APPLICATION OF NON-LINEAR TIME SERIES ANALYSIS TECHNIQUES TO THE NORDIC SPOT ELECTRICITY
MARKET DATA
Fernanda Strozzi, Eugénio Gutiérrez Tenrreiro
and José-Manuel Zaldívar Comenges
Update on WP5:Task 5.2
Deliverable 5.3
Hourly spot prices in the Nordic electricity market (Nord Pool)
from May 1992 until December 1998. (NOK/MWh)Hourly spot prices in the Nordic electricity market (Nord Pool)
from January 1997 until January 2007. (EUR/MWh)
SwedenNord Pool
Finland W Denmark E Denmark KontekNorway
Statnett Market
DATA
1996- EU Electricity Directive starts to have impact: EU countries open their electricity markets to competition ( high consumers can choose their provider).1993- Nord Pool (Nordic Electricity Market) was created by Norway.……2005-KT area. (Kontek cable connection Zealand-Germany). A competition starts between Nord Pool and European Energy Exchange (EEX)
Electricity Price
• Participants trade power contract one day for each hours of the next day• At the end of the trade all possible congestions or insufficient capacities are checked.• If some congestion or insufficient capacities occurs with these flows the market system established different “prices area” and TSOs ask to generators to increase (reduce) production or to buyer to increase (decrease) demand.• Then sometimes prices are the price of all Nordic region (“system price”) sometimes different “areas prices” exist.
• In this work we will consider “System price”: the price if no transmission constrains are considered
• Nordic Market has a single price less than half of the time (due to congestions)
Electricity Price: dependencies
• The variation of the prices in the Nord Pool system is well correlated with the variations in precipitations because of its dependence from hydropower generation.• In the “dry” periods the price and its volatility increase due to the dependence from other source of energy (petrol)
Electricity Price
Goals
1. Characterizing time series spot price dynamicscomparing them with- Linear Gaussian dynamic with the same fft testing nonlinearity- Time series with the same probability distribution but not temporally correlated (shuffled)2. Detect important events using nonlinear time series analysis
Data treatment
)(
)(ln)(
ttx
txtr tHourly logarithmic return
Hurst exponent: H
H is a tool for studying long-term memory and fractality of a time series
A long memory process is a process with a random component, where a past event has a slow decaying effect on future events.
H provides a measure of whether the data are a pure random walk or have some trend (i.e. some degree of correlation exists)
Hurst exponent: R/S analysis
Hurst measures how the range of cumulative deviations from the mean of the series is changing with the time
n
tn
ntnt
n rtrn
ntXntX
S
R
1
2
11
))((1
),(min),(max
(R/S)0 is a constant, n is the number of years to calculate the mean value H is the Hurst exponent
t
in
rirntX1
))((),(.
(R/S)n→(R/S)0 nH
A variety of techniques exist for estimating H but the accuracy of estimation is a complicate issue
as n →log2(R/S)n
log2n
Hurst exponent: properties
0 H 1 H=0.5 for random walk time series
fractional Brownian motion (fBm) is a random walk with H≠0.5 H < 0.5 for anticorrelated time seriesH > 0.5 for positively correlated series
.
Data set H
NOK 0.4406
EUR 0.2673
Power Spectral Density
deR f
fPSD2
)( )(
PSD describes how the power of the signal is distributed with frequency
The PSD is the Fourier transform of the autocorrelation function R() of the signal s(t) if the signal con be treated as a stationary random process
Data set
NOK 1.4612
EUR 1.4562
f
1fPSD )(
Stable Distribution
0,2
1 ,2/1
0 ,1
A stable probability distribution is defined by the Fourier transform of its characteristic function t
:
dtetxf itx
2
1),,,;(
)sgn(1||exp tititt
1 t
1
)log()/2(
)2/tan(
Gaussian22 2
Chauchy
Levy
(0,2], [-1,1] [0,) (-,)
Stable Distribution
Stable distribution
Data set
NOK 0.412 -0.365 0.035 -0.00018
NOK(0) 1.116 0.127 0.242 -0.0514
EUR 1.308 0.164 0.268 -0.068
EUR(0) 1.315 0.173 0.272 -0.069
Nord Pool data fitted parameters using STABLE (Nolan, 1999).
Fitted density plot for the Nord Pool Norwegian Krone and Euro time series data (blue line): a/Original time series, first difference; b/ without zero values (23962 values).
Surrogate datak S..., S ,1
Surrogate data is an ensemble of data sets similar (same mean, variance, etc) to the observed data and consistent with some null hypothesis
A discriminating statistic Q is chosen
If only a single realization exists for the original data we cannot compare the distributions of Q then we define a level of “significance”
If the null hypothesis is not true for k=19 surrogate data sets we can reject it with a probability of =0.05 (95% level of significance)
Surrogate data: Null hypothesis
1. Temporally uncorrelated noise. The null hypothesis: any correlation at all. Surrogate data are generated by a random shuffling of the original time series.
