Slide 1

Slide 1 text

1 High-dimensional time series Rob J Hyndman robjhyndman.com

Slide 2

Slide 2 text

Outline 1 Sub-daily time series analysis 2 Time series feature analysis 3 Time series anomaly detection 4 Forecast reconciliation 5 Probabilistic electricity demand analysis 6 NUMBAT 2

Slide 3

Slide 3 text

Pedestrian counts 3

Slide 4

Slide 4 text

Pedestrian counts 3 −37.83 −37.82 −37.81 −37.80 −37.79 144.93 144.94 144.95 144.96 144.97 144.98 Longitude Latitude

Slide 5

Slide 5 text

Pedestrian counts 3 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec M T W T F S S M T W T F S S M T W T F S S M T W T F S S

Slide 6

Slide 6 text

Sub-daily time series analysis How to visualize many series of sub-daily data over several years? How to identify unusual patterns/incidents? How to forecast sub-daily data taking account of public holidays and special events? 4

Slide 7

Slide 7 text

Sub-daily time series analysis How to visualize many series of sub-daily data over several years? How to identify unusual patterns/incidents? How to forecast sub-daily data taking account of public holidays and special events? Di Cook Earo Wang Mitchell O’Hara-Wild 4

Slide 8

Slide 8 text

Outline 1 Sub-daily time series analysis 2 Time series feature analysis 3 Time series anomaly detection 4 Forecast reconciliation 5 Probabilistic electricity demand analysis 6 NUMBAT 5

Slide 9

Slide 9 text

Walmart weekly sales data 6

Slide 10

Slide 10 text

Time series feature analysis Can we use time series features for fast identification of forecasting models? How to generate new time series with specified feature vectors? What can we say about the feature space of time series? 7

Slide 11

Slide 11 text

Time series feature analysis Can we use time series features for fast identification of forecasting models? How to generate new time series with specified feature vectors? What can we say about the feature space of time series? Kate Smith-Miles George Athanasopoulos Thiyanga Talagala 7

Slide 12

Slide 12 text

Outline 1 Sub-daily time series analysis 2 Time series feature analysis 3 Time series anomaly detection 4 Forecast reconciliation 5 Probabilistic electricity demand analysis 6 NUMBAT 8

Slide 13

Slide 13 text

Security monitoring 9

Slide 14

Slide 14 text

Security monitoring 10

Slide 15

Slide 15 text

Time series anomaly detection How to identify anomalous behaviour within streaming data? How to define an anomaly in a large multivariate data set? 11

Slide 16

Slide 16 text

Time series anomaly detection How to identify anomalous behaviour within streaming data? How to define an anomaly in a large multivariate data set? Kate Smith-Miles Mario Andrés Muñoz Acosta Sevvandi Kandanaarachchi Dilini Talagala 11

Slide 17

Slide 17 text

Outline 1 Sub-daily time series analysis 2 Time series feature analysis 3 Time series anomaly detection 4 Forecast reconciliation 5 Probabilistic electricity demand analysis 6 NUMBAT 12

Slide 18

Slide 18 text

Forecast reconciliation 13 Total A AA AAA AAB AAC AB ABA ABB ABC AC ACA ACB ACC B BA BAA BAB BAC BB BBA BBB BBC BC BCA BCB BCC C CA CAA CAB CAC CB CBA CBB CBC CC CCA CCB CCC

Slide 19

Slide 19 text

Forecast reconciliation Forecasts at all nodes must be coherent Bottom level typically has thousands or millions of time series How to define coherence probabilistically? How to visualize so many time series? 14

Slide 20

Slide 20 text

Forecast reconciliation Forecasts at all nodes must be coherent Bottom level typically has thousands or millions of time series How to define coherence probabilistically? How to visualize so many time series? 14 George Athanasopoulos Anastasios Panagiotelis Shanika Wickramasuriya Puwasala Gamakumara Earo Wang

Slide 21

Slide 21 text

Outline 1 Sub-daily time series analysis 2 Time series feature analysis 3 Time series anomaly detection 4 Forecast reconciliation 5 Probabilistic electricity demand analysis 6 NUMBAT 15

Slide 22

Slide 22 text

Electricity demand 16

Slide 23

Slide 23 text

Electricity demand 16

Slide 24

Slide 24 text

Electricity demand 17 Monday Tuesday Wednesday Thursday Friday Saturday Sunday 1539 1549 0 6 12 18 24 0 6 12 18 24 0 6 12 18 24 0 6 12 18 24 0 6 12 18 24 0 6 12 18 24 0 6 12 18 24 0 2 4 6 0 2 4 6 Time of day Demand (kWh) Percentile 10 25 50 75 90

Slide 25

Slide 25 text

Electricity demand How to forecast future demand by household? How to reconcile household demand forecasts with state and national demand forecasts? How to identify unusual demand patterns? How to measure forecast accuracy when forecasts are probability distributions within a hierarchy? 18

Slide 26

Slide 26 text

Electricity demand How to forecast future demand by household? How to reconcile household demand forecasts with state and national demand forecasts? How to identify unusual demand patterns? How to measure forecast accuracy when forecasts are probability distributions within a hierarchy? Souhaib Ben Taieb Cameron Roach 18

Slide 27

Slide 27 text

Outline 1 Sub-daily time series analysis 2 Time series feature analysis 3 Time series anomaly detection 4 Forecast reconciliation 5 Probabilistic electricity demand analysis 6 NUMBAT 19

Slide 28

Slide 28 text

NUMBAT: Non-Uniform Monash Business Analytics Team 20 Test Post-docs Nick Tierney Sevvandi Kandanaarachchi Shanika Wickramasuriya