AMINE GAS TREATMENT CASE PROACTIVELY PREDICT COLLAPSE OF AN AMINE PROCESSING CONTACTOR

AMINE GAS CASE PREDICT COLLAPSE OF PROCESSING CONTACTOR Save $200M USD per year Foaming / Flooding events: 2 to 4 events per contactor per year Costing $250,000 per event 200 contactors worldwide Impact: Lost revenue High unplanned maintenance costs Poor amine utilization 2

AMINE GAS CASE SAS DAY AT TEXAS A&M: PREDICT COLLAPSE OF PROCESSING CONTACTOR Project Objectives Data Overview Agenda Analytical Approach Summary of Analytical Findings Lessons Learned Requirements for Big Data Predictive Analytics

AMINE GAS CASE AMINE GAS TREATMENT CASE OBJECTIVES Business Challenge Reduce collapse of Amine processing contactor due to foaming and flooding Business Impact Proactively predict foaming and flooding events Predict and monitor processing liquid levels Optimize Amine utilization Value Driver Improve up-time and processing efficiency

AMINE GAS CASE GOAL: PROACTIVELY PREDICT PROCESSING FAULTS Lean Amine to Contactor Foaming/Flooding Event Excess Amine Pickup Low Amine Gas Pickup

AMINE GAS CASE PROCESSING CONTACTOR SYSTEM OVERVIEW Contactor C101 CO 2 treated C-101 C-102 CO 2 feed gas Storage Tanks 41, 42 & 43

AMINE GAS CASE ANALYTICAL APPROACH SAMPLE EXPLORE MODIFY MODEL ASSESS Consolidate multiple data sources Quality Check Cleanse Visualize sensor and fault data Collaborate with subject matter experts Identify candidate anomalies Transform variables Filter outliers Cluster similar acting groups Variable selection Identify early fault indicators Root cause analysis Generate predictive models Decision trees Logistic regression Neural networks Review results with subject matter experts Score new data Automate ongoing assessment of prediction accuracy Flag model needing tuning Alert when operation becomes unstable

AMINE GAS CASE DATA OVERVIEW Source: PI System Systems: Contactor C-101: 22 sensor variables & 18 calculated variables Contactor C-201: 22 sensor variables & 18 calculated variables Contactor Feed Flow: 1 sensor variable Storage tanks: 3 sensor variables (1 per tank) Time range: January 1, 2010 to April 12, 2014

OSIsoft Overview Privately held software company One Product. Singular Focus. Agnostic Leader in Operational Intelligence Founded 1000 250+ Employees Developers Of Global Fortune 500 65% Customer Support # of Sites 16,000 C o p yr i g h t 2 0 1 4 O S Is o f t, L L C. 9

OSIsoft is trusted by the world s leading companies Oil & Ga s Pha rm a c e utic a ls, Food & Life Sc ie nc e s Pulp & Pa pe r Critic a l Fa c ilitie s, Da ta Ce nte rs & IT Innovative uses to monitor complex IT environments De ve loping Ma rke ts & Industrie s Connected Services, Smart Cities, Transportation, Healthcare, Academia C o p yr i g h t 2 0 1 4 O S Is o f t, L L C. 10

C o p yr i g h t 2 0 1 4 O S Is o f t, L L C.

C o p yr i g h t 2 0 1 4 O S Is o f t, L L C. 13

C o p yr i g h t 2 0 1 4 O S Is o f t, L L C. 14

C o p yr i g h t 2 0 1 4 O S Is o f t, L L C. 15

ANALYTICAL APPROACH www.sas.com

AMINE GAS CASE ANALYTICAL WORKFLOW Identify Time periods to investigate Events to investigate Explore See potential trends/areas to investigate Leading cause of failures Identify potential explanatory variables Analyze Perform root cause analysis Statistical signatures for events Stability modeling and monitoring

AMINE GAS CASE EXPLORE SENSOR DATA AND FAILURE TIMES February 10, 2013 to April 12, 2014

AMINE GAS CASE ANALYTICAL METHODS EMPLOYED Pareto Analysis Correlation Analysis Variable Cluster Analysis Principal Component Analysis Association Analysis Root Cause Analysis Decision Trees Logistic Regression Stability Monitoring