2. Linearly correlated noise. The null hypothesis : the time series are originated by a linear random process with the same autocorrelation function or, equivalently, with the same Fourier Power Spectrum.
t
q
kktkt
exax
1
et is un uncorrelated Gaussian noise of unit varianceis chosen so that the variance of the surrogates matches with the one of original dataak contain information on Correlation function
Surrogate data: discriminating statistic Q
The dynamic is chaotic? Q = Correlation dimensionQ = Lyapunov exponentQ = Forecasting error
The structures of recurrence plots?Q = RQA measures
State Space reconstruction
),...,,( 21 nxxxx ),( xFx
dt
d
),...,,( 21 Edyyyy ),( yGy
dt
d
,.../,/, 21
211 dtxddtdxxy
)(),.....2(),(),( 1111 Edtxtxtxtx ydE=embedding dimension
=time delay
Embedding Parameters
First minimum of the Average Mutual Information
Tnn nxPnxP
nxnxPnxnxPI
,2 ))(())((
)(),((log))(),(()(
NOK/MWh time series occurs at =15 EUR/MWh time series for =13.
=time delay
dE=embedding dimension
NOK/MWh dE=10 EUR/MWH dE=10
E1=relative increment of the mean distance between nearest points
E2=another way to measure relative
increment .
Cao (1997)
Stationarity test: space time separation plot
Sn=Random walkRn=Gaussian Random Number
Stp=Contour plot of Probability Density Function in the space: separation in space versus separation in time in the reconstructed space
Stationary = flat profile
Separation in time
Random walk
The reconstruction parameters do not change in the time if the series is stationary
Provenzale et al. (1992)
Stationarity test: space time separation plot
Space-time separation plot of the Nord Pool spot prices (NOK/MWh).
Space-time separation plot of the Nord Pool spot prices (EUR/MWh).
Space-time separation plot of Australian-US dollar foreign exchange time series.
Space-time separation plot Belgium Franc-US dollar foreign exchange time series.
Stationarity test: space time separation plot
Recurrence Plots: Introduction
N1,...,ji, xx
xx
ji
ji
ji
:0
,:1,R
N
i
ix1
Recurrence: Henri Poincaré 1890. when a system recurs many times as close as one wish to its initial state
Recurrence Plots (RPs): Eckmann, 1987.A trajectory of a system in its phase space
can produce a RP
i.e. a matrix given by:
Real data
Recurrence Plots
EUR/KMh =13, dE =10, =10 NOK/MWh =15, dE=10, =40
Recurrence Plots: Surrogate linearly correlated
EUR/KMh =13, dE =10, =10 NOK/MWh =15, dE=10, =40
Recurrence Plots: Surrogate Temporally uncorrelated
EUR/KMh =13, dE =10, =10 NOK/MWh =15, dE=10, =40
Recurrence Quantification Analysis
a/ Measures based on recurrence density
N
jijiN
RR1,
,2)(
1)( R
where )(, jiR is one if the state of the system at time i and the one at time j have a distance less than
and zero otherwise.
b/ Measures based on diagonal lines
N
l
N
ll
llP
llP
DET
1
)(
)(min%determinism (DET) is then the percentage of recurrent points forming diagonal line structures
lmin=100
%recurrence
lN
iilL 1max max max
1
LDIV
the length maxL of the longest diagonal line found in the RP, or its inverse, the divergence (DIV)
N
ll
lplpENTRmin
)(ln)(
Shannon entropy of the probability lNlPlp /)()( to find a diagonal line of length l in RP.
Recurrence Quantification Analysis
c/ Measures based on vertical lines
% vertical lines of length in RP
N
v
N
vv
vvP
vvP
LAM
1
)(
)(min
The average length of vertical structures is given by
N
vv
N
vv
vP
vvP
TT
min
min
)(
)(
Data set %recur %deter maxline entropy trend % laminar TrapTime
Bnok 16.095 67.13 3545 8.593 -8.687 69.994 308.044
Surr001 8.150 6.129 4808 6.740 2.306 1.796 123.511
Surr002 1.926 4.521 1355 4.913 -0.142 0.000 -1
Surr003 2.807 8.026 4808 6.028 -1.616 0.000 -1
Surr004 30.218 36.309 4808 7.994 -3.360 35.521 214.805
Surr005 1.735 13.216 1844 6.117 -0.983 0.055 110
Surr006 1.007 32.018 1178 6.287 -0.752 16.980 166.134
Surr007 4.785 13.279 2674 6.895 -0.802 7.533 153.511
Surr008 14.122 17.880 4350 7.357 -4.479 9.293 154.815
Surr009 5.934 13.528 3130 7.195 -2.458 6.301 159.498
Surr010 1.193 5.900 1064 4.696 -0.677 0.347 119.500
Surr011 4.860 51.638 4808 7.918 -1.542 52.444 266.162
Surr012 31.899 52.675 4808 8.415 12.407 54.522 218.168
Surr013 4.795 9.417 4808 6.880 0.524 0.724 143.393
Surr014 5.725 9.169 4154 6.783 -2.980 1.774 144.963
Surr015 4.972 6.340 2370 6.606 -2.341 1.720 114.988
Surr016 18.050 23.404 4808 7.678 -4.183 12.845 161.470
Surr017 10.846 43.188 4614 8.799 -7.222 38.298 338.899
Surr018 4.956 8.523 4808 6.596 -2.654 3.484 141.553
Surr019 6.323 4.462 4808 6.184 -1.989 0.375 114.208
Surrogate: linear surrogate
Data set %recur %deter maxline entropy trend % laminar TrapTime
Beur 7.12 35.33 2094 7.658 -4.587 33.94 263.525
Surr001 12.524 3.665 3340 6.355 -6.259 2.539 149.367
Surr002 1.643 5.894 2238 5.270 -1.100 1.872 119.367
Surr003 3.840 1.397 2150 4.533 -0.998 0.000 -1
Surr004 4.377 1.105 1324 3.970 -0.286 0.000 -1.000
Surr005 10.677 1.825 4187 5.730 -5.483 1.527 126.613
Surr006 8.658 18.813 4826 7.538 -5.638 9.854 146.364
Surr007 0.491 3.888 690 2.807 -0.346 0.000 -1.000
Surr008 23.790 11.105 4826 7.509 -7.639 9.252 162.159
Surr009 30.269 10.831 4826 7.393 -1.830 7.108 151.053
Surr010 20.536 4.700 4826 6.845 -7.466 6.416 150.611
Surr011 2.336 3.777 1888 5.094 -1.160 1.529 134.161
Surr012 3.715 1.475 3517 4.059 -1.627 0.108 117.250
Surr013 4.994 3.736 3721 5.972 -3.343 1.886 135.457
Surr014 21.649 9.020 4826 7.162 -2.900 9.664 154.810
Surr015 20.052 8.142 2669 7.247 -4.243 4.171 146.484
Surr016 6.811 5.384 3998 6.574 -4.098 0.758 125.312
Surr017 3.161 4.113 1964 5.641 -2.076 0.715 131.650
Surr018 7.809 3.369 2429 6.204 -0.473 2.766 132.437
Surr019 12.185 1.330 4826 5.429 1.503 0.088 125.600
Surrogate: linear surrogate
Data set %recur %deter maxline entropy trend % laminar TrapTime
Bnok 16.095 67.13 3545 8.593 -8.687 69.687 308.044
Surr001 0 0 -1 -1 0 0 -1
Surr002 0 0 -1 -1 0 0 -1
Surr003 0 0 -1 -1 0 0 -1
Surr004
Surr005
Surr006
Surr007
Surr008
Surr009
Surr010
Surr011
Surr012
Surr013
Surr014
Surr015
Surr016
Surr017
Surr018
Surr019
Surrogate: temporally uncorrelated
RQE analysis
RQE analysis is able to detect important changes?
RQE=RQA quantities on moving windows
Norway SwedenFinland W Denmark E Denmark Kontek
NOK/MWh EUR/MWh
720 point window (one month), data are shifted 720 points
Volatility
Inverse of standard deviation and %determinism (top) and %laminarity (bottom) for NOK/MWh
Inverse of standard deviation and %determinism (top) and %laminarity (bottom) for EUR/MWh
Volatility: EUR/MWh
RQA measures of EUR/MWh: Values are computed from a 720 point window (one month), shifted of 720 points. RQA parameters: =13, dE=10, distance cutoff: max. distance between points/10, line
definition: 100 points (~4 days). Vertical lines correspond to the following dates: 1 st October 2000, 5th October 2005 (see historical background).
Volatility: NOK/MWh
Nonlinear metrics of the Nord Pool spot prices time series in NOK/MWh: Values are computed from a 720 point window (one month), data are shifted 720 points. RQA parameters: t =15, dE=10, distance cutoff: max. distance between points/10, line definition: 100 points (~4 days). Vertical lines correspond to the following dates: 1st January 1993, 1st January 1996, 29th December 1997 and 1st July 1999 (see historical background).
