data, power data, surface data during the normal operating conditions of an ESP (normal operating patterns)


Methodology: proposed solution framework

June 27, 2018 6

ESP Operation Expertise

Pattern detection, alarm,

model and Prescriptive

recommendation

Data Access

• Right data is transmitted in the right time - reliably

• Data is available to people, tools, model calibrator where it can be use• No guess work on data -

Fuzzy logic based analysis for finding whether data is increasing, decreasing or constant

• Easy to recognize ESPs with issues – exception based surveillance (Exceptions are cross verified with the Prescriptive rules)

• Once the analyzed data and rules/pattern matches, alarms are triggered based on the prescription

• Subject Matter Expert experienced on operating ESPs in different application

• Support model institutional learning and continuous model calibration


Methodology: proposed deployment service platform

June 27, 2018 7


Methodology: proposed system architecture

June 27, 2018 8

CSV-Time Series

Shutdown Prediction Engine

(Rule based)&

(ML Apache Spark)

Real Time Data

History Data

DTS Datamining

DSS Datamining

DNP3

SCPI

Demand Trac

<PRODML/>1.2, 1.3, 2.0(DAS)

RTAWS Engine

Training Data from Gateway

Gateway UI

History Data

Configuration parameters

Visual alerts

Cluster

Real Time Data

Data Acquisition Data Streaming (RT/HA) and Storage Data Computation Data Visualization


Methodology: proposed spark computational cluster architecture

June 27, 2018 9

Spark Driver

Cassandra NameNode

Spark Worker

Cassandra DataNode

Spark Worker

Cassandra DataNode

Spark Worker

Cassandra DataNode

Master VM

Slave VM-6Slave VM-1 Slave VM-2


Methodology: ESP shutdown prediction model

June 27, 2018 10

Identify Problem Domain

ESP Data Selection ESP Investigate

Data Set

Normal ESP Conditions

ESP Shutdown Conditions

Classification

Clustering Algorithm

Low

Pattern Recognition Algorithm for Anomaly

Detection

Medium High

ESP Shutdown Prediction


Applications – case studies

June 27, 2018 11

Dataset for the case study

Normal Conditions

Testing set

Normal Conditions

Shutdown Conditions

02000 4000 6000 8000

Training set

Shutdown Conditions

0 100 200 400 600300 500

testing

training

data set

0 2000 4000 6000 8000 10000 12000

Distribution of dataset

TOTAL NUMBER OF WELLS ~350

Total number of shutdown 3,984

Maximum shutdowns per month 597

Minimum shutdowns per month 111

Average shutdowns per month 332

Total SCADA shutdowns 1,751

Total input power quality 1,516

Total operations/downhole 635

Total external signal shutdown 46

Total internal drive component 36


Important operating parameters that may be Inferred include ...

• Seal section oil expansion

• Driveline torque

• Thrust bearing loading

• Motor terminal voltage

• Motor terminal current

• Motor load factor

• Pump operating range

• Stator temperature

• Seal oil temperature

Important operating parameters that is typically monitored include ...

• Frequency

• Pump intake pressure and temperature

• Pump discharge pressure and temperature

• Internal motor operating temperature

• Flow rate

• Surface current and voltage

• Voltage imbalance

• Tubing pressure and casing pressure

• Current leakage

• Unit vibration

June 27, 2018 12




June 27, 2018 13

0

20

40

60

80

100

120

140

160

Total

Trip (Non-ESP trip) Underload Low Intake Pressure

High Motor Temp. High Intake Temp. High Discharge Pressure

Motor Stall Overload High Downstream Pressure

Troubleshooting Leak Low Speed Trip

Over Current

OPERATIONAL/DOWNHOLE TOTAL

Trip (non-ESP trip) 151

Underload 148

Low intake pressure 136

High motor temp. 57

High intake temp. 41

High discharge pressure 38

Motor stall 33

Overload 12

High downstream pressure 6

Troubleshooting 5

Leak 1

Low speed trip 6

Over current 1



June 27, 2018 14

44.0%

38.1%

15.9%

0.9%1.2%

Shutdown Category Breakdown

SCADA Shutdown

Input Power Quality

Operations/DownHole Issues

Internal Drive Component Issues

External Signal Intervention

197

5190 110 130 110

242 223

150

207

107134

4116

332 320343

6842 47

21

101 118

6746 41 54 53

114

26 37 47 3073 64 50

3 2 2 0 3 2 5 8 1 7 0 33 1 4 2 7 1 6 3 7 6 2 40

100

200

300

400

JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC

Monthly Num. of Shutdowns per Category

SCADA Shutdown Input Power Quality Operations/DownHole Issues Internal Drive Component Issues External Signal Intervention



June 27, 2018 15

0

5

10

15

20

25

30

35

40

45

JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC

Monthly number of operations/downhole issues per category

Trip (Non-ESP trip) Underload Low Intake Pressure High Motor Temp. High Intake Temp. High Discharge Pressure Motor Stall



June 27, 2018 16

151

148

136

57

41

38

3312 65161

Trip (Non-ESP trip)23.78%

Underload23.31%

Low Intake Pressure21.42%

High Motor Temp.8.98%

High Intake Temp.6.46%

High Discharge Pressure5.98%

Motor Stall5.20%

Overload1.89%

High Downstream Pressure0.94%

Troubleshooting0.79% Leak

0.16% Low Speed Trip0.94%

Over Current0.16%

Operations/downhole issues breakdown

Trip (Non-ESP trip)

Underload

Low Intake Pressure

High Motor Temp.

High Intake Temp.

High Discharge Pressure

Motor Stall

Overload

High Downstream Pressure

Troubleshooting

Leak

Low Speed Trip

Over Current


Conclusions

June 27, 2018 17

• An automated detection and prediction of shutdown has been designed and implemented for real-time Electrical Submersible Pumps (ESP) performance monitoring and diagnostics

• The main novelties are that the proposed approach extracts features from operating conditions based on clustering analysis and predict via data classification

• Test, verify and demonstrate the effectiveness and efficiency of the novel methodology for ESP shutdown prediction on a number of real-life case studies from the field

• Robust performance is guaranteed and versatile to various ESP applications – onshore, offshore, SAGD, etc.

• Future directions – more sophisticated anomaly detection algorithms for online self-learning on cloud-based distributed clusters

ESP shutdown prediction using hybrid ... - spe … · for ESP shutdown prediction on a number of...

Documents

Transcript of ESP shutdown prediction using hybrid ... - spe … · for ESP shutdown prediction on a number of...