The Effect of Wind and Solar Energy on Dutch Hourly
Average Electricity Spot Prices
David Ruhe (10385029)
Bachelor’s Thesis and Thesis Seminar Economics and Finance
(6013B0345)
Supervisor: Damiaan Cheng June 29th, 2015
Statement of Originality
All of the statements and work included in this portfolio is original and is entirely my work. Please keep the work in this portfolio confidential.
Abstract
During the last two decades sustainable energy resources have been on a constant growth path. Especially solar energy is going through an exponential growth phase. On the Dutch wholesale electricity spot market prices are completely dependent on supply and demand. In this study we tried to examine the effects of a change in weather conditions on the supply of electricity and hypothesised a change in price due to this change in price. Confirmed variables were wind speed, average hours of sunshine, night-time and weekend effects. Expected seasonal effects were not found. We also tried to quantify the effects of changing weather conditions on the volatility of this market. Confirmed predictors were prices, hours of sunshine, weekend, night-time and an expected winter effect. Expected summer and wind speed effects were not confirmed.
1. Introduction
On the Dutch electricity market different parties can participate. Examples are producers, distributors, traders and network operators. Network operators have a monopoly in their district and charge electricity consumers with a (by government) regulated tariff. Most people only know about network operators and the regulated tariff. However, on July 1st 2004 the market for energy was liberalized and since then consumers and firms can feed in or feed out power to the spot market. Most power is traded on the Amsterdam Power Exchange (APX). On this exchange, electricity for all hours of the next day is bought at a day-ahead auction. A day later spot prices will be known. On the spot market the price of electricity is completely
dependent on the effects of supply and demand. When network operators require more power to supply to their district, the spot price of a Megawatt-hour (MWh) of electricity will increase. When demand for electricity is high, prices will rise and more parties will supply more MWhs because of increased revenues during times of high energy prices. Because of the extra supply spot prices will drop and the market will return to equilibrium. Likewise, when demand suddenly is low, the grid can get overloaded and prices will drop to eliminate the excessive supply. The economy always requires a certain supply of power which can differ in magnitude because of certain circumstances. Previous research has shown that there are several systematic effects that influence spot prices. At the same time sudden market frictions like weather changes or power outages will have influence on the price of electricity. The Netherlands are an excellent country to study the effects of weather conditions because of its sizable electricity market relative to the small geographical surface. Moreover, weather conditions are roughly the same throughout the country. The Netherlands are striving to receive 14% of their total energy supply from sustainable resources in 2020. By striving for higher sustainable energy output a lower CO2 emission rate can be reached. Wind and solar power are the main resources that are aimed for. In 2014 the total sustainable energy supply accounted for 11.4% of the total energy supply.1 Using data from the Central Bureau of Statistics of the Netherlands we can plot the increase of renewable energy produced in the Netherlands. In figure 1a we observe that renewable energies are on a long-term growth path. Figure 1b shows us the tenfold increase of solar powered energy supply over the last 5 years. In spite of this boost solar energy only makes up 3% of total renewable energy supplied which itself is only 11.4% of the total energy supply. Energy supplied by natural resources will have an immediate impact on supply and demand curves because of the fact that electricity is hard to store. An extra surge of supply because of high wind velocities will have an immediate effect on supply of electricity and lower prices are
1 Centraal Bureau voor de Statistiek. 2015. Centraal Bureau voor de Statistiek. Accessed June 28, 2015. http://www.cbs.nl/.
presumably the result. In this research we want to study the effect of weather conditions on hourly average spot prices through an increase of energy supply. More specifically, our research question is: do the supply effects of wind velocity and solar power affect hourly average spot prices of the Dutch electricity market? We expect that a higher wind speed will increase supply and therefore prices will decrease. Likewise, during extended periods of sunshine we expect prices to decrease. After we set up our initial model we will look at an alternative model that tries to predict the same effect. We also look at the effect of changing weather conditions on the volatility of the spot prices to identify if uncertainty of the market is dependent on these extra energy supplies. Since network operators, small companies and even consumers can sell energy these days, these effects can also be of economic interest to a wider public. Because weather conditions are the only factors that drive solar-panel and wind turbine outputs the forecasting abilities of weather conditions on electricity prices will become more precise when more sustainable energy is produced. Therefore, this research could be an important little step to widen the scale of research in this field and keeping track of the magnitude of these effects. Part 2 covers the earlier research done in this field. Part 3 sets out the dataset we used for this paper. In part 4 we explain the methodology used for this research. In part 5 the empirical results are shown and in part 5 we draw some conclusions.
2. Literature Review
In this paper the effects of weather conditions on the prices of one MWh on the electricity spot market are studied. Because the market for electricity is just eleven years old the literature on it is limited compared to other markets. Moreover, most of the research that has been published is about day-ahead prices, not about spot markets. In 2001 Knittel and Roberts worked on an article about liberalizing power markets. They constructed one of the
first models for predicting power prices and confirm temperature as significant coefficient for prices. However, in their study it has little forecasting ability.2 After liberalizing electricity markets in parts of the United States and Europe research got more extended. In 2005 Knittel and Roberts extend their research and find strong time of the day, seasonal and temperature effects.3 Because electricity cannot be stored, extreme spikes can occur. Prices can increase tenfold in under fifteen minutes. Deng and Huisman & Mahieu early established Markov regime-switching models (MRS) for analysing spike behaviour which became the basis for research in energy markets.45 It is also commonly known that the market shows mean-reversing patterns. Therefore autoregressive conditional heteroscedasticity models (GARCH) that are modelled to analyse mean-reverting processes have been constructed.6 However, these models initially failed being very precise and Janczura and Weron find that an independent spike 3-regime model with time-varying transition probabilities, heteroscedastic diffusion-type base regime dynamics and shifted spike regime distributions are the best models for predicting electricity spot prices.7
Several effects of weather conditions have been confirmed. A working paper that studies the effects of wind and temperature on German electricity spot prices finds significant wind velocity and temperature effects (Kosater 2006).8 A study of electricity prices in Texas finds that an increase of wind generation reduces prices on the Texan electricity spot market (Woo,
2 Knittel, Christopher R., and Michael R. Roberts. 2001. "An Empirical Examination of Deregulated Electricity Prices." PWP-087.
3 Knittel, R. Christopher, and Michael R. Roberts. 2005. "An Empirical Examination of Restructured Energy Prices." Energy Economics 791–817.
4 Deng, Shijie. 1998. "Stochastic Models of Energy Commodity Prices and Their applications: Mean-reversion with Jumps and Spikes." University of California Energy Institute.
5 Huisman, Ronald, and Ronald J. Mahieu. 2003. "Regime Jumps in Electricity Prices." Energy Economics 425-434.
6 Escribano, Alvaro, Juan Ignacio Peña , and Pablo Villaplana . 2011. "Modelling Electricity Prices: International Evidence." Oxford Bulletin of Economics and Statistics 622-650.
7 Janczura, Joanna, and Rafal Weron. 2010. "An Empirical Comparison of Alternate Regime-switching Models for Electricity Spot Prices." Energy Economics 1059-073.
8 Kosater, Peter. 2006. "On the impact of weather on German hourly power prices." Discussion papers in
et al. 2011).9 Moreover, they find that an increase of wind velocity increases spot-price variance. An Australian study suggests that while demand continues to be the greatest influence on spot prices, wind speed is becoming a significant secondary influence (Cutler, et al. 2011).10 Cutler et al. also found that during winter and summer times volatility of spot prices increases. Research of this kind has not been performed regarding solar energy because of its very young nature. But since 2010 solar energy has been growing exponential in the Netherlands we expect to find a similar negative effect of solar power on energy prices.
3. Methodology
In this paper we first discuss the effects of wind speed and time of sunshine on the hourly average electricity spot prices. Second, we set up an alternative model that predicts the same effects. Third, we discuss the effects of weather variables on hourly variability of spot prices.
1. The Effect of Weather Conditions
In this research we do not incorporate GARCH models because heteroscedasticity is caused by spikes, but at the same time by periodic autocorrelations.11 Kosater looks at the logarithm of spot prices because they tend to follow a lognormal distribution. In this study we do not follow this methodology because of the possibility of spot prices being negative. Neither do we use a workaround in order to keep data trustworthy. The models that were tested by Janczura & Weron as best fitting are not usable for the scope of this research. Instead, we try to model the effect of weather conditions by setting up a dependent variable and using Ordinary Least Squares regression to estimate the effect of the independent variables. Let 𝑃𝑡 be the average
9 Woo, C. K., I. Horowitz, J. Moore, and A. Pacheco. 2011. "The impact of wind generation on the electricity spot-market price level and variance: The Texas experience." Energy Policy 3939-3944.
10 Cutler, Nicholas J., Nicholas D. Boerema, F. Macgill Iain, and Hugh R. Outhred. 2011. "High Penetration Wind Generation Impacts on Spot Prices in the Australian National Electricity Market." Energy Policy 5939-949. 11 Carnero, Angeles M., J. S. Koopman, and M. Ooms. 2003. "Periodic Heteroskedastic RegARFIMA models for daily electricity spot prices." Discussion Paper Tinbergen Institute Amsterdam.
spot price of hour 𝑡 with 𝑡 ∈ {1, … , T}. ß0 is an intercept. Following Kosater (2006) we include a lagging effect of the price of the hour before: 𝑃𝑡−1. Following Knittel and Roberts (2005) we add seasonal dummies. We also include weekend and night-time dummies. Finally we add the variables we want to observe.
Model 3.1
𝑃𝑡 = 𝛽0+ 𝛽1𝑃𝑡−1+ 𝛽2𝑓ℎ𝑡+ 𝛽3𝑠𝑞𝑡+ 𝛽4𝑑𝑢𝑚𝑚𝑦𝑤𝑒𝑒𝑘𝑒𝑛𝑑+ 𝛽5𝑑𝑢𝑚𝑚𝑦𝑛𝑖𝑔ℎ𝑡+ 𝛽6𝑑𝑢𝑚𝑚𝑦𝑤𝑖𝑛𝑡𝑒𝑟+ 𝛽7𝑑𝑢𝑚𝑚𝑦𝑠𝑝𝑟𝑖𝑛𝑔+ 𝛽8𝑑𝑢𝑚𝑚𝑦𝑠𝑢𝑚𝑚𝑒𝑟 + 𝜖, 𝜖 ~ 𝑁(0, 𝜎2)
With 𝑓ℎ𝑡 as the mean wind speed of hour 𝑡 and 𝑠𝑞 as time of sunshine variable. Betas are regression coefficients and epsilon represents an error term, assumed to be normally distributed with a mean of zero. Following the empirical results of Woo et al. (2011) and Cutler et al. (2011) our alternative hypotheses include 𝛽2 < 0 and 𝛽3 < 0. We expect the extra hours of sun and a higher wind velocity to have negative effects on the spot price due to the higher supply of electricity they both generate.
2. An Alternative Model
To be surer about the causal relationship between sunny weather and a lower spot price we can set up a model with, instead of the minutes of sunshine, a cloudiness variable. Let 𝑛𝑡 be a score of cloudiness following an ordinal distribution from 0 to 9. Our hypothesis for this model is 𝛽3 > 0.
Model 3.2
𝑃𝑡 = 𝛽0+ 𝛽1𝑃𝑡−1+ 𝛽2𝑓ℎ𝑡+ 𝛽3𝑛𝑡+ 𝛽4𝑑𝑢𝑚𝑚𝑦𝑤𝑒𝑒𝑘𝑒𝑛𝑑+ 𝛽5𝑑𝑢𝑚𝑚𝑦𝑛𝑖𝑔ℎ𝑡+ 𝜖,
𝛽6𝑑𝑢𝑚𝑚𝑦𝑤𝑖𝑛𝑡𝑒𝑟+ 𝛽7𝑑𝑢𝑚𝑚𝑦𝑠𝑝𝑟𝑖𝑛𝑔+ 𝛽8𝑑𝑢𝑚𝑚𝑦𝑠𝑢𝑚𝑚𝑒𝑟+ 𝜖 ~ 𝑁(0, 𝜎2)
Because of the negative correlation of the cloudiness variable with time of sunshine we expect its predictor 𝛽3 to be positive. Next, we set up a model with an interaction effect of the
night-time dummy with cloudiness to exclude the possibility of cloudiness being a random variable correlating with electricity prices.
Model 3.3
𝑃𝑡 = 𝛽0+ 𝛽1𝑃𝑡−1+ 𝛽2𝑓ℎ𝑡+ 𝛽3𝑑𝑢𝑚𝑚𝑦𝑛𝑖𝑔ℎ𝑡∗ 𝑛𝑡+ 𝛽4𝑑𝑢𝑚𝑚𝑦𝑤𝑒𝑒𝑘𝑒𝑛𝑑 + 𝛽5𝑑𝑢𝑚𝑚𝑦𝑛𝑖𝑔ℎ𝑡 + 𝛽6𝑑𝑢𝑚𝑚𝑦𝑤𝑖𝑛𝑡𝑒𝑟+ 𝛽7𝑑𝑢𝑚𝑚𝑦𝑠𝑝𝑟𝑖𝑛𝑔+ 𝛽8𝑑𝑢𝑚𝑚𝑦𝑠𝑢𝑚𝑚𝑒𝑟+ 𝜖, 𝜖 ~ 𝑁(0, 𝜎2)
If we find statistical insignificance of the interaction effect we can be surer that statistical evidence of lower spot prices due to more sunshine is the result of solar panels supplying more energy.
3. The Volatility of Spot Prices
To measure the effect of weather conditions on the volatility of spot prices we set up a new model. This model is no result of earlier empirical research. Therefore we cannot say anything about the causality of all of these variables. Only Woo et al. have concluded that an increased wind velocity will result higher price volatility and Cutler et al. found increased volatility during winter and summer months. The model includes a variable that accounts for residual lagging variance, variables that look for a relationship of prices and variance, seasonal and time of the day dummies. Finally, we include weather variables 𝑓ℎ𝑡 and 𝑠𝑞𝑡 which represent mean wind velocity and tenths of hours of sunshine, respectively. Our hypotheses are 𝛽3 > 0, 𝛽7 > 0 and 𝛽8 > 0. We expect wind velocity to have a positive relationship with spot price variances and we expect higher variances during the summer and winter months.
Model 3.4
𝜎𝑃2𝑡 = 𝛽0+ 𝛽1𝜎𝑃2𝑡−1+ 𝛽2𝑃𝑡+ 𝛽3𝑃𝑡−1 + 𝛽4𝑓ℎ𝑡+ 𝛽5𝑠𝑞𝑡+ 𝛽6𝑑𝑢𝑚𝑚𝑦𝑤𝑒𝑒𝑘𝑒𝑛𝑑 + 𝛽7𝑑𝑢𝑚𝑚𝑦𝑛𝑖𝑔ℎ𝑡+ 𝛽8𝑑𝑢𝑚𝑚𝑦𝑤𝑖𝑛𝑡𝑒𝑟+ 𝛽9𝑑𝑢𝑚𝑚𝑦𝑠𝑢𝑚𝑚𝑒𝑟 + 𝛽10𝑑𝑢𝑚𝑚𝑦𝑠𝑝𝑟𝑖𝑛𝑔+ 𝜖, 𝜖 ~ 𝑁(0, 𝜎2)
4. Data
Weather data was retrieved from the Royal Netherlands Meteorological Institute (KNMI) which is the Dutch national weather forecasting service which supplies a lot of accurate data about meteorological and climatological factors, including weather conditions.12 For this research we used data of 19 Dutch weather stations. They are located at De Kooy, Schiphol, Hoorn, De Bilt, Stavoren, Lelystad, Leeuwarden, Deelen, Hoogeveen, Eelde, Twenthe, Vlissingen, Rotterdam, Gilze-Rijen, Herwijnen, Eindhoven, Volkel, Ell and Maastricht. Because the electricity wholesale market is national, weather stations were selected by location with the intention to get an even geographical spread of data. By averaging the data we observed from all these weather stations we tried to get an as accurate as possible estimation of weather conditions in the Netherlands. Moreover, these weather stations all provide data about the variables we want to analyse, which are mean wind speed and sunshine hours. We also made sure to include weather stations located near the North Sea to analyse the effect of wind turbine parks that are located there. Unfortunately only actual weather data are available, so we cannot make any predictions about effects of forecasted weather conditions on prices. We observe some descriptive statistics in table 4.1. 𝑓ℎ represents the mean wind velocity in the Netherlands during the last hour in 0.1∙m/s. 𝑡 shows the mean temperature of the Netherlands during the last hour, measured in 0.1∙C°. 𝑛 is the average cloudiness in the Netherlands. It follows an ordinal distribution with a scale from 0 to 9. 𝑠𝑞 represents the average time of sunshine during the hour, measured per 6 minutes. These variables do not include any outliers. Temperature and wind velocities represent trustworthy averages for the Netherlands.
12Koninklijk Nederlands Meteorologisch Instituut. 2015. Koninklijk Nederlands
Table 4.1: Descriptive statistics about weather variables
Variable | Obs Mean Std. Dev. Min Max ---+--- fh | 8760 42.58976 20.95913 5.2632 135.7895 t | 8760 115.5475 58.0837 -32.3158 325.1053 n | 8760 4.894128 2.30977 0 8 sq | 8760 2.078683 3.035924 0 10
TenneT is the Dutch transmission system operator owned fully by the Dutch government prices. It is fully responsible for the 380 and 200 kV high-voltage grid throughout the Netherlands. For this research we used data downloaded from their website.13 It includes the electricity spot prices for every quarter of an hour from December 31st 2013 to December 31st 2014 measured in euros. Because we could get electricity spot prices per 15 minutes and weather info only per hour we had to streamline the data. Table 4.2 exhibits some descriptive statistics about the market. The minima and maxima show the huge volatility of the market. In one hour spot prices can vary with tens of their own standard deviation and return to equilibrium multiple times. Therefore we use the average of four 15 minute spot prices in order to get a more general understanding of influence the weather conditions of that particular hour on the prices of that hour. When we would use spot prices a price could be the result of the market being in a spike regime. In a sudden spike regime because of (for example) a short power outage, prices will deviate far from their long-term equilibrium. These huge spikes of prices can never be the effect of weather conditions. The long term equilibrium price of an MWh tends to be somewhat around 40 euros. Very interesting phenomena are the negative prices. When sudden extreme demand drops occur very inflexible power generators such as coal stations cannot stop producing in a cost effective manner for the short duration of that demand drop. To prevent the whole grid from overloading, parties are tempted into tapping more electricity from the market by decreased prices to the extent that they get paid for doing just that. For extended
periods (longer than an hour) the average can sit around its mean plus tens of its standard deviation. Table 4.3 shows the outliers of our dataset. From this table we can conclude that the fact that the maximum price has about the same deviation in positive direction as the minimum has in negative direction is a randomness. We conclude this because of the simple fact that the minimum occurred 163 days before the maximum and data from 2015 showed that prices can get even higher than these values. We have no reason to drop any of these outliers since it is in the very nature of the spot market to be vulnerable to such huge spikes. Figure 2a and 2b show the distributions of the same variables as presented in table 4.2. The distributions are heavily centred on the long-term equilibrium. By looking at these figures we can see why Janczura & Weron (2010) advocate the use of mean-reverting models. Prices tend to fluctuate around an equilibrium but do variate significantly. For this research we have chosen not to build on the suggested models. These models are too complicated for the mere two weather effects we want to analyse.
Table 4.2: Descriptive statistics of spot prices in MWh
Variable | Obs Mean Std. Dev. Min Max ---+--- meanprice | 8760 40.16502 37.31922 -434.2 436.74 meanpricet~1 | 8760 40.16204 37.32137 -434.2 436.74 spotprice | 8760 38.97064 49.32463 -435.1 446.12 spotpricet~1 | 8760 38.96835 49.32522 -435.1 446.12
Table 4.3: Outlier analysis for mean hourly prices. +---+ | obs: meanpr~e | |---| | 3845. -434.2 | | 3844. -385.1 | | 2631. -277.99 | | 1222. -211.4 | | 32. -178.59 | +---+ +---+ | obs: meanpr~e | +---+ | 4462. 360.77 | | 1835. 386.33 | | 4460. 403.02 | | 1836. 428.34 | | 7770. 436.74 | +---+
Moreover, these models all assume a lognormal distribution of spot prices (figure 4c). We will not make use of this assumption because of the significant amount of data lost when using logarithms of positive and negative prices. Table 4.4 shows that the main variables of model 3.1 do not correlate and we have no reason to exclude variables due to collinearity. The highest of these is the correlation between the night-time dummy and the sunshine variable, which is expected and of no concern to our research.
Table 4.4: Correlation matrix of variables used in model 3.1
| meanpr~e meanpr~1 fh sq nightt~y weeken~y ---+--- meanprice | 1.0000 meanpricet~1 | 0.4663 1.0000 fh | -0.0910 -0.0910 1.0000 sq | -0.0398 -0.0339 0.0435 1.0000 nighttimed~y | -0.1691 -0.1280 -0.1427 -0.3734 1.0000 weekenddummy | -0.0822 -0.0806 0.0257 0.0242 -0.0000 1.0000
To analyse model 3.4 we calculated the variance of the electricity spot prices of every hour during 2014. Table 4.5 shows descriptive statistics of the hourly variances of spot prices. These are the hourly variances calculated using the prices per 15 minutes. Table 4.6 exhibits the hourly standard deviation of electricity spot prices calculated by taking the square root of variances. We observe 70 hours during which the volatility was 0. We also observed hours in which the mean deviation was more than 250 euros. Again, we have no reason to neglect these values because of the volatility of this market. Finally, from the correlation matrix in table 4.7 we can conclude that there is no collinearity in the model. Figure 4d shows the distribution of standard deviation which has the shape of a standard Χ2 distribution but with a heavy tendency towards zero. This means that often the market is quite stable, but is prone to some extreme spikes.
Table 4.5: Descriptive statistics of the variance of electricity spot prices per hour Variable | Obs Mean Std. Dev. Min Max ---+--- var | 8760 2023.44 6146.205 0 113211.8
Table 4.6: Outlier analysis of the hourly standard deviation of electricity spot prices +---+ | obs: sd | |---| | 86. 0 | | 154. 0 | | 483. 0 | | 556. 0 | | 575. 0 | +---+ +---+ | 1847. 250.105 | | 1750. 250.45 | | 408. 258.9185 | | 8360. 326.7564 | | 1056. 336.4696 | +---+ note: 70 values of 0 Table 4.7: Correlation matrix of non-dummy variables of model 3.4
| var vartmin1 meanpr~e meanpr~1 fh sq ---+--- var | 1.0000 vartmin1 | 0.1794 1.0000 meanprice | 0.3626 0.1042 1.0000 meanpricet~1 | 0.0514 0.3626 0.4663 1.0000 fh | -0.0135 -0.0076 -0.0910 -0.0910 1.0000 sq | -0.0837 -0.0600 -0.0398 -0.0339 0.0435 1.0000
4. Empirical Results
1. The effect of weather conditions
After running OLS regression we found significant evidence for an effect of hours of sunshine and wind velocity on hourly average electricity spot prices. The 𝑅2 is 0.24. 24% of the variance of the spot prices is determined by this model. This is low, but expected because of the relatively low sustainable energy sources in use together with the fact that demand is still the main driving factor of spot prices (Cutler et al., 2011). The ANOVA 𝐹2 is 349.94 and significant. This means that a significant proportion of the variance is accounted for by the predictors. Our model for average spot prices predicts a constant of 35.78. For every euro of the average price of the last hour this is increased with .43. This lagging effect is expected because demand for electricity often overlaps several hours. Earlier research also predicted a night-time effect and a weekend effect which are significant in our research as well. In both periods the economy isn’t operating at full capacity and as a result demand for electricity is
lower. On average this results in a price decrease of €13.04 euros during night-time and €3.60 euros in the weekend.
Table 5.1: OLS regression of model 1
Source | SS df MS Number of obs = 8760 ---+--- F( 8, 8751) = 349.94 Model | 2956650.42 8 369581.303 Prob > F = 0.0000 Residual | 9242220.01 8751 1056.13301 R-squared = 0.2424 ---+--- Adj R-squared = 0.2417 Total | 12198870.4 8759 1392.7241 Root MSE = 32.498 --- meanprice | Coef. Std. Err. t P>|t| [95% Conf. Interval] ---+--- meanpricetmin1 | .4321528 .0095274 45.36 0.000 .4134769 .4508287 fh | -.1315341 .0180141 -7.30 0.000 -.166846 -.0962223 sq | -.886763 .1277559 -6.94 0.000 -1.137195 -.6363315 nighttimedummy | -13.04049 .8861769 -14.72 0.000 -14.77761 -11.30338 weekenddummy | -3.604173 .7720608 -4.67 0.000 -5.117593 -2.090752 springdummy | -3.352082 1.021285 -3.28 0.001 -5.354042 -1.350123 summerdummy | -.6626821 1.021537 -0.65 0.517 -2.665135 1.339771 winterdummy | -.951897 .9933628 -0.96 0.338 -2.899122 .9953276 _cons | 35.77872 1.280733 27.94 0.000 33.26819 38.28926 ---
Most importantly we found significant results for weather conditions. Every .1∙m/s of wind speed accounts for a 13 eurocent decrease in average spot price with a t-score of -7.3. Every 6 minutes of sunlight accounts for a decrease of 87 eurocents of the spot price with t-score of -6.94. Seasonal variables are not significant except for spring season. This is unexpected compared to earlier research and follow-up research could look at the cause of these phenomena. Concluding, the result of our model without seasonal dummies is equation 5.1 with regression table 5.2. The effect of sunshine and wind are significant but do not have a huge impact on the final spot price. Following Woo et al. (2011) & Cutler et al. (2011) we can say that these effects are the effect of increased supply due to more durable energy supply. The effects are so small because of the relative small size of these supplies. When these markets get more sizable we expect to see an increase of these predictors. We can reject hypotheses 𝐻0: 𝐵𝑠𝑞 = 0 and 𝐻0: 𝛽𝑓ℎ = 0. Results of the model are visualized in figure 5a and 5b by plotting
the best fitting lines of the influence of wind speed and hours of sunshine on average hourly spot prices.
Equation 5.1
𝑃𝑡 = 34.23 + .43𝑃𝑡−1− .12𝑓ℎ𝑡− .96𝑠𝑞𝑡− 3.59𝑑𝑢𝑚𝑚𝑦𝑤𝑒𝑒𝑘𝑒𝑛𝑑 − 13.14𝑑𝑢𝑚𝑚𝑦𝑛𝑖𝑔ℎ𝑡+ 𝜖
Table 5.2: Regression of model 3.1 without seasonal dummies.
Source | SS df MS Number of obs = 8760 ---+--- F( 5, 8754) = 556.77 Model | 2943332.94 5 588666.589 Prob > F = 0.0000 Residual | 9255537.49 8754 1057.29238 R-squared = 0.2413 ---+--- Adj R-squared = 0.2408 Total | 12198870.4 8759 1392.7241 Root MSE = 32.516 --- meanprice | Coef. Std. Err. t P>|t| [95% Conf. Interval] ---+--- meanpricet~1 | .4343558 .0095122 45.66 0.000 .4157096 .453002 fh | -.122376 .0168589 -7.26 0.000 -.1554235 -.0893285 sq | -.9586154 .123903 -7.74 0.000 -1.201494 -.7157363 nighttimed~y | -13.14106 .8853392 -14.84 0.000 -14.87654 -11.40559 weekenddummy | -3.594587 .7724794 -4.65 0.000 -5.108829 -2.080346 _cons | 34.23451 1.075645 31.83 0.000 32.12599 36.34302 ---2. An alternative model
To investigate the weather effect of sunshine a little further we can alter the equation of model 3.1. We add a variable 𝑛𝑡, which is a scale of cloudiness that ranges from 0 to 9 and drop variable 𝑠𝑞𝑡. Because cloudiness and sunshine are highly correlated we expect to see the same effect of cloudiness on the spot prices, but in the opposite direction. In table 5.3 we see the OLS results. Seasonal dummies are excluded from the model. All the unchanged variables do show similar results as they do in the previous table. The effect of cloudiness indeed works in opposite direction of the sunshine variable does in the previous model. With a coefficient of 1.02 and a t-value of 6.48 we conclude that cloudiness has roughly the same effect on spot prices as sunshine has. Is this the effect of lack of sunshine? Cloudiness could occur during the night as well. The result of model 3.2 is shown in equation 5.2.
Equation 5.2
To analyse the effect of cloudiness we set up a model with an interaction effect of the night-time dummy and cloudiness. If this effect turns out to be significant, cloudiness also correlates with spot prices during the night, which has no causal background yet. The results are exhibited table 5.4.
Table 5.3: OLS regression results of model 3.2
Source | SS df MS Number of obs = 8760 ---+--- F( 8, 8751) = 348.92 Model | 2950104.88 8 368763.111 Prob > F = 0.0000 Residual | 9248765.55 8751 1056.88099 R-squared = 0.2418 ---+--- Adj R-squared = 0.2411 Total | 12198870.4 8759 1392.7241 Root MSE = 32.51 --- meanprice | Coef. Std. Err. t P>|t| [95% Conf. Interval] ---+--- meanpricetmin1 | .4317546 .0095398 45.26 0.000 .4130544 .4504548 fh | -.1659058 .018506 -8.96 0.000 -.2021819 -.1296297 n | 1.023722 .1580557 6.48 0.000 .7138956 1.333548 nighttimedummy | -10.89643 .820854 -13.27 0.000 -12.50549 -9.287359 weekenddummy | -3.489478 .7729272 -4.51 0.000 -5.004597 -1.974359 springdummy | -4.186744 1.003925 -4.17 0.000 -6.154672 -2.218816 summerdummy | -1.254515 1.00891 -1.24 0.214 -3.232216 .723186 winterdummy | -.3644755 1.005311 -0.36 0.717 -2.335122 1.606171 _cons | 30.0488 1.425867 21.07 0.000 27.25377 32.84384 ---
Table 5.4: OLS regression results of model 3.3
Source | SS df MS Number of obs = 8760 ---+--- F( 5, 8754) = 541.21 Model | 2880489.87 5 576097.975 Prob > F = 0.0000 Residual | 9318380.56 8754 1064.47116 R-squared = 0.2361 ---+--- Adj R-squared = 0.2357 Total | 12198870.4 8759 1392.7241 Root MSE = 32.626 --- meanprice | Coef. Std. Err. t P>|t| [95% Conf. Interval] ---+--- meanpricetmin1 | .440804 .0095077 46.36 0.000 .4221667 .4594412 fh | -.1209101 .0170272 -7.10 0.000 -.1542874 -.0875328 interact | .1932216 .2989347 0.65 0.518 -.3927607 .7792039 nighttimedummy | -11.48678 1.67891 -6.84 0.000 -14.77784 -8.195722 weekenddummy | -3.704314 .7750094 -4.78 0.000 -5.223514 -2.185113 _cons | 31.304 1.011327 30.95 0.000 29.32156 33.28644 ---
As expected, cloudiness during the night has no significant influence on the spot price. The 𝑅2 is 0.24 and does not deviate a lot from the first model. Likewise, the ANOVA 𝐹2 is significant. The t-statistic of the interaction of cloudiness with night-time is 0.65. The effect of cloudiness
is not significant anymore. We can conclude that the positive effect of cloudiness is the result of the decrease of sunshine, which reinforces our hypothesis that an increased supply of solar energy does result in a drop of energy prices.
3. The volatility of spot prices
The third and last model inspects the effect of weather conditions in hour 𝑡 on the volatility 𝜎𝑡2. The alternative hypotheses we wrote were 𝐻1: 𝛽𝑓ℎ > 0, 𝐻1: 𝛽𝑑𝑢𝑚𝑚𝑦𝑤𝑖𝑛𝑡𝑒𝑟 > 0 and
𝐻1: 𝛽𝑑𝑢𝑚𝑚𝑦𝑠𝑢𝑚𝑚𝑒𝑟 > 0. The results are shown in table 5.5. The model is significant with an
F-score of 308.74 and 20% of the variance explained. We observe that the variance in 𝑡 is dependent on the variance in 𝑡 − 1. This means that uncertain periods in this market tend to last for several hours. Another effect we observe is that of prices. When prices are high, the market tends to be very uncertain. However, the effect of the hour before 𝑡 tends to be negative. This means that in extended times of high prices the market becomes calmer. Another significant effect is the effect of the average 0.1 hours of sunshine in the last hour. It has a very strong negative relationship with the variance of that hour. When the sun shines the market is calmer. We also observe that the market is less variable during the night and weekends. However, what we expected to observe, 𝛽𝑓ℎ> 0, we do not observe significantly. We have no empirical explanation for this and further research is needed to explain why this is not the case on the Dutch electricity market. The hypothesis we wrote about the market being more volatile during the summer cannot be confirmed significantly. However, we can confirm the market being more volatile in the winter months. Regular traders on this market do confirm these seasonal results.14
Table 5.5: OLS regression of variance of spot prices
Source | SS df MS Number of obs = 8760 ---+--- F( 10, 8749) = 219.54 Model | 6.6373e+10 10 6.6373e+09 Prob > F = 0.0000 Residual | 2.6451e+11 8749 30232682.3 R-squared = 0.2006 ---+--- Adj R-squared = 0.1997 Total | 3.3088e+11 8759 37775837.6 Root MSE = 5498.4 --- var | Coef. Std. Err. t P>|t| [95% Conf. Interval] ---+--- vartmin1 | .2098437 .0103667 20.24 0.000 .1895225 .2301648 meanprice | 72.4074 1.813484 39.93 0.000 68.85255 75.96226 meanpricetmin1 | -39.71274 1.918833 -20.70 0.000 -43.4741 -35.95137 fh | -2.464532 3.057405 -0.81 0.420 -8.457765 3.528701 sq | -146.0891 21.68013 -6.74 0.000 -188.5872 -103.5909 weekenddummy | -701.1937 131.0503 -5.35 0.000 -958.0831 -444.3043 nighttimedummy | -550.5031 151.9985 -3.62 0.000 -848.4559 -252.5504 winterdummy | 775.7223 168.3361 4.61 0.000 445.7439 1105.701 summerdummy | 103.2617 172.8398 0.60 0.550 -235.545 442.0683 springdummy | -6.711807 172.9019 -0.04 0.969 -345.6401 332.2165 _cons | 815.9626 226.1792 3.61 0.000 372.5983 1259.327 ---
6. Conclusion
In this paper we wanted to analyse the effects of some weather variables on the Dutch wholesale electricity market. In particular, our intentions were to find any effects of the extra supply of durable energy on the stock price. Likewise, because solar energy increased tenfold from 2010 to 2014 we expected a similar pattern. Firstly, we gave an introduction to the relatively unknown electricity wholesale market. Then we looked into previous research. We learned that previous research found seasonal, time of the day and weekend influences. As early as 2001 weather influence was concluded as well. Research from Texas and Australia also found evidence of the direct impact of an extra supply of wind energy on the spot price of electricity. Although never previously confirmed, we tried to find similar effects of solar energy. Thereafter we set up three models. The first model included variables to explain the effects previous research had confirmed and included solar and wind variables. The second model looked at the inverse of sunniness: cloudiness. We used an interaction variable to conclude that this effect only occurred during day time and hence a sunshine effect could be confirmed. Finally, we set up an hourly variance model, not backed by a lot of previous research. Only a Texan study found an increase of volatility of spot prices due to an increase of wind speed and an Australian study found increased seasonal volatility. Data was retrieved from TenneT and KNMI and streamlined using the
hourly average of 15 minute spot prices. When we ran OLS regression on the first model we found significant effects on the average hourly spot prices of all previous confirmed variables except for seasonal effects. Wind velocity and hourly time of sunshine have significant influence on spot prices. We conclude that they both shift supply curves to the right and result in a lower spot price. This got confirmed when we used a cloudiness variable and let it interact with the night-time dummy, this predictor turned out to be insignificant. Finally, when we ran the OLS regression on hourly variances we found a significant lagging effect, significant price effects and just one significant seasonal effect: winter. We found no expected increased volatility in the summer and no expected significant wind effect. This is unexpected because of the hypothesis that wind velocity should increase spot price volatility. We did find a sunshine effect that reduced spot price volatility significantly. Follow-up research should look into these effects with better fitting models. Suggested research could try to explain why we did not find any seasonal effect in the Netherlands or why increased wind speed did not increase volatility. In the long run this kind of research will get more important since the Dutch government plans to receive 100% of the electricity on the Dutch market from natural resources by 2050. Therefore we expect to find increasingly stronger negative price effects of both sunshine and wind velocity in the future to the point where forecasted weather conditions will be the main predictors of wholesale electricity spot prices.
Bibliography
Blom, J., and Ruhe, J. June 28th 2015. Personal Communication.Carnero, Angeles M., J. S. Koopman, and M. Ooms. 2003. "Periodic Heteroskedastic RegARFIMA models for daily electricity spot prices." Discussion Paper Tinbergen
Institute Amsterdam.
Centraal Bureao voor de Statistiek. 2015. Centraal Bureao voor de Statistiek. Accessed June 28, 2015. http://www.cbs.nl/.
Cutler, Nicholas J., Nicholas D. Boerema, F. Macgill Iain, and Hugh R. Outhred. 2011. "High Penetration Wind Generation Impacts on Spot Prices in the Australian National Electricity Market." Energy Policy 5939-949.
Deng, Shijie. 1998. "Stochastic Models of Energy Commodity Prices and Their applications: Mean-reversion with Jumps and Spikes." University of California Energy Institute. Escribano, Alvaro, Juan Ignacio Peña , and Pablo Villaplana . 2011. "Modelling Electricity
Prices: International Evidence." Oxford Bulletin of Economics and Statistics 622-650. Huisman, Ronald, and Ronald J. Mahieu. 2003. "Regime Jumps in Electricity Prices." Energy
Economics 425-434.
Janczura, Joanna, and Rafal Weron. 2010. "An Empirical Comparison of Alternate Regime-switching Models for Electricity Spot Prices." Energy Economics 1059-073.
Knittel, Christopher R., and Michael R. Roberts. 2001. "An Empirical Examination of Deregulated Electricity Prices." PWP-087.
Knittel, R. Christopher, and Michael R. Roberts. 2005. "An Empirical Examination of Restructured Energy Prices." Energy Economics 791–817.
Koninklijk Nederlands Meteorologisch Instituut. 2015. Koninklijk Nederlands
Meteorologisch Instituut. Accessed June 28, 2015. http://www.knmi.nl/.
Kosater, Peter. 2006. "On the impact of weather on German hourly power prices." Discussion
papers in statistics and econometrics.
TenneT. 2014. TenneT. Accessed June 28, 2015. http://www.tennet.nl.
Woo, C. K. , I. Horowitz, J. Moore, and A. Pacheco. 2011. "The impact of wind generation on the electricity spot-market price level and variance: The Texas experience." Energy
Appendix
Figure 1a: Increase of total renewable and wind energy in 106 kWh (Source: CBS)
Figure 1b: Increase of energy supplied by solar panels in 106kWh (Source: CBS)
0 5 0 0 0 1 0 0 0 0 1 5 0 0 0 T o ta l Su st a in a b le En e rg y Pro d u c e d i n Mi lli o n s o f kW h s 1990 1995 2000 2005 2010 2015 year 0 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 En e rg y Su p p lie d By So la r Pa n e ls in Mi lli o n s kW h 1990 1995 2000 2005 2010 2015 year
Figure 4a: Histogram of the average spot prices per hour
Figure 4b: Histogram of the hourly average spot prices of every last 15 minutes of the hour
0 .0 0 5 .0 1 .0 1 5 .0 2 D e n si ty -400 -200 0 200 400 MeanPrice 0 .0 0 5 .0 1 .0 1 5 .0 2 .0 2 5 D e n si ty -400 -200 0 200 400
Figure 4c: Histogram of the logarithm of mean hourly prices
Figure 4d: Histogram of hourly standard deviations.
0 .2 .4 .6 .8 1 D e n si ty -4 -2 0 2 4 6 logmeanprice 0 .0 2 .0 4 .0 6 D e n si ty 0 100 200 300 400 sd
Figure 5a: Two-way plot of best fitting line of prices (y-axis) and mean wind speed in .1∙m/s
Figure 5b: Two-way plot of best fitting line of prices (y-axis) and average time of sunshine
25 30 35 40 45 F it te d va lu e s 0 50 100 150
Mean Wind Speed 0.1MS
36 37 38 39 40 41 F it te d va lu e s 0 2 4 6 8 10 Time of Sunshine 0.1Hrs