Conclusions and Future Developments
• H < 0.5 anticorrelation
RQE: (RQA measures on moving windows) are able to detect changesRQE→%determ, %Lam give a new measure of volatility
New Developments: analyse the possible correlation between the volatility of energy market and blackouts
Characterize the underlying dynamics using surrogates:
• Space Time Separation Plot: the data seems stationary •RQA Measures (%determ, %Lam) are able to distinguish between data: - surrogate linearly correlated - temporally uncorrelated
Detect important changes in the time series not evident from their time representation
Measuring long term correlation of the time series:
• Stable distribution
Bibliography
Data History:
Amundsen, E.S. and Bergman, L., 2007, Integration of multiple national markets for electricity: The case of Norway and Sweden, Energy Policy, doi.1016/j.enpol.2006.12.014.
Byström H.N. E. 2005. Extreme value theory and extremely large electricity price changes. International Review of Economics and Finance 14, 41-55.
Hsieh, D. A., Chaos and nonlinear dynamics: Application to financial markets, 1991, The Journal of Finance 46, 1839-1887Kristiansen, T., 2007. Pricing of monthly contracts in the Nord Pool market. Energy Policy 35, 307-316.
Kristiansen, T., 2006, A preliminary assessment of the market coupling arrangement on the Kontek cable, Energy Policy (in press)
Perelló, J., Montero, M., Palatella, L., Simonsen, I. and Masoliver, J., 2007. Entropy of the Nordic electricity market:anomalous scaling, spikes, and mean-reversion. J. Stat. Physics (in press).
Vehviläinen I. and Pyykkönen, T. 2005. Stochastic factor model for electricity spot price-the case of the Nordic market. Energy Economics 27, 351-357.
Weron, R., Bierbrauer, M. and Truck, S. 2004. Modelling electricity prices: Jump diffusion and regime switching. Physica A 336, 39-48.
Hurst:Cannon, M. J. Percival, D. B., Caccia, D. C., Raymond, G. M. and Bassingthwaighte, J. B., 1997, Evaluating scaled windowedvariance methods for estimating the Hurst coefficient of time serie
Haldrup, N., Nielsen, M. Ø., 2006. A regime switching long memory model for electricity prices. Journal of econometrics 135,349-376.
Hurst, H. E., 1951, Long-term storage capacity of reservoirs, Trans. Am. Soc. Civ. Eng. 116, 770-779.
Simonsen, I., 2003. Measuring anti-correlations in the Nordic electricity spot market by wavelets. Physica A 322, 597-606.Weron, R. and Przybyłowicz, B., 2000. Hurst analysis of electricity price dynamics. Physica A 283, 462-468.
Embedding parametersCao, L., 1997, Practical method for determining the minimum embedding dimension of a scalar time series, Physica D 110, 43-50.
Bibliography
Surrogate data:
Schreiber, T. and Schmitz, A., 2000, Surrogate time series, Physica D 142, 346-382.
Theiler, J., Eubank, S., Longtin, A., Galdrikian, B., and Farmer, J. D., 1992, Testing for nonlinearity in time series: the method of surrogate data. Physica D 58, 77-.
Space time separation plot:Hegger, R., Kantz, H., Schreiber, T., 1999, Practical implementation of nonlinear time series methods: The TISEAN package. CHAOS 9, 413-. The software package is publicly available at http://www.mpipks-dresden.mpg.de/~tisean .
Provenzale, A., Smith, L. A., Vio, R. and Murante, G., 1992, Distinguishing between low-dimensional dynamics and randomness in measured time series. Physica D 58, 31-49.
Stable distributions:
Nolan, J.P., 1999. Fitting data and assessing goodness of fit with stable distributions. In Proceedings of the Conference on Applications of Heavy Tailed Distributions in Economics, Engineering and Statistics , American University, Washington DC, June 3-5.
Nolan, J.P., 1997. Numerical computation of stable densities and distribution functions. Commun. Stat.: Stochastic models 13, 759-774.
PSD
Theiler, J., 1991, Some comments on the correlation dimension of noise. Phys. Lett A 155, 480-493.f/1
Recurrence plot:
Eckmann, J. P., Kamphorst, S. O. and Ruelle, D., 1987, Recurrence plots of dynamical systems, Europhys. Lett. 4, 973-977.
Marwan, N., Romano, M. C., Thiel, M. and Kurths, J., 2007. Recurrence plots for the analysis of complex systems. Physics Reports 438, 237-329.
Strozzi, F., Zaldivar, J, M., Zbilut, J., P., 2007, Recurrence quantification analysis and state space divergence reconstruction for financial time series analysis. Physica A 376, 487-499
Strozzi, F., Zaldívar, J. M., & Zbilut, J. P., 2002. Application of nonlinear time series analysis techniques to high frequency currency exchange data, Physica A 312, 520-538.
Zbilut, J. P. and Webber Jr. C. L., 1992, Embeddings and delays as derived from quantification of recurrence plots. Phys. Lett. A 171, 199-203
Bibliography