You are on page 1of 6

S* International Conferenceon ProbabilisticMethods Applied to Power Systems, Iowa State University, Ames, Iowa, September 12-16,2004

The Use of PAR@) model in the Stochastic Dual Dynamic Programming Optimization Scheme Used in the Operation Planning of the Brazilian Hydropower System
M.E.P.Maceira, J.M.Damtkio reservoirs, corresponding to the four Brazilian regions (South, Southeast, North and Northeast illustrated in Figure I), and for the thermal plants, plus the energy interchanges among regions. The state system includes the stored energy in the beginning of each month for each of the four aggregated reservoirs or subsystems and information about hydrologic trend, that is, the last p observed energy inflows in each of the four subsystems. This number, p, can vary from month to month and from subsystem to subsystem.

Abstrou-In September 2000, the Brazilian system dispatch and spot prices were calculatrd twice, using different inflow forecasts for that month, as in the last five days of August the inflows to the resewoirs in South and Southeast region change 200%. The first run used a smaller forecasted e n e w inflow, and the second B higher energy inflow. Contrary to expectations, the spot price in the second run, with the higher energy inflow, was higher than the one found in the first run. This piper describes the problem, presents the special features of PAR(p) model that allow the descdbcd behavior and shows the solution taken to avoid the problem.

Index Term- Time Series Models, Operation Planning, Bydrothermal Systems, Dynamic Programming.

I. INTRODUCTION cham of optimization models with different planning horizons and degrees of detail in system representation composes a computational system for the electric-energetic operation planning and programming of the Brazilian generating system [I]. The chain was proposed to the Brazilian Independent System Operator (ONS) and currently is being gradually implemented. ONS is responsible for central system optimization and dispatch according to clearly defined rules agreed by all industry members and approved by the Regulatory Body (ANEEL). The optimization models also compute the water values that form the basis for determining the Wholesale Energy Market (MAE) spot price. On the top of chain, the long term operation planning Z ] , applies a stochastic dual model, called NEWAVE [ dynamic programming algorithm, which uses Benders Decomposition, to produce an operation strategy for each stage (month) of the planning period [3]. The operation strategy, given the system state at the beginning of the stage, produces generation targets for each of the four aggregated

A :

Fig 1 Brazilian Subsystem

M.E.P.Maceira, is with the Electric Power Rereareh Center,CEPEL, Rio de Janeiro, Brazil (e-mail: with Rio de Janeiro State University, UERI, Bradl. 1.M.Damizio, is with the Electric Power Research Center, CEPEL,Rio de Janeiro, Brazil (email: and with Rio de Janeiro State University, UERI, B r a d .

In stochastic streamflow modeling, this corresponds to the use of the well known and widely used Periodic Autoregressive PAR(p) model [4], which use as forecast model for monthly flows has been demonstrated in the comprehensive study of Noakes et al. [5], which considered several models. Noakes et al. used thirty monthly unregulated river flow time series. The last thirty-six observations were omitted in the model identification and estimation phases of the study. One step ahead forecasts statistics of the obtained forecasts in the omitted period were used to compare the models. The PAR(p) model obtained by identification of the models order in each month using the partial autocarrelation function (PACF), and estimating the parameters using the Yule-Walker equations [4] was found to be the hest model for the overall data in the study. Noakes et al., also mmment Although other

