SLIDE 1
Today’s presentation
Problem definition Data discovery and transformation Model building and selection State of play
2
Machine Learning for rare events Peter Condon Business Intelligence - - PowerPoint PPT Presentation
Machine Learning for rare events Peter Condon Business Intelligence - - PowerPoint PPT Presentation
Machine Learning for rare events Peter Condon Business Intelligence and Data Analytics Todays presentation Problem definition Data discovery and transformation Model building and selection State of play 2 Who we are State
SLIDE 2
SLIDE 3
State Government-owned corporation that builds, maintains and operates electricity network throughout majority of south Western Australia Governed by an independent Board and reports to Minister for Energy, as owner’s representative Serving more than one million customers across a network area of 254,920 km2 Customer-orientated organisation that provides a safe, reliable and affordable electricity supply to Western Australians Provides an essential service through transmission and distribution of electricity across our vast infrastructure of poles, wires, substations and depots
Who we are
3
SLIDE 4
Generators make electricity at power plants. The network grid is made up
- f transmission and
distribution assets. The network enables electricity to flow from generators to consumers. Retailers issue accounts to collect revenue for the whole supply chain.
What we do
4
SLIDE 5
Problem definition
SLIDE 6
Problem definition
Pole top fires affect any tiny fraction of the network each year, but occasionally cluster geographically and temporally Modelling of pole top fires has traditionally been for proactive maintenance There is anecdotal evidence of leading indicators, however a formal short term forecast could allow for more fault crews to be available when needed
1
6
SLIDE 7
Pole top fire definition
“Pole top fires occur when the electrical energy from leakage currents between any two phases and/or between any phase and earth passing through conductive paths in the pole top components, generates sufficient localised heat to ignite or char flammable cross-arm and/or pole material, nominally timber.”
- Electricity Networks Association
1
7
SLIDE 8
Why does a pole top fire occur?
Weather patterns
–Dust build up –Light rain
Asset condition
–Insulator defects –Cross-arm defects –Voltage
8
SLIDE 9
Data discovery and transformation
SLIDE 10
Pre-existing datasets
Pole top fire master list Weather observations and forecasts
–Temperature, humidity, rain, wind, and air pressure
NRMT risk
–Insulator model
Unexplained outages
–Recloser trips and unknown cause outages
10
SLIDE 11
Unavailable Datasets
Wind gust forecast Air quality
–Particulate matter
2
11
SLIDE 12
Collating the Data
Collate incidents and risk factors by feeder
–950 feeders across the network offer natural grouping based
- n network topology
Aggregate to nearest weather station
–17 weather stations with sufficient historic observations and forecasts
2
12
SLIDE 13
Model building and selection
SLIDE 14
Fail fast
Deterministic functional form (Proc NLIN)
–Too many combinations of factors (gave up after 25 million iterations)
Generalised Additive Model (Proc GAM)
–Over fitting issues
AutoNeural(Enterprise Miner)
–Unable to optimise on so few non zero targets
2
14
SLIDE 15
Gradient Boosting pros
Actually worked
–After increasing default Node Sample
Fast
–Run time typically measured in minutes to a couple of hours
No need to specify data interactions
2
15
SLIDE 16
Gradient Boosting cons
Limited explicability
–Scoring code and variable importance lists can be difficult to interpret
Art to hyperparameter tuning
–Setting N Iterations too high or too low will lead to a poor model
2
16
SLIDE 17
Model Selection
Weather stations cover large geographic areas
–Fires may occur during relatively benign conditions at point
- f measurement
Actual predictions from the model cover a very small range (e.g. between 0 and 0.03) Champion model deemed to be the one that does the best job of ranking riskiest days
2
17
SLIDE 18
State of play
SLIDE 19
Model validation
Seven independent models are being scored each day A champion model will be selected in November for display
- ver summer
Next steps to be reviewed in April
2
19
SLIDE 20
Dashboard
2
20
SLIDE 21
Dashboard
2
21
SLIDE 22
Dashboard
2
22
SLIDE 23