Master Electrabel 2003

1
Electricity gas load forecastingin Belgium
ULB - 27/02/2007 Olivier DucarmeLoad
forecasting manager
2
Agenda

• Electrabel key figures
• Load forecast why ?
• Impact of liberalization
• Risk factors impacting the load
• Forecasting methods
• Modelling techniques
• From load to price
• Key issues for the future

3
Agenda

• Electrabel key figures
• Load forecast why ?
• Impact of liberalization
• Risk factors impacting the load
• Forecasting methods
• Modelling techniques
• From load to price
• Key issues for the future

4
Key figures Sales 2005
Electricity Total sales GWh 145
447 Benelux 100 149 Europe outside Benelux
45 298 Number of final customers 5 485
903 Benelux 3 758 922 Europe outside
Benelux 1 726 981 Natural gas Total
sales GWh 73 337 Number of final customers 2
027 254
5
Key figures Generation 2005
Electricity Net generating capacity MW 29
084 Benelux 18 252 Europe outside
Benelux 10 832 Net generation GWh 130
742 Benelux 89 197 Europe outside
Benelux 41 545 Heat Net generation GWh
12 724 Benelux 7 224 Europe outside
Benelux 5 500
6
Key figuresStaff Finance 2005
Staff Number of employees 15 794 Benelux
12 282 Europe outside Benelux 3
512 Finance million Revenue 12
218 EBITDA 2 378 Result from operations 1 444
7
Revenue million - 2005
Electricity Benelux Electricity outside
Benelux Natural gas Services and others
2005
12 218
8
Electrabel 2005
Germany Sales 7 722 GWhGeneration 296
MWStaff 165 Poland Sales 8 079 GWh
Generation 1 654 MWStaff 1 415 Hungary Sales
3 763 GWhGeneration 1 676 MWStaff
438 Italy Sales 13 667 GWhGeneration 2 225
MWStaff 1 125
The Netherlands Sales 22 325 GWh Generation 4
711 MW Staff 810 Belgium Sales 75 230
GWh Generation 13 165 MW Staff 11 452
Luxembourg Sales 2 594 GWh Generation 376
MW Staff 20 France Sales 12 027 GWh
Generation 4 818 MW Staff 319 Spain Sales 3
GWh Generation under construction Staff 51
Portugal Sales 37 GWh Generation 164 MW
Electrabel is an active trader on all of
Europes energy markets.
9
Generating capacity in EuropeRenewable energy
sources 2005
The NetherlandsWind 3.5 MWBiomass 65
MW BelgiumWind 58 MWHydroelectric 22 MWBiomass
255 MW ( 174 MW)Wind 16 MW FranceHydroelectric
3 710 MWWind 22 MW PortugalWind 131
MWHydroelectric 33 MWWind 90 MWWind 356 MW
PolandBiomass 160 MW ItalyHydroelectric 129 MW
Commissioned in 2005Under construction end 2005
10
Generating capacity in the Benelux2005
Combined cycle gas turbine (CCGT) Cogeneration
with gas turbine Conventional thermal power
station Nuclear power station Pumped-storage
power station
11
Generating capacity in the BeneluxRenewable
energy sources 2005
Wind farm Biomass co-combustion inconventional
thermal power station Hydroelectric power
station
12
Generating capacity and generation Per type of
unit Net 2005
Nuclear Hydroelectric and wind
Combined cycle gas turbine Cogeneration Convention
al thermal
Capacity 29.1 GW
Generation 131 TWh
13
Generation By fuel type Net 2005
Gas Coal, biomass Fuel oil Nuclear Hydroelectric
and wind Energy recovery
0.4
12.4
35.1
131 TWh
37.1
13.8
1.2
14
Electrabel business model
Market
Fuels
Marketprices
Trading
Procurement
Internal Portfolio Management
Generation
MS
Loadforecasting
Customers
Power plants
15
Agenda

• Electrabel key figures
• Load forecast why ?
• Impact of liberalization
• Factors impacting the load
• Forecasting methods
• Modelling techniques
• From load to price
• Key challenges for the future

16
Load forecast why ?

• To insure a stable network (stable tension 50
  Hz frequency), the production must always
  perfectly fit the load.
• Any imbalance (difference between load
  production) induces a deviation of frequency
  or/and tension with a risk of black-out by a
  domino effect (at European level - cf. incident
  in Germany on 4th Nov. 2006 impacting France
  Belgium)
• More transparency between countries for loadflow
  simulation, N-1 constraint

