A.I. in Power Systems Alarm Processing

Transcription

A.I. in Power Systems Alarm Processing
A.I. in Power Systems Alarm Processing
Michael Nagy – PhD Scholarship Student
[email protected]
Affiliation:
NICTA Optimisation Group, NICTA Machine Learning
and Australian National University
Supervisors: Prof. Sylvie Thiébaux, Adj/Prof. Hanna Suominen,
Dr. Alban Grastien - Senior Researcher.
Email: [email protected]
Introduction and Background
Timed Hybrid Automaton for Critical Alarms
The Electricity Industry and its provision of power is ubiquitous. For a High
Voltage Power System (HVPS) to function safely and efficiently, a
number of automated systems are required to function in unison. One such
system is the Alarm System that notifies human operators of status and
condition of the Power System.
A power system in a metropolitan area can deliver thousands of alarms
per day that can overwhelm human operators. This is an alarm stream
which, in periods of high inflow, becomes an ’Alarm Avalanche’ or ’Alarm
Flood’ and has been around for 25 years since the advent of the Supervisory
Control and Data Acquisition System (SCADA). Much of the research has
been conducted in critical alarm conditions (3%) with very little on the
non-critical alarm conditions (97%). This Thesis will address both
conditions using an Finite State Timed Hybrid Automaton, Machine
Learning, Natural Language processing and Information Visualisation.
In this Thesis, critical alarms are defined as alarms produced by the HVPS.
In our novel approach, the Finite State Timed Hybrid Automaton uses the
alarms generated by the HVPS to identify critical alarms and create a casual
inference graph. A critical alarm, if not attended to, will result in
disintegration and ultimately failure of power system electrical apparatus.
Figure 3: Timed Hybrid Automaton
Machine Learning for Non-Critical Alarms
Figure 1: Alarm Breakdown: Critical to Non-Critical
Alarms
Problem to Solve
The problem is stated as:
’How to use Artificial Intelligence (A.I.) methodologies to assist human
operators to digest and respond to the significant volume of power system
alarm events given critical and non-critical power system conditions.’
A non-critical alarm is the complement of a critical alarm. Natural
Language Processing, Machine Learning and Information Visualisation
methods process the non-critical alarms generated by a HVPS. In our novel
approach, the edit distance of alarm messages are related to the same device
and location; Conditional Random Fields to recognise the named entities
(e.g., device name, geographic location, device status, data and time);
information visualisation to produce cross-sectional summaries (i.e.,
geospatial maps) and longitudinal trends of the incoming alarm stream with
an associated probability.
A significant portion of research for solving this problem has been based on
dealing with critical alarms which, from empirical evidence, is about 3% of
the total alarm volume.
The purpose of this Thesis is to find a complete solution for 100% of alarms
generated in order to support human operators’ situational awareness.
Figure 2: Breakdown of Alarm Sequence
Figure 4: Alarm Geospatial Mapping
Research Excellence in ICT Wealth Creation for Australia