Upload
doankhanh
View
226
Download
2
Embed Size (px)
Citation preview
Forecasting electricity spot prices usingtime-series models with a double temporal
segmentation
Marie Bessec∗, Julien Fouquau∗∗, Sophie Meritet∗
∗LEDa-CGEMP, University Paris Dauphine∗∗NEOMA Business School
IAEE 37th International Conference - New York 2014
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Aim of the paper: short-run forecast of day-aheadelectricity spot prices in France
Features of electricity prices⊲ Seasonality⊲ Sudden and fast-reverting price spikes
Modeling and forecasting procedure⊲ Double temporal segmentation⊲ Non-linear models: MS and STR models⊲ Forecast combinations
Extensive evaluation: 2880 models, 1728 combinations
Results of the out-of-sample evaluation on French data⊲ Considering each season separately improves the results⊲ Non-linear models leads to better forecasts⊲ Pooling results provide more reliable forecasts
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Related literature
Table: Related papers on electricity price forecasting
Authors Market
Combinations Bordignon et al. (2013) UKPXNowotarski et al. (2013) EEX, NP, PJM
Non-linear models Karakatsani and Bunn (2008) UKPXMisiorek et al. (2006) CaliforniaWeron and Misiorek (2008) California and NPKosater and Mosler (2006) EEX
Seasonality Nowotarski et al. (2013) NSW, EEX, ISONP, NYISO, PJM
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Spikes and drops
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Seasonality
0
10
20
30
40
50
60
70
80
90
100
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24Hour
Pric
e (�
/MW
h)
0
10
20
30
40
50
60
70
80
90
100
1 2 3 4Season
Pric
e (�
/MW
h)
intraday cycle intrayear cycle
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Seasonality
High thermal sensitivity in France: 35% of the total stock of housingheated with electricity (RTE, 2012)
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
DataTemporal segmentationLinear versus non-linear modelsIndividual versus pooled forecasts
Dependent variable: Hourly day-ahead prices (e/MWh) in EPEX
Explanatory variables (available before the day-ahead auction)
Demand forecast (in MW): available for each trading hour at 0:00in t-1,
Capacity margin (in MW): available at 8 p.m. in t-1,
Volatility of spot prices (in e/MWh): the coefficient of variation ofthe hourly prices over the last 5 days,
Past values of spot prices (in e/MWh),
Gas price (in e/MWh): the Dutch day-ahead price (TTF); toavoid endogeneity problem, we use lag-1 price,
Forecasted balance of exchange programs with Germany (inMW): available for each hour at the end of t-1.
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
DataTemporal segmentationLinear versus non-linear modelsIndividual versus pooled forecasts
Temporal segmentation
Significant variation of coefficients of the regressors acrosstrading hours and seasons
Figure: Coefficients of the drivers in linear modelspt = α + φ1pt−1 + φ2pt−2 + β1Capat−1 + β2Demandt + β3Volatt−1 + β4Gast−1 + β5exchgt−1 + εt
2 4 6 8 10 12 14 16 18 20 22 24−2
−1
0
1
2
3intercept
2 4 6 8 10 12 14 16 18 20 22 240
0.5
1
1.5price(−1)
2 4 6 8 10 12 14 16 18 20 22 24−0.5
0
0.5
1price(−2)
2 4 6 8 10 12 14 16 18 20 22 24−0.2
−0.1
0
0.1
0.2margin(−1)
2 4 6 8 10 12 14 16 18 20 22 240
0.1
0.2
0.3
0.4first−differenced demand
2 4 6 8 10 12 14 16 18 20 22 24−10
−5
0
5price volatility
2 4 6 8 10 12 14 16 18 20 22 24−3
−2
−1
0
1first−differenced gas price(−1)
2 4 6 8 10 12 14 16 18 20 22 24−0.04
−0.03
−0.02
−0.01
0
0.01exchange with DE(−1)
WinterSpringSummerFallNon seasonal
⇒ Double temporal segmentation: each hour of the day and eachseason of the year is considered separately.
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
DataTemporal segmentationLinear versus non-linear modelsIndividual versus pooled forecasts
Non linear models
Markov-switching models
pt = α(St ) + β′(St )xt + εt(St )
where εt (St ) → NID(0, σ2(St )), St = 1, 2, . . . ,M a first-order Markovchain.