17
Load forecast why ?
Load( off-take)
Production( injection)
Imbalance (real time)
50 Hz
18
Agenda

• Impact of liberalization

19
Agenda

• Electrabel key figures
• Load forecast why ?
• Impact of liberalization
• Risk factors impacting the load
• Forecasting methods
• Modelling techniques
• From load to price
• Key issues for the future

20
Impact of liberalization

• Split between
• Transmission grid (gt 11 kV) TSO (Transmission
  System Operator) Elia
• Distribution DGO (Distribution Grid Operators)
• Production Sales
• Elia responsible to maintain the equilibrium
  between load and production in Belgium
• buying ancillary services (primary, secondary
  tertiary reserves) to use flexible assets in
  real time (balancing market liberalized on 1st
  Jan. 06)
• using imports/exports (mutual solidarity in
  Europe coordinated by ETSO)
• transferring these balancing costs to the
  different market players (ARPs Access
  Responsible Party) which are responsible for this
  imbalance. This is done by ¼ hour.gt imbalance
  invoice

21
Impact of liberalization
Load( off-take)
Production( injection)
real time production adjustment
EBL
EBL
Imbalances
50 Hz
Prov. X
Prov. Y
Prov. Y
Prov. X
ARP (Access Responsible Party) balance
responsible
22
Impact of liberalization

• BUT
• the production is metered on a 1/4 hourly basis
• the load is metered on a 1/4 hourly basis only
  for the largest customers !

Elia grid
¼ hourly
dist. grid
¼ hourly
¼ hourly
yearly
yearly
¼ hourly
300 customers on Elia grid
8000 large clients
4.000.000 small clients
Load ( off-take)
Production ( injection)
23
Impact of liberalization
Elia grid
Distr. grid
¼ hourly
¼ hourly
¼ hourly
?
?
yearly
50 Hz
Production ( injection)
Load ( off-take)
24
Impact of liberalization
solution ?
25
Impact of liberalization

• 2 solutions
• either an AMR (Automated Meter Reading) for each
  customer (cf Italy Netherlands)
• either an allocation process to allocate the
  infeed (1/4 hourly metered) to the different
  market players
• Load profile EAV (Estimated Annual Volume)
  SLP (Synthetic Load Profile)
• Step 1 Install AMR devices in a representative
  sample of customers (0,1 of customers in
  Belgium)
• Step 2 Clustering customer load profiles gt
  Synthetic Load Profiles (SLPs)
• Step 3 Collect historical ¼ hourly load
  profiles
• Step 4 modeling these load profiles
• Step 5 Computation of model output for each
  customer
• Step 6 Correction to fit the infeed residue
  factor ( infeed S SLP)

26
Impact of liberalization load side
AMR
EBL
Provider X
Provider Y
total infeed - AMR
Distribution grid total infeed
Elia grid
27
Impact of liberalization clients without AMR
S11
S12
S21
S22
S11 small prof. S12 large prof. S21 small
resid. S22 large resid.
EBL
Provider X
Provider Y
total infeed - AMR
28
Impact of liberalization clients without AMR
S11
S12
S21
S22
S11 small prof. S12 large prof. S21 small
resid. S22 large resid.
EBL
Provider X
Provider Y
total infeed - AMR
Allocation ex-post estimation of the load
/provider, /SLP, /DGO ( municipality)
29
Impact of liberalization allocation
Key issue up-to-date access registry (for
each DGO)
quality ?
Up-to-date ?
30
Impact of liberalization SLP

• Following slides show key characteristics of
  SLPs for electricity
• S11 small professionals
• S12 large professionals
• S21 small residentials
• S22 large residentials.

31
Impact of liberalization SLP - yearly pattern
32
Impact of liberalization SLP - weekly pattern
33
Impact of liberalization SLP - daily pattern
34
Impact of liberalization SLP - daily pattern
35
Impact of liberalization allocation SLPs
Précision 2 s(résidus) / moyenne (Le facteur
2 correspond à environ 95 de confiance.)
36
Impact of liberalization allocation SLPs

• Compare to SLP model performances, residu level
  seems high. Possible explanations
• Is the panel representative ? (0,1 of the
  clients)
• Are SLPs correctly assigned ?
• Yearly consumption difference between last 2
  yearly metering gt can be 1 year old ! This is
  improved with reconciliation process (similar to
  allocation but with updated yearly consumptions).
  Exercise ended some weeks ago for the year 2004
  !

37
Impact of liberalization allocation SLPs
For forecasting purposes, marketing segments have
to be converted in SLPs
38
Impact of liberalization on forecasting

• Allocation process reality estimated after 1
  month (at least !). On 27th February 2007, the
  allocation for December is still not known !

Models (SLPs) residue
forecast
Representative (?) panel
3 months blind
estimated reality (allocation)
30/06/04
27/02/07
01/01/06
31/10/05
30/11/06
39
Impact of liberalization on forecasting
Before liberalization process We know we are
long
Since liberalization process We know we were
long one month ago !
40
Impact of liberalization on forecasting

• Solution to use real-time and pseudo-real
  time data at market level and use the
  correlation with our load
• Elia load
• Infeed-AMR (new project)
• Main goals of forecasting today
• To reduce our imbalance (contributing to network
  stability)
• To optimize our assets (merit order)
• gt to reduce error to anticipate (ex. moteur
  de réglage)

41
Agenda

• Electrabel key figures
• Load forecast why ?
• Impact of liberalization
• Factors impacting the load
• Forecasting methods
• Modelling techniques
• From load to price
• Key challenges for the future

42
Factors impacting the load

• Explanatory variables
• LT MT (till month ahead)
• Market share (or churn) Marketing campaigns
• Macro-economical factors
• Organic growth
• Arbitrage between fuel costs
• Sales of air-co, etc.
• New tarifs, 5 min without light, etc.
• Calendar data
• Day-of-week, days-off, bridges holidays
• Sunrise sunset hours
• ST (after week-ahead)
• Weather data temperature, cloud cover,
  radiation, wind speed direction,
• Special events strikes,

M -1
W -1
D
Y -1
LT MT
ST
43
Factors impacting the load

• Most difficult forecasts are due to
• strikes
• Christmas holidays (depends on day of the week of
  Christmas day)
• Extreme weather
• Highly increasing sales of airco.
• 1st Feb. 2007 no light from 1955 till 2000
• Market transitions off-peak tariff during
  week-end, etc.
• Sales forecast in number of customers per
  marketing segments gt to be converted in volumes
  per SLP (and accounting/billing view different
  than consumption view !)

44
Factors impacting the load yearly profile (elec)
1800 MW
45
Factors impacting the load Elia load
Day(2 years)
Hour of the week
46
Factors impacting the load yearly pattern (gas)
47
Factors impacting the load temperature
48
Factors impacting the load temperature
49
Factors impacting the load cloud cover
50
Factors impacting the load precipitations
51
Factors impacting the load precipitations
52
Factors impacting the load t (U.S. example)
Week days
Week-end
impact of inertia !
53
Factors impacting the load weather

• 2 similar days with the same temperature

54
Agenda

• Electrabel key figures
• Load forecast why ?
• Impact of liberalization
• Factors impacting the load
• Forecasting methods
• Modelling techniques
• From load to price
• Key challenges for the future

55
Forecasting method

• Top-down vs bottom-up
• Proxy days vs modeling

56
Forecasting method T-D vs B-U
Top Down vs. Bottom Up

• Top Down one model for whole portfolio
• EAV profile (total load)
• Bottom Up one model per segment
• S EAVi profile (SLPi)

i
57
Forecasting method T-D vs B-U

• Top Down vs. Bottom Up
• Top Down
• Deduce our load from system load ( Elia load)
  using the churn
• total system load history of 8 years, smooth,
  exactly measured in real-time gt easier to
  forecast
• - Getting worse overtime with customer churn
• Bottom Up
• forecasting different customer segments
  separately to be aggregated
• to account for customer churn in the individual
  segments
• Use different models for AMR SLPs
• - Limited history of segment load to be used for
  model calibration due to changing segments
  definition data warehouses

58
Forecasting method proxy days

• Proxy days
• either previous days
• either total Belgian load
• either estimation of our load
• either similar days in history
• limited to existing combinations of factors
• limited to a reduced number of risk factors
• difficult to consider market transitions

59
Forecasting method proxy days
60
Agenda

• Electrabel key figures
• Load forecast why ?
• Impact of liberalization
• Factors impacting the load
• Forecasting methods
• Modelling techniques
• From load to price
• Key challenges for the future

61
Modeling techniques

• 6 stages
• Data collection preprocessing
• Selection of explanatory variables
• Technique selection linear regression, ANNs,
  SARIMA, etc.
• Error function selection (objective function to
  be minimized MSE, MAE, MAPE, etc.)
• Calibration with history (back-testing) avoiding
  local optima !
• Performance assessment (based on an independent
  validation data set)

62
Modeling techniques

• Preprocessing
• Data Integrity (Bad data?)
• Stationarity (transitions ?)
• Do not estimate what you already know
  (deterministic pattern) but be careful mindless
  preprocessing can remove vital information or add
  wrong information !
• Multiple techniques
• Linear regression (with non-linear variable
  transformations) safe, easy to use and to
  interpret
• Auto-regressive models (SARIMA)
• Artificial Neural networks (ANNs)
• etc.

63
Modeling techniques

• Linear regression
• Autoregressive model ARMA (p,q)

64
Modelling techniques

• Generic formulation of autoregressive models
  SARIMA(p,d,q)(P,D,Q)
• S Seasonal
• AR Autoregressive (p autoregressive terms)
• I differencing/integration (d differences to
  achieve stationarity)
• MA Moving Average (q moving average terms)
• Fp (B)(1- B)d Zt d Tq (B)et

65
Modeling techniques

• Neural network

66
Modeling techniques

• Subdivision of dataset
• Training dataset vs test dataset
• Diagnostic tests
• Are all selected variables relevant ?

67
Modelling techniques

• The load is made up of a
• a deterministic part yearly, weekly daily
  patterns
• a random part
• Key issues
• to isolate the deterministic part (no more
  deterministic pattern in the random part)
• additive vs multiplicative or combined model
• Y(t) W(t) D(t) R(t)
• Y(t) W(t) D(t) R(t)

68
Modelling techniques

• Example of split between deterministic and
  stochastic components temperature

69
Agenda

• Electrabel key figures
• Load forecast why ?
• Impact of liberalization
• Factors impacting the load
• Forecasting methods
• Modelling techniques
• From load to price
• Key issues for the future

70
From load to price
71
From load to price
72
From load to price
73
From load to price

• 38 /MWh to 55 /MWh for a yearly base forward in
  6 months !!!

74
From load to price

• To forecast wholesale market prices load is a
  key driver
• To budget costs/revenues and check actuals
  (backcasting)
• The liberalization also imposes to better know
  the actual costs per segment / large customers
  (use of clustering techniques).
• Financial risk analysis
• Each ST load adjustment increases the costs, each
  LT/MT load adjustment increases the financial
  risks gt load volatility to be converted in a
  risk premium volume risk implies market risk
  with highly volatile market prices (ex. impact
  of credit risk).
• Impact of different load scenarios extreme
  weather, client bankruptcy, economical growth,
  etc. use of weather derivatives
• Portfolio effect diversification in Europe
  (exple impact of a cold wave in Europe, )

75
Agenda

• Electrabel key figures
• Load forecast why ?
• Impact of liberalization
• Factors impacting the load
• Forecasting methods
• Modelling techniques
• From load to price
• Key challenges for the future

76
Load Key challenges for the future

• Increasing randomness in production with
  renewable energies (hydro, wind sun)
• To contribute to improve allocation modelling,
  SLP assignment clustering
• Use of an on-line AMR sample for residential
  customers ?
• Use of auto-regressive models for large customers
  with on-line AMR?
• Optimize the use of newly pseudo real-time data
  available at market level (infeed-AMR).

77
Load Key challenges for the future

• Generalize probabilistic forecast associate
  confidence intervals depending on
• Day of the week, season, day-off vs normal day
• Weather forecast confidence level
• Etc.
• Refine the forecast at DGO level.
• Manage large database (i.e. 3 years of ¼ hourly
  historical data for 8000 AMR customers
  800.000.000 values)

78
Further infos
www.electrabel.com www.suez.com
www.elia.be www.rte-france.com www.etso-net.org
79
Contact
Olivier Ducarme Energy forecasting
manager Electrabel Rue Souveraine - Office S
406 1050 Brussels e-mail olivier.ducarme_at_electra
bel.com Tel [phone] Fax 32
(0)2 518 64 59 Mobile [phone]
80
Questions ?
81
Back-up slides
82
Liberalization of balancing market
Imbalance market
End-user market
Primary reserve (VFR  very fast reserve)
Belgian Imbal.
ImbalARP n
Electrabel
Elia NRV

Secondary reserve
ImbalARP 2
Tertiary reserve (probids)
Bal. player 2
ImbalEBL
Planned imports/exports
Bal. player n
Unplanned imports/exports
Net Regulation Volume
83
Operational forecast the global picture
The later, the more expensive !