IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 23, NO.

1, FEBRUARY 2008

25

Forecasting the Mean and the Variance of Electricity Prices in Deregulated Markets
Claudio M. Ruibal and Mainak Mazumdar
AbstractA fundamental bid-based stochastic model is presented to predict electricity hourly prices and average price in a given period. The model captures both the economic and physical aspects of the pricing process, considering two sources of uncertainty: availability of the units and demand. This work is based on three oligopoly modelsBertrand, Cournot, and supply function equilibrium (SFE) due to Rudkevich, Duckworth, and Rosenand obtains closed form expressions for expected value and variance of electricity hourly prices and average price. Sensitivity analysis is performed on the number of rms, anticipated peak demand, and price elasticity of demand. The results show that as the number of rms in the market decreases, the expected values of prices increase by a signicant amount. Variances for the Cournot model also increase, but the variances for the SFE model decrease, taking even smaller values than Bertrands. Thus, if the Rudkevich model is an accurate representation of the electricity market, the results show that an introduction of competition may decrease the expected value of prices but the variances may actually increase. Finally, using a renement of the model, it has been demonstrated that an accurate temperature forecast can reduce signicantly the prediction error of the electricity prices. Index TermsAverage prices, Bertrand model, Cournot model, deregulated electricity markets, Edgeworth expansion, electricity price variance, electricity prices, hourly prices, method of cumulants, Rudkevich, Duckworth, and Rosens formula, stochastic load, supply function equilibrium.

Bivariate cumulant of order .

of

Actual electricity load at time (assumed to be normally distributed) . Equivalent load at time . Marginal unit at daily peak. Number of companies in the market. Number of generating units in the market. Price of electricity at time . Average price between hours and .

Proportion of time that generating unit is up. Joint probability of .

Proportion of time that generating unit is down, known as forced outage rate (FOR) . Averaging weight of the hourly load at time . Excess of load not met by the available generated power up to generating unit at time . Generating unit state at hour ( in case of outage) . working, if

NOTATION

Mean time to failure of generating unit . Capacity of unit . Sum of capacities of the rst units. Total system demand at time as a function of . Variable cost of generating unit . Derivative of time . with respect to price at Mean time to repair of generating unit . Mean of load at time . Correlation coefcient. Correlation coefcient between . Variance of load at time . Covariance between the loads at time and . I. INTRODUCTION . and

Marginal cost at time . Marginal unit at time . Cumulant of order # of

Manuscript received March 14, 2007; revised July 20, 2007. This work was supported by the National Science Foundation under Grant No. ECS-0245355. Paper no. TPWRS-00175-2007. C. M. Ruibal is with the Universidad de Montevideo, Montevideo, Uruguay (e-mail: cruibal@um.edu.uy). M. Mazumdar is with the University of Pittsburgh, Pittsburgh, PA 15261 USA (e-mail: mmazumd@engr.pitt.edu). Digital Object Identier 10.1109/TPWRS.2007.913195

N THIS PAPER, we outline a procedure for forecasting the mean and variance of the average price of electricity over a specied time interval in a deregulated market. Such information would be found useful in nancial forecasts, risk management, derivative pricing, investment, and operational decisions. The computations are based on a system model in which the physical and engineering processes and the bidding strategies are simultaneously considered. The price of electricity depends

0885-8950/$25.00 2007 IEEE

26

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 23, NO. 1, FEBRUARY 2008

on physical factors such as production cost, load, generation reliability, unit commitment, and transmission constraints. It also depends on economic factors such as strategic bidding and load elasticity. We consider here a model that captures the dependence of the price on costs, load, reliability, and bidding strategies. Many of these factors are stochastic in character which we have characterized by their probability distributions. A potential advantage of our approach is that it can be used to consider changes in systems structure over time (e.g., entry of additional generators or a change in load). We do not consider factors related to transmission congestion, transmission outages, and unit commitment. While these factors should be accounted for in a full and complete description of the movement of electricity prices, we believe that the system-based approach in this paper is an important rst step in the construction of a comprehensive model. The inclusion of unit commitment and transmission constraints will make the system model very complicated. The degree of complexity that will ensue can be comprehended by referring to Hobbs et al. [6] and to the papers contained in Hobbs et al. [7]. The emphasis in the current paper is on the use of analytical methods to forecast the statistical distributions of prices. When unit commitment and transmission constraints are included in the system model, it appears that there will remain no alternative other than using Monte Carlo methods for such forecasts. The approach that we have taken for modeling the prices is as follows. First we have provided the formulation for computing the mean and variance of price for any specied hour given its generation and load characteristics using a traditional production costing-type model. Then we have given the framework for computing the mean and variance of the time-average of the price over a specied time interval. This latter step is much more difcult because in addition to the results obtained from the rst step, estimates on covariances of hourly loads and production quantities become necessary. In this effort, we were aided by the method of cumulants-based formulation (in Valenzuela [16]) used for nding means and variances of production costs over a given interval. The estimates of variances in addition to those of the expected values would allow computation of prediction intervals for the price as well as individual rms prots. By comparing these prediction intervals yielded by different models with the actually realized prices, a judgment can be made about the accuracy of the individual models. This is the essence of what is known as backcasting (Paehlke [11]). Also an estimate of variance will be useful for the purposes of risk management, for example, in the computation of the value-at-risk and con ditional value-at -risk indices (Pilipovic [12], Rockafellar and Uryasev [13]). Its use in the derivative market is also apparent. We consider three bidding models for the market price of electricity. In each model, we assume that the electricity traded within the region of interest is unconstrained by transmission. The market under consideration is one where marginal bid pricing is assumed to prevail. We consider three bidding models: the Bertrand model [1] in which rms offer their marginal costs, the Cournot model [2] in which rms offer quantities that maximize expected prots, and a supply function equilibrium (SFE) model (Klemperer and Meyer [10]) in which rms offer a supply curve (quantity versus price) based on

Rudkevich, Duckworth, and Rosens equilibrium formula [14]. These three models are based on Nash equilibrium1 solutions for different bidding strategies. Other papers, like Kang et al. [9], explicitly include the bidding strategies in the model. We assume that the deregulated market will eventually arrive at an equilibrium. So, using the equilibrium solutions becomes a more realistic analysis. The rst two models have been considered extensively in the context of deregulated electricity markets. The elegant analytical expression developed in the Rudkevich model is applicable when a) the competing rms are identical in every respect, b) the load has zero elasticity with respect to price, and c) the price at the maximum load is equal to the marginal cost of supplying it. Because of our desire to use the simplied closed form expressions given by the Rudkevich formula, we have retained the assumptions used in this work. No doubt, the rst assumption is far removed from reality. It can perhaps be partially justied by noting that only a few rms realistically inuence the price: namely, those that usually own the marginal unit. It is assumed that those are not the small rms which are price-takers, but the large ones. Only the large rms participating in the market are considered identical. Even if the rms were actually identical, generator outages would break the theoretical and assumed symmetry. We have performed sensitivity analysis on the statistics related to electricity prices varying the number of rms, anticipated peak demand, and price elasticity of demand. The results show that as the number of rms in the market decreases, the expected values of prices increase by a signicant amount. Variances for the Cournot model also increase, but the variances for the SFE model decrease, taking even smaller values than Bertrands. Thus, if the Rudkevich model is deemed to be an accurate representation of the electricity market, the results suggest that an introduction of perfect competition may decrease the expected value of prices but the variances may actually increase. We also address the following question: can the electricity prices be more accurately predicted in the sense of obtaining smaller prediction intervals if accurate predictions of ambient temperature are available for the period for which the average price is being computed? This paper is organized into the following sections. Section III briey describes the three economic models. Section IV gives the formulas for the mean and variance of hourly price. The expressions for the mean and variance of the time averaged price over a given interval are given in Section V. Section VI describes a stochastic model for factoring in the forecast information about ambient temperatures. Section VII gives the numerical results for a hypothetical market. Section VIII states the conclusions. II. BASIC MODELS ON ELECTRICITY PRICING In the current literature, three major models are in use for (imperfect) electricity markets: the Bertrand model, Cournot model, and supply function equilibrium (SFE) model. Cournot and Bertrand models constitute the two often used paradigms of imperfect competition. In the Bertrand (1883) model, rms compete in price. They simultaneously choose prices and then
1Nash equilibrium is a prole of strategies such that each players strategy is an optimal response to the other players strategies.

RUIBAL AND MAZUMDAR: FORECASTING THE MEAN AND THE VARIANCE OF ELECTRICITY PRICES IN DEREGULATED MARKETS

27

must produce enough output to meet demand after the price choices become known. Under the assumption that each rm has enough capacity to meet demand, the Nash equilibrium price in this model is the marginal cost which is the same as the case of perfect competition. The other basic noncooperative equilibrium is the Cournot (1838) model. In this model, competition is in quantities. Firms simultaneously choose the quantities they will produce, which they then sell at the market-clearing price (the price for which demand is met by supply). Firms choose the quantities that optimize their prot. The Cournot model is a more accurate representation of the market. The assumption underlying the Bertrand model that competition is over prices and the rms have enough capacity to meet demand is not sustainable. Cournot models prevail over Bertrand models in the current literature on electricity markets. Yet another model has been used in the recent literature. This approach is based upon the work of Klemperer and Meyer [10] and was applied to a pool model by Green and Newbery [5]. In this model, competition is neither over price (as in Bertrand models) nor quantity (as in Cournot models) but in supply functions. A supply function relates quantity to price. It shows the prices at which a rm is willing to sell different quantities of output. The SFE model applies very well to the market structure of many restructured electricity markets, such as New Zealand, Australia, Pennsylvania-New Jersey-Maryland Interconnection (PJM), and California Power Exchange. In these markets, the bid format is precisely a supply function. SFE models can better explain the markups of electricity prices which empirical studies have shown to be above the Bertrand equilibrium but below the Cournot model. The problem with the use of SFE models is that in general, there is not a unique equilibrium. There is often an innite number of solutions lying between the Cournot and Bernard equilibria, which represent their upper and lower limits in price, respectively. The existence of many equilibria makes it difcult to predict the likely outcome of strategic interaction among players. There are some factors that reduce the range of feasible equilibria: uncertainty of demand and capacity constraints are among them. Rudkevich, Duckworth, and Rosen[14] calculated the electricity prices that would result from a pure pool market with identical prot-maximizing generating rms, bidding stepwise supply functions. Under the assumption that the price at peak demand is the marginal cost of the peak marginal unit, they obtain the unique Nash equilibrium marketclearing price of electricity in a pool, given by a closed formula expression.

dispatch the units from the cheapest to the more expensive ones, until the demand is met. This is the case with PJM and many other electricity markets. to get the exWe will condition on the marginal unit2 pected value and the variance of the price. The expected value can be written as follows: (1) . The where is the merit order index. is needed. Following Valenprobability mass function of zuela and Mazumdar [18], we will express (2) being the cumulative distribution . function of the auxiliary variable Cramer [3] provides the Edgeworth expansion of the distribution as follows: function of

where , is the standard normal cumulative probability distribution function, is the standard normal probability density function with are the cumulants of mean zero and unit variance; and order # of

III. MEAN AND VARIANCE OF THE HOURLY PRICE Ignoring unit commitment constraints, it is assumed that the generating units, which are dispatched system consists of in ascending merit order, based on the offered price of each one. st unit refers to a hypothetical generator with inThe nite capacity and perfect reliability that can be accessed for additional power in the event that the available capacity of the market is insufcient to meet the load. Utilities will offer energy (quantity and price), unit by unit, to the independent system operator (ISO). The latter will order the units by offered price, and

where and are the mean and variance of , respectively; is the nominal capacity of unit ; is the proportion of time that unit is up; and is the proportion of time that unit is . This formula is known as the method of down. cumulants in the power system literature. A random variable, called equivalent load, is used to consider the uncertainty of the load and the reliability of the units as well. It is dened as the load that could have been delivered if all the units up to the marginal unit were working and can be expressed as (3)
2Marginal unit is the last unit called on to produce electricity to meet demand.

28

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 23, NO. 1, FEBRUARY 2008

We use the approximation justied , the difference by considering that, if is smaller than . In practice, a power market has many units, such that an error smaller than the capacity of a single unit is negligible, from a practical point of view, for typical load values. Detailed results on the accuracy of this ap, the proximation are given in Ruibal [15]. Conditioning on expected value of the equivalent load is (4) Now we compute the expected value and variance of hourly load under the three economic models. Under the Bertrand model, the Nash equilibrium market-clearing price is the marginal cost. . Considering the s to be known That is, and deterministic constants, the expected price and variance at hour for the Bertrand model are

generalized for the stochastic case as follows: (10) where is the dispatch order number of the most expensive unit expected to run during the 24-h period. Conditioning on and using the approximation given in (4), the expected price and variance at hour for the Rudkevich model are

(11)

(12) (5) IV. MEAN AND VARIANCE OF THE TIME-AVERAGE PRICE (6) Green and Newbery [5] derive the Cournot model Nash equilibrium price for the basic case of a symmetric duopoly: , where and is the derivative of the total system demand with respect to price. Following Green [4], it is assumed in this is a linear function of price. work that the total demand The outage of units has the same effect on price as a shift upwards of the load, in the same amount of the power that cannot be delivered. In order to consider the uncertainty of the load and the availability of the units simultaneously, we use the equivadened in (3) instead of the actual load . lent load It can be seen that the Nash equilibrium market-clearing price for the Cournot model can be generalized to (7) Once again, conditioning on and using the approximation given by (4), the expected price and variance at hour for the Cournot model are This section derives formulas for the expected value and variance of average prices. The objective is to predict accurately the average for some specic hours of a given day. A loadweighted average is considered, as a more general approach. The weighted average price between the initial hour and nal is . The expected value and hour variance can be expressed as

(13)

(14) where the expressions for and have been derived in the preceding section for each bidding model. Note that and are correlated for any pair . To get an expression for , we need to compute . Conand , the following expression holds: ditioning on

(8) (15) (9) As discussed earlier for the Cournot model, the Nash equilibrium market-clearing price following Rudkevich, Duckworth, and Rosens [14] formula for deterministic load can be Valenzuela [16] has obtained an approximate expression using the Edgeworth exfor pansion formula, which is summarized below. The events and are equivalent. So, using the variables

RUIBAL AND MAZUMDAR: FORECASTING THE MEAN AND THE VARIANCE OF ELECTRICITY PRICES IN DEREGULATED MARKETS

29

dened before, . Denoting by the joint probability of the two events, the following equality holds:

and

TABLE I SUPPLY MODEL

(16) can be The joint probability distribution of approximated by using the Edgeworth expansion. Iyengar and Mazumdar [8] give the Edgeworth approximate expan, sion of the joint probabilitydistribution of shown in the equation at the bottom of the page, where , , and the and correlation coefcient between . Also, is the probability density function of the bivariate standard normal distribution with corare the bivariate cumulants of order relation coefcient . of and are the bivariate Hermite polynomials that can be found in Ruibal [15]. V. STOCHASTIC MODEL OF THE LOAD as In the preceding section, we consider the hourly load a random independent variable. This basic case is called model 1 in this section and in the following section. A question arises at this moment: Can we predict electricity prices more accurately if we can better explain the variability of demand? Some previous work (see Valenzuela, Mazumdar, and Kapoor [19]) showed that part of the load variance can be explained by the effect of temperature. Valenzuela and Mazumdar [17] calibrated a stochastic model based on a data set containing hourly load and temperature readings for weekdays during March to September 1996. The stochastic model is expressed by the following regresis the response and sion equations in which the hourly load is the independent variable: the hourly temperature

. The new expected value randomness comes from the term and variance of hourly load and covariances are given by

(17) (18) (19) The are computed from the historical data set, once the effect of temperature is removed. On the other hand, as a time seriesdenoted by the symbol model 3 considers following an ARIMA (1,120,0) process of the form , where is Gaussian white noise with mean zero and variance , and is the autocorrelation coefcient for a 1-h lag. This model has been validated by Valenzuela and Mazumdar [17]. Given a temperature forecast, the new expected values, variances, and covariances of the load . are used as , , and VI. NUMERICAL RESULTS FOR A HYPOTHETICAL MARKET The example system comprises 12 identical sets of eight generators each. The total number of units in the system is 96. Table I shows the characteristics of the generating units: capacity, production cost, mean time to failure, mean time to repair, and the steady-state proportion of time that it is able to generate power. The total nominal capacity of the system is 18 000 MW. The model assumes that innite amount of energy can be bought outside the system; four ownership scenarios of the system: 3, 4, 6, and 12 identical rms, with 4, 3, 2, and 1 eight-unit groups each, respectively; and that all the rms forecast the load with the same accuracy. It is assumed that the generators are dispatched in a prearranged merit order, based on the offered prices. There exists a positive correlation between bids and production costs. Load data from PJM for weekdays of Spring 2002 (March 21 to June 20, 2002) are used in these illustrations. Table II shows the mean and standard deviation of the hourly load. It is assumed

if and , the remaining term. if We study two models based on this expression: model 2 considers as normally distributed for each ; model 3 considers with a time series approach, in which is correlated to and to other terms of the series. In model 2, the effect of temperature is subtracted from the load, to compute the exfor each . Tempected value and variance of the remaining perature is considered to be a deterministic variable, so all the where

30

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 23, NO. 1, FEBRUARY 2008

TABLE II ACTUAL AGGREGATE LOAD MODEL

that hourly loads follow a normal probability distribution. Standard deviations are small enough with respect to the mean so that the probability of negative loads can be neglected. The data points used in the model were scaled by a factor of 0.75 to t into the supply model. For the Cournot model, a nonzero price elasticity of the demand is proposed. A linear demand function with is used having the form being deterministic and constant across all hours , and being a random variable different for every hour. For the Rudkevich model, zero price elasticity of demand is required. A code written in Matlab is used to run the model. Sensitivity analysis is then performed on number of rms (3 12); demand slope MWh ); and peak-defor Cournot models ( mand-to-full-capacity ratio (PDFCR) for Rudkevich model (0.6 1.0). The PDFCR expresses the belief about the anticipated peak demand with respect to full capacity. Bertrand model results play the role of benchmarks. The number of rms in the market affects the expected value of the price and the variance. In all the cases, when the number of rms increases, the results tend to the Bertrand solution. Fig. 1 shows that the expected values and variances of hourly prices for the Cournot model follow a similar prole to the Bertrand model, but always stay above it. Only two cases of demand elasticity are shown (the highest and the lowest), to keep the graphics clear. The other three cases fall between them. Both expected values and variances can reach high values when the elasticity of demand is low. For different values of demand elasticity, expectations and variances with 12 rms remain around one half those when there are only three rms in the market. For the Rudkevich model, for low anticipated peak demand, market concentration does not affect the results in both expectations and variances to a marked degree. The expected values and variances of hourly prices for the Rudkevich model are shown in Fig. 2. Only two cases of anticipated peak load are given for were equal the sake of clarity. The results for or very close to the Bertrand solution; therefore, they are not , the curve of exincluded. In Fig. 2(a), for pected values is atter than in the other cases, and between hour 9 and hour 22, the differences between ownership scenarios are very small. As was expected, all the prices are above Bertrand hourly prices. When the PDFCR is high (close to 1), then the differences are striking. Rudkevich expected prices for low demand hours are more affected by the uctuation in demand than

Fig. 1. Expected values and variances of hourly prices (Cournot model).

Fig. 2. Expected values and variances of hourly prices (Rudkevich model).

the expected prices for peak hours. This produces the effect of leveling of prices. Rudkevich model variances of hourly prices [see Fig. 2(b)] are less disparate for the different ownership sce. The lower the PDFCR, the closer the narios for solutions are to Bertrands curve. Note that, for peak hours, except in the case of 12 rms, Rudkevichs variances are smaller than Bertrands. The reason for this is that Rudkevichs supply functions have smaller slopes than the marginal cost at peak hours. Furthermore, the fewer the number of rms in Rudkevichs model, the higher the expected values, but the smaller the variances, something which is not intuitive. Fig. 3 shows the expected values and variances of average price between hours 13 and 16 for the three bidding models, with sensitivity analysis. Results for the Bertrand model are insensitive to all the parameters and are taken as references. As could be intuited, in all the cases, the expected values increase when the number of rms decreases. When the number of rms is large, the behavior tends to the perfect competition case. As also could be expected, Cournot average prices [see Fig. 3(b)] increase when demand is more inelastic (i.e., decreases in absolute value). The increase may be very high

RUIBAL AND MAZUMDAR: FORECASTING THE MEAN AND THE VARIANCE OF ELECTRICITY PRICES IN DEREGULATED MARKETS

31

Fig. 4. Expected values and variances of hourly prices. Fig. 3. Expected values and variances of average prices between hours 13 and 16.

with respect to Bertrand prices. Rudkevich expected values [see Fig. 3(a)] increase with the PDFCR, implying that a big peak load for a given day will drag up all the hourly prices of that day. However, markups are not that large for anticipated peak loads less than or equal to 90% of total capacity. For the Cournot model, variances of average prices [see Fig. 3(d)] are always above the Bertrand model case. They also increase when the number of rms decreases, and when demand is more inelastic. On the contrary, Rudkevich model variances of average prices [see Fig. 3(c)] have quite a different behavior. The rst thing to point out is that the variances are below that for the Bertrand model for most of the chosen values of peak-demand-to-full-capacity ratio. Second, the variances increase with the number of rms. The explanation for this is again that when the number of rms increase, the market tends to the perfect competition scenario, so the variances get close to that of the Bertrand model. In addition, with fewer companies in the market, Rudkevich prices go up and atten more quickly, and therefore, the slopes of the supply curves are smaller for relatively higher values of the peak-demand-to-full-capacity ratio. Third, for values of PDFCR increasing from 0.6 to 0.9, the variances of average prices decrease in this range of on-peak load. Considering the load stochasticity, and the effect of temperature on the load, after running the three load models described in Section VI, the outputs are compared to extract some conclusions. Fig. 4 shows the expected values and variances of hourly prices under three load models and three bidding models. Rudkevich model was selected with a PDFCR of 0.8 and Cournot model was selected with a demand-to-price slope of MWh . Rudkevich models expected values are close to Cournot models for low demand hours, and closer to Bertrand models for peak hours. There do not appear to be great differences between load models. On the contrary, the variances of hourly prices show a huge difference between load models. Variances at peak hours are extreme for model 1, being twice to ve times larger than in the other two models. Across

Fig. 5. Expected values and variances of average prices.

bidding models, Rudkevich models variances for hours 6 to 23 are half of Cournot models. Fig. 5 depicts the expected values and variances of average prices between hour 13 and hour 18 for the three load models and the three bidding models. In this case, all Rudkevichs and Cournots scenarios are shown. There are no big changes in expected values of average prices across load models. Rudkevich models expected value of average price increases a lot for a forecasted peak demand close to full capacity. Also, in this case, variances of average price are much larger for model 1. As was anticipated, temperature plays an important role in the expected value and variance of hourly prices and average prices. Forecasting temperature accurately can reduce the variance of prices considerably. VII. CONCLUSIONS The numerical results of the previous section are detailed enough to derive some conclusions in the following respects: price behavior with regard to market concentration, price reaction to demand elasticity, and installed capacity. Computations made using the models of Section VI allow us to arrive at some

32

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 23, NO. 1, FEBRUARY 2008

conclusions on the effect of temperature on expected prices and variances. Market concentration is an important factor in the determination of the expected value and the variance of hourly and average prices, especially in the Cournot model for all values of demand elasticity. In the Rudkevich model, the greater the number of rms, the lower the prices and the greater the variances. With a small number of rms in the market, prices tend to level off across hours. The Cournot model helps us to understand and to measure the effect of price elasticity of demand. As is to be expected, a large elasticity brings the prices down and the variances as well. A signicant part of the demand is totally inelastic because it is needed irrespective of price. The remaining part of the demand shows more elasticity. A key factor is to design the market structure in such a way that it provides this elasticity. In order to do this, it is necessary to allow the end consumers to react to different prices in the wholesale market, even though they buy energy in the retail market. This change should be carefully considered by the market designers as an important part of the deregulation process. The Rudkevich model has the advantage of showing the effect of the entire supply system on the prices. Prices are affected by the cost-capacity structure of the market, even by those units that are not running in a given hour. It is clear that if the market has much more capacity than needed, it can assure a better service because it has a lot of energy reserve, and the buyers will appreciate that up to a certain point. Eventually, the rms will charge a bit more to compensate for the investment on the excess capacity. Even for a market which does not have a large excess capacity, the perception of the rms about the daily peak demand affects the price, under the Rudkevich model. The difference between an anticipated peak demand of 90% and of 100% turns out to be important. Then, the question to think about is how to inuence the rms beliefs. Temperature can explain in great part the variance of the load. In the example shown in the previous section, for on-peak hours, temperature explains up to 75% of the variance of the prices. Temperature plays a more important role in determining the hourly load than the load in the preceding hour. What is also true is that the hourly temperatures are very correlated among themselves. REFERENCES [1] J. Bertrand, Theorie mathematique de la richesse sociale, J. Savants, vol. 45, pp. 499508, 1883. [2] A. Cournot, Recherches sur les Principes Mathematiques de la Theorie des Richesses (in English) Transl.:Translation by N. T. Bacon published in Economic Classics [Macmillan, 1897] and reprinted in 1960 by Augustus M. Kelly. Paris, France: Hachette, 1838.

[3] H. Cramer, Mathematical Methods of Statistics. Princeton, NJ: Princeton Univ. Press, 1946. [4] R. J. Green, Increasing competition in the British electricity spot market, J. Ind. Econ., vol. 44, no. 2, pp. 205216, 1996. [5] R. J. Green and D. M. Newbery, Competition in the British electricity spot market, J. Polit. Econ., vol. 100, no. 5, pp. 929953, Oct. 1992. [6] B. F. Hobbs, C. B. Metzler, and J. S. Pang, Strategic gaming analysis for electric power system: An MPEC approach, IEEE Trans. Power Syst., vol. 15, no. 2, pp. 638645, May 2000. [7] B. F. Hobbs, M. P. Rothkopf, R. P. ONeil, and H.-P. Chao, Eds., The Next Generation of Electric Power Unit Commitment Models Boston, MA, Kluwer, 2000. [8] S. Iyengar and M. Mazumdar, A saddle point approximation for certain multivariate tail probabilities, SIAM J. Sci. Comput., vol. 19, pp. 12341244, 1998. [9] C. Kang, L. Bai, Q. Xia, J. Jiang, and J. Zhao, Incorporating reliability evaluation into the uncertainty analysis of electricity market price, Elect. Power Syst. Res., vol. 73, no. 2, pp. 205215, Feb. 2005. [10] P. D. Klemperer and M. A. Meyer, Supply function equilibria in oligopoly under uncertainty, Econometrica, vol. 57, pp. 124377, Nov. 1989. [11] R. Paehlke, Conservation and Environmentalism: An Encyclopedia. New York: Garland, 1995. [12] D. Pilipovic, Energy Risk. Valuing and Managing Energy Derivatives. New York: McGraw-Hill, 1997. [13] R. T. Rockafellar and S. Uryasev, Optimization of conditional value-at-risk, J. Risk, vol. 2, no. 3, pp. 2141, 2000. [14] A. Rudkevich, M. Duckworth, and R. Rosen, Modeling electricity pricing in a deregulated generation industry: The potential for oligopoly pricing in a poolco, Energy J., vol. 19, no. 3, pp. 1948, 1998. [15] C. Ruibal, On the variance of electricity prices in deregulated markets, Ph.D. dissertation, Univ. Pittsburgh, Dept. Ind. Eng., Pittsburgh, PA, 2006. [16] J. Valenzuela, Stochastic optimization of electric power generation in a deregulated market, Ph.D. dissertation, Univ. Pittsburgh, Dept. Ind. Eng., , 2000. [17] J. Valenzuela and M. Mazumdar, Statistical analysis of electric power production costs, IIE Trans., vol. 32, pp. 11391148, 2000. [18] J. Valenzuela and M. Mazumdar, A probability model for the electricity price duration curve under an oligopoly market, IEEE Trans. Power Syst., vol. 20, no. 3, pp. 12501256, Aug. 2005. [19] J. Valenzuela, M. Mazumdar, and A. Kapoor, Inuence of temperature and load forecast uncertainty on estimates of power generation production costs, IEEE Trans. Power Syst., vol. 15, no. 2, pp. 668674, May 2000.

Claudio M. Ruibal received the Ph.D. degree in industrial engineering from the University of Pittsburgh, Pittsburgh, PA, in 2006. He is currently a faculty member in the School of Business and Economics at the Universidad de Montevideo, Montevideo, Uruguay.

Mainak Mazumdar received the Ph.D. degree in industrial engineering from Cornell University, Ithaca, NY, in 1966. He worked as a Research Scientist at the Westinghouse R&D Center during the period 19661981. Since 1981, he has been a faculty member in the Department of Industrial Engineering at the University of Pittsburgh, Pittsburgh, PA.