Smooth Threshold Regressive models
pt = α0 + β′
0xt + (α1 + β′
1 xt ) G(qt ; γ, c) + εt
with εt i.i.d .(
0, σ2ε
)
and:
G(qt ; γ, c) = [1 + exp(−γ(qt − c))]−1 , γ > 0
Linear model pt = α+ β′xt + εt , t = 1, . . . ,T
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
DataTemporal segmentationLinear versus non-linear modelsIndividual versus pooled forecasts
Individual Models Code
Autoregressive model ARExponential model EXPOAR model augmented with forecasted demand AR-XLinear regression with regressors selected with a stepwise procedure LSTEPMS model, 2 regimes, FTP, variable intercept and variance MS 1MS model, 2 regimes, FTP, variable intercept, AR coefficients and var. MS 2MS model, 2 regimes, FTP, variation of coefficients MS 3MS model, 2 regimes, TVTP, variable intercept and variance MSV 1MS model, 2 regimes, TVTP, variable intercept, AR coefficients and var. MSV 2MS model, 2 regimes, TVTP, variation of all coefficients MSV 3MS model, 3 regimes, FTP, variation of intercept and variance MS3 1MS model, 3 regimes, FTP, variable intercept, AR coefficients and var. MS3 2STR model, variable intercept STR 1STR model, variable intercept and AR coefficients STR 2STR model, variation of all coefficients STR 3
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
DataTemporal segmentationLinear versus non-linear modelsIndividual versus pooled forecasts
Forecast combinations
Combination of the individual forecasts pct =
∑Kk=1 ωt,k p(k)
twith alternative weights for k = 1, . . . ,K :
ωt,k = 1K (simple average) [C1]
ωt,k =(∑t−1
τ=t−l e2τ,k )
−1
∑Kj=1(
∑t−1τ=t−l e2
τ,j)−1
[C2]
ωt,k =v−1/2
t−1,k exp[−e2
t−1,k2vt−1,k
]ωt−1,k
∑Ki=1 v−1/2
t−1,i exp[−e2
t−1,i2vt−1,i
]ωt−1,i
[C3]
where eτ,k = pτ − p(k)τ , ω1,k = 1/K and vt−1,k = 1
t−1
∑t−1τ=1 e2
τ,k .
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Empirical designOut-of-sample results
Real time evaluation: rolling forecasts of the last 35observations of each season in 2012
Evaluation of the forecast accuracy:
⊲ RMSE =√
1N
∑Nt=1 et (1)2
⊲ MAE = 1N
∑Nt=1 |et (1)|
⊲ MAPE = 1N
∑Nt=1 |
et (1)pt+1
|
Tests for comparing predictive accuracy⊲ Diebold-Mariano test, Giacomini-White test⊲ Encompassing test
Extensive evaluation: 15 individual models, 9 combinationsover each trading hour h = 1, . . . ,24, 4 seasons
2880 individual models and 1728 combinations
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Empirical designOut-of-sample results
Figure: Estimation and forecasting windows
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Empirical designOut-of-sample results
Temporal segmentation
Comparison of seasonal and non-seasonal models (in %)Winter Spring Summer Fall
% of times the seasonal model yields a lower MSE or MAE than the non-seasonal model
Seasonal beats non seasonal 69.1 77.6 30.4 54.5Seasonal beats non seasonal (AR & EXPO) 95.8 77.1 52.1 70.8Seasonal beats non seasonal (linear models) 89.6 80.9 36.1 64.6
Significance of difference with the Diebold-Mariano test
Seasonal beats non seasonal 32.4 42.7 13.4 24.5Seasonal beats non seasonal (AR & EXPO) 50.0 51.0 20.8 41.7Seasonal beats non seasonal (Linear models) 46.9 50.4 16.7 33.0
Encompassing test (Harvey, Leybourne and Newbold, 1997)
H0: NS encompasses SA 34.1 66.5 31.3 60.7H0: NS encompasses SA (AR & EXPO) 52.1 56.3 33.3 50.0H0: NS encompasses SA (Linear models) 45.8 66.0 27.8 54.9
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Empirical designOut-of-sample results
Linear versus non-linear models
Comparison of linear and non-linear models (in %)Winter Spring Summer Fall
SA NS SA NS SA NS SA NS
% of times the non-linear model yields a lower MSE or MAE than the linear model
Non-linear beats linear 31.8 36.1 72.6 67.7 42.0 58.5 55.4 57.3MS3 1 beats linear 41.7 54.9 77.8 77.8 55.6 77.8 69.4 79.9STR beats linear 22.0 35.4 66.9 65.5 42.1 42.8 44.0 33.1
Significance of difference with the Diebold-Mariano test
Non-linear beats linear 9.2 11.7 41.2 29.2 15.6 21.2 20.6 29.7MS3 1 beats linear 14.6 13.9 52.1 47.2 27.1 27.1 20.8 52.8STR beats linear 4.2 13.0 37.3 23.4 14.4 12.3 19.0 14.6
Encompassing test (Harvey, Leybourne and Newbold, 1997)
H0: Linear encompasses NL 34.7 40.0 71.6 68.2 61.2 64.4 57.2 61.9H0: Linear encompasses MS3 1 38.9 47.2 73.6 76.4 76.4 75.0 65.3 79.2H0: Linear encompasses STR 31.9 39.4 68.1 69.0 59.3 60.7 46.3 47.7
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Empirical designOut-of-sample results
Table: Best models - RMSE
Winter Spring Summer FallModel SA NS SA NS SA NS SA NS Total
Linear 13 7 2 2 12 7 5 4 52AR 9 6 1 1 6 4 2 1 30expo 4 0 1 1 6 1 3 3 19
AR-X 0 1 0 0 0 2 0 0 3MS 7 13 18 16 9 16 12 17 108MS 1 4 3 0 2 2 0 1 2 14MS 2 0 1 3 0 1 0 1 2 8MS 3 1 0 2 1 1 0 2 0 7MSV 1 0 1 1 3 2 2 2 0 11MSV 2 1 0 2 1 0 3 2 3 12MSV 3 0 3 2 0 1 0 1 0 7MS3 1 0 5 7 7 1 8 3 9 40MS3 2 1 0 1 2 1 3 0 1 9STR 4 4 4 6 3 1 7 3 32STR 1 3 0 3 0 3 1 4 3 17STR 2 1 2 1 1 0 0 1 0 6STR 3 0 2 0 5 0 0 2 0 9Total 24 24 24 24 24 24 24 24 192
(•) = number of times each model gives the best RMSE.