AMINE GAS CASE CORRELATION ANALYSIS Purpose: Measures strength of a linear relationship between numerical variables Benefits: Improves accuracy and performance of predictive models

AMINE GAS CASE VARIABLE CLUSTER ANALYSIS AND PRINCIPAL COMPONENT ANALYSIS Cluster Analysis Principal Component Analysis Purpose: Identifies similar acting variables or groups Variable reduction technique Benefits: Identifies most meaningful sensor data and events to further analyze

AMINE GAS CASE MULTIPLE LINEAR REGRESSION Purpose: Predicts contactor processing liquid levels Explains variables or combination of variables affecting contactor processing liquid levels Benefits: Analyzes more information generating greater predictive power Generates better understanding of relationships between variables Dependent Var: Gas Treating Area Pickup Ratio Explanatory Var: Flash Drum Level Feed Filter Separator Level CO 2 in treated gas Tray 21 Level Dependent Var: Contactor Level Feed Gas Pressure DSG Stripper Explanatory Temp Var: DSG Stripper Temp D101 NGL Train Separator Feed Gas Temp Feed gas - Lean DGA temp Feed gas Lean DGA temp Contactor Level D201 NGL Train Separator Lean DGA % Strength Dependent D101 Var: NGL Feed Train Gas Separator Flow NGL Train Separator Level Lean DGA % Strength Explanatory Var: Lean DGA % Strength Feed Gas Temp Feed Filter Separator Level DGA Antifoam Tank Level Tray 21 Level Contactor Level Flash Drum Level Flash Drum Level Lower Tray #5 Temp NGL Train Separator Level

AMINE GAS CASE CONTACTOR FOAMING/FLOODING DECISION TREE Purpose: Illustrates all possible outcomes and highlights typical path Benefits: Visually identifies variables considered relevant to target event occurrence or non-occurrence

AMINE GAS CASE ROOT CAUSE ANALYSIS & ASSOCIATION ANALYSIS Period: Jan 1/2013 to Dec 31/2013 Foaming/Flooding Event Analytical rule Lift PDI205_v_gt95p & PDI022_v_gt95p ==> TARGET_EVENT 17.86 PDI205_v_gt95p & PDI006_v_gt95p ==> TARGET_EVENT 16.76 TI028_v_gt95p & PDI205_v_gt95p ==> TARGET_EVENT 15.94 TI028_v_gt95p & FIC004_v_gt95p ==> TARGET_EVENT 15.59 PDI022_v_gt95p & FIC004_v_gt95p ==> TARGET_EVENT 15.52 TI026_v_gt95p & FIC004_v_gt95p ==> TARGET_EVENT 15.45 TI026_v_gt95p & PDI022_v_gt95p ==> TARGET_EVENT 15.12 Purpose: Helps understand why processing fault occurred Identifies variables potentially contributing to fault/failure event Benefits: Identifies association rules to predict occurrence of future faults/failures

AMINE GAS CASE ALERT FOAMING/FLOODING ROOT CAUSE PATTERN RECOGNITION WARNING

AMINE GAS CASE CONTACTOR LEVEL STABILITY MONITORING MODELS ARIMA REGRESSION Purpose: Provides predictive capability to show current state and alert of potential problems Benefits: Automatically alerts when operation becomes unstable Provides better forecasting accuracy Detects and accounts for daily and seasonal operational differences Provides better explanatory power Traditional approach used by engineers

AMINE GAS CASE SCORE NEW DATA AGAINST PREDICTION MODELS

AMINE GAS CASE ALERT UNSTABLE OPERATION TO PREVENT FUTURE FOAMING/FLOODING EVENTS

AMINE GAS CASE SUMMARY OF ANALYTICAL FINDINGS Proactively predict foaming and flooding events Discovered foaming/flooding fault signatures Predict and monitor processing liquid levels Developed prediction models and stability monitoring forecast models Optimize Amine utilization Developed stability model to feed optimization model

AMINE GAS CASE LESSONS LEARNED Sensor and fault data can easily be analyzed Wide variety of analytical methods available Tremendous business value discovering root cause of faults and failure initiating variables

AMINE GAS CASE REQUIREMENTS FOR BIG DATA PREDICTIVE ANALYTICS Data: High-Performance Architecture: Scalable Performance Data Server Analysis: