ModelBased Reasoning a Case Study Jan Eric Larsson
Model-Based Reasoning – a Case Study Jan Eric Larsson WASP Autonomous Systems 2, HT 2018 Lund University, Lund, October 5 th 2018 Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com
Ordinary Alarm list • Too much information – – – Different types of alarms Different levels of importance Many alarms on the same thing • Beyond human capacity to effectively understand the complete situation • The alarm list becomes useless (really!) during incidents • Look elsewhere to really understand what happened… Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 2
Goal. Art Alarm Display Root Causes Consequences • Alarms are grouped per equipment • Suppression of chattering alarms • Suppression of long-standing alarms • Suppression of non-grid alarms • Real-time root cause analysis algorithm • 100 % automatic configuration • Based on CIM or other EMS model • Integration with SCADA/EMS system Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 3
Blackout September 23 rd 2003 • • • Large blackout in Scandinavia September 23 rd 2003, 12. 35 PM Root causes – – • Consequences – – • Two lines for all of southern Sweden Southern Sweden collapsed (5 -15 min) Eastern Denmark collapsed Lasted 1 -5 hours Actions – – • 12: 30 OKG 3 nuclear reactor trip (east) 12: 35 Internal station short-circuit (west) Second root cause unknown for 4 hours Helicopters looking for line faults Cost – – – Lost ~ 10 000 k. Wh Cost ~ 500 000 USD Largest disturbance in 22 years Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 4
2003 -09 -23 12: 35: 30 Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 5
2003 -09 -23 12: 37: 00 Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 6
2003 -09 -23 12: 38: 00 Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 7
The Old Horred Sub-Station Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 8
The Real Root Cause Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 9
Integration • Typical integration uses CIM, OPC DA, and OPC A/E Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 10
HOPS New Control Rooms, Croatia Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 11
How Does It Work? Pump • Simple example – • • Tank Pump and closed tank Described as transport and storage Four consequence propagation rules are valid for this connection Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 12
Model Properties • Discrete model – Either correct or not correct, never approximately correct • Generated automatically from CIM or EMS • Complete and correct algorithm – – – • If the model is correct and the SCADA data are correct, the results are correct Can handle any fault combination Multiple, indepedent faults Efficient – Sub-second calculation times for large transmission grids Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 13
Artificial Intelligence • Goal. Art is AI but not machine learning or big data • Model-based reasoning • Difficult fault situations are – – – Unexpected combinations of several faults Will happen once and never again No possibility of training or statistics • Called “once in a lifetime” events • The Goal. Art system will be right the first time • Learning algorithms are unpredictable in this case Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 14
Goal. Art Algorithms Based on MFM • Root Cause Analysis • • Fault Diagnosis Probability and Safety Analysis Failure Mode Analysis Action Planning • • Measurement Consistency Analysis, analog/discrete Redundancy Analysis • Dynamic Detection of Congestion Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 15
General MFM Algorithm Properties • Completeness – – • Easy modeling – – • Covers all cases (if the model and inputs are correct, the result is correct) First time right Handles multiple faults and circular dependencies Predictable in unexpected situations (so-called “once-in-a-lifetime” events) Simple concepts, discrete logic Automatic model generation from CIM or other EMS/DMS model Efficient – Can handle large systems (such as large power grids) • All algorithms work from the same MFM model • Artificial intelligence, but not learning or big data Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 16
Comparison with Rule-Based Systems • Goal. Art • Rule-based (typical) • • Based on CIM or similar data Compilation (100 % automatic) • • Based on human knowledge Large engineering effort • • Complete knowledge base Handles all potential cases First time right Handles unexpected combinations • • Incomplete knowledge base Difficulties if several root causes No guarantee of first time right Unpredictable in unexpected situations Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 17
Comparison with Learning Algorithms • Goal. Art • Learning algorithms (typical) • • Based on CIM or similar data Compilation (100 % automatic) • • Based on saved case data Manual effort to find test cases • • Complete knowledge base Handles all potential cases First time right Handles unexpected combinations correctly • • Incomplete knowledge base Difficulties if many root causes No guarantee of first time right Unpredictable in unexpected situations • Automatic reconfiguration if new or changed system • New training period if new or changed system Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 18
Current Application Areas Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 19
Two Questions • What is the customer benefit? • What is the cost of – – – • Implementation/installation Educating human users Maintenance and upgrades when target system changes These questions are more important for success than technological properties Goal. Art, Scheelevägen 17, 223 70 Lund, Sweden, Jan Eric Larsson, janeric@goalart. com 20
- Slides: 20