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Empirical designOut-of-sample results
Table: Rank of individual and pooled forecasts in term of RMSE
Winter Spring Summer Fall AllModel SA NS SA NS SA NS SA NS all
AR 6 8 23 23 12 20 23 20 22EXPO 3 6 22 22 10 23 22 21 19AR-X 15 19 10 19 21 14 7 14 14MS 1 10 12 19 17 14 13 13 6 12MS 2 21 20 10 16 20 16 16 9 16MS 3 20 23 6 18 16 15 9 18 15
MSV 1 11 14 16 7 11 8 17 10 11MSV 2 16 18 6 14 23 12 15 4 13MSV 3 19 16 8 20 22 18 14 16 17MS3 1 12 9 9 6 9 10 12 2 8MS3 2 17 22 14 21 18 17 20 12 18STR 1 18 21 12 13 17 21 19 19 20STR 2 22 15 15 12 15 19 18 22 21STR 3 23 17 21 11 18 22 21 23 23
C1 7 5 3 1 7 3 3 7 3C1 L 9 7 2 3 5 1 1 3 2
C1 NL 13 11 4 2 6 5 6 1 6C2 5 3 5 5 4 2 2 5 1
C2 L 2 2 17 9 1 6 5 15 5C2 NL 4 4 13 8 2 4 4 13 4
C3 8 13 18 10 8 11 8 8 10C3 L 1 1 20 15 3 9 10 17 7
C3 NL 14 10 1 4 12 7 11 11 9
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Empirical designOut-of-sample results
Table: Rank of individual and pooled forecasts in term of RMSE
Winter Spring Summer Fall AllModel SA NS SA NS SA NS SA NS all
AR 6 8 23 23 12 20 23 20 22EXPO 3 6 22 22 10 23 22 21 19AR-X 15 19 10 19 21 14 7 14 14MS 1 10 12 19 17 14 13 13 6 12MS 2 21 20 10 16 20 16 16 9 16MS 3 20 23 6 18 16 15 9 18 15
MSV 1 11 14 16 7 11 8 17 10 11MSV 2 16 18 6 14 23 12 15 4 13MSV 3 19 16 8 20 22 18 14 16 17MS3 1 12 9 9 6 9 10 12 2 8MS3 2 17 22 14 21 18 17 20 12 18STR 1 18 21 12 13 17 21 19 19 20STR 2 22 15 15 12 15 19 18 22 21STR 3 23 17 21 11 18 22 21 23 23
C1 7 5 3 1 7 3 3 7 3C1 L 9 7 2 3 5 1 1 3 2
C1 NL 13 11 4 2 6 5 6 1 6C2 5 3 5 5 4 2 2 5 1
C2 L 2 2 17 9 1 6 5 15 5C2 NL 4 4 13 8 2 4 4 13 4
C3 8 13 18 10 8 11 8 8 10C3 L 1 1 20 15 3 9 10 17 7
C3 NL 14 10 1 4 12 7 11 11 9
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Empirical designOut-of-sample results
Forecast combinations
Table: Comparison of individual and combined forecasts (in %)
Winter Spring Summer FallSA NS SA NS SA NS SA NS
% of times the pooled models yield a lower MSE or MAE than the individual model
Combined beats individual 67.3 70.4 61.9 70.9 74.6 73.7 71.9 66.6
Significance of difference with the Diebold-Mariano test
Combined beats individual 25.2 26.9 27.1 36.4 43.9 38.9 35.2 37.5
Encompassing test (Harvey, Leybourne and Newbold, 1997)
H0: indv encompasses comb 29.3 39.7 43.3 56.3 66.8 59.4 54.4 52.5
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices
Brief summaryFeatures of electricity prices
Modeling and forecasting proceduresDay-ahead forecasting
Conclusion
Main resultsThe double temporal segmentation improves theforecasting ability of the models.Non-linear models designed to capture the sudden andfast-reverting spikes in the price dynamics improve theforecast accuracy.Pooling forecasts gives more reliable results.
Possible extensionsEnlarge the comparison of the specifications with a largerset of regressors.Explore how revised weather forecasts available in themorning before the auction could add extra value beyondthe midnight release of demand forecasts.
M. Bessec, J. Fouquau, S. Meritet Forecasting electricity spot prices