Copyright Iowa State University, 2004


model may be more parsimonious, the PARPACF model gives the most adequate model due to the seasonal correlation effect" and "Besides being used for forecasting, PAR@) models can also be employed for simulating hydrologic sequences", that is, to generate a great number of energy inflow scenarios [6]. The good results obtained with the PARRACF modeling approach, both in one-step ahead forecasts and simulation of drought sequences reasonably similar to the historical ones together with the easy mathematical tractability, recommended the use of this approach in NEWAVE algorithm to model the energy inflow stochasticity. It was considered suflicient to use a maximum order p equal to 6, although the effective order varies for each month, according to the PARPACF modeling approach. This model is called PAR(6). The PAR(6) model generates energy inflow scenarios that will he used in the simulation process of NEWAVE model and in the construction of the expected costto-go function of each month produced by the stochastic dual dynamic programming algorithm. At the end of each month, the dispatch model is run by ONS, based on the inflow forecast for the neat month. Besides being used to determine the hydm and thermal generation, plus the energy interchanges among subsystems, the model calculates the spot price for each subsystem. The spot price is used for settling transactions in the Wholesale Energy M a r k e t , and is therefore of great commercial importance. In September 2000, the system dispatch and spot prices were calculated twice, using different inflow forecasts for that month, as in the last five days of August the inflows to the reservoirs in South and Southeast region change 200%. The first run used a smaller forecasted energy inflow, and the second a higher energy inflow. Contrary to expectations, the spot price in the second run, with the higher energy inflow, w a s higher than the one found in the first run. The analysis of the case made by [7]suggested that the problem occurred because of an error in the calculation of Bender's cut that constitutes the expected cost-to-go-function of each month. The problem was further investigated and it was detected that the expected energy inflows produced by the PAR(p) model in February 2001 decrease sharply when the energy inflow in September 2000 was increased and this was the reason of the increase in the spot price. The suggestion in 171 proved not be the case. The objective of this work is to describe the problem occurred in September 2000 in the operation planning of the Brazilian hydropower system, present the special features of PAR@) model that allow the described behavior and show the solution taken to avoid the problem.

stochastic dual dynamic programming [ Z ] , [31, [81, represented by the following backward recursive equation, can solve the operation dispatch problem:

V t = T, T-1, ..., 1 subject to Storage balance equation in each subsystem aggregated reservoir V,+l = V, + FC y A, -Q, -S, -EV, Demand s~iuulv eauation in each subsvsteni k

Bounds in storage in each subsystem aggregated resemoir

v,,, 5 iT
A4aximiim hydro production in each subsystem aggregared reservoir

Q, sG
Lower bounds on total outj7ow in each subsystem aggregated reservoir

A4axinnim generation in each thermalplant GT, < GT max, Flow limits among subsystems Fmin, <F1 <Fmax, Set of mriltivariate linear constraints representing the costto-gofunction (Bender's cut)

c.",t+Iv,+I + %+lA, +...+ .ap,t+lAt-p+l+ ~ n s t




indexes the stages state vector at the beginning of stage t (Vf, At.l, ...,A,.

The operation strategy minimizes the expected value of the operation cost during the planning period, which is composed of fuel costs plus penalties for failure in load supply. If the inflow volumes are known at the beginning of each stage,

subsystem energy storage level energy inflow to subsystems expected value of the operation cost from stage t to the end of the planning period under the optimal operation policy represents the probability distribution of energy inflows A, conditioned by the state XI expected value thermal generation of each thermal plant discount rate aggregated hydro production in each subsystem generation cost of each thermal plant aggregated energy spillage in each subsystem evaporated energy on aggregated reservoirs proportion of energy inflow due to conrrollable energy irflow, which represents the inflow volumes that can be stored in the system reservoir



proportion of energy inflow due to uncontrollable energv inflow, which represents lateral inflow volumes arriving at run-of-the-river plants, which have no associated reservoir FC controllable energy inflow correction curve, associated to the energy storage level, to take into account the effect of head variation FkkJ energy flow from subsystem k to subsystem I D, energy demand V maximum reservoir storage Q maximum aggregated hydro production Qmin minimum aggregated hydro production GTmaxmaximum thermal generation Fmin maximum flow between subsystems Fmax minimum flow between subsystems R. Bender's cut coefficient associated to subsystem energy storage level Bender's cut coeffcient associated to energy inflow in x* the k previous stage The periodic autoregressive model, denoted by PAR(p), can he written as:

Bender's cut coefficient is positive. Next, it will he detailed the Bender's cut coefficient calculation responsible for a z , , ~ positive in the dispatch operation problem of September 2000. The adopted horizon of the backward recursive simulation was ten years, starting in December 2009 and ending in September 2000. To illustrate the problem, it will he sufficient to carry out the application of the recursive equation (1) since February 2001 to September 2000 considering only one energy inflow scenario. Tables 1 to 5 show the PAR@) coefficients of Southeast subsystem from February to October.


where Z,, t = 1,2, ...,is the seasonal time series with period
s (in the case of monthly time series s is equal to 12). The time







index I may he regarded as a function of the year T (T = 1, ..., N), and the season m (m = I , ..., s): t = (T - 1) s + m. p m is the mean period m, a,,, is the standard deviation period m and a, is the time uncorrelated series which is independent of Z,, and it has also zero mean and variance equal to a : @ " .

4 1



* (Om1Uwd





Take p,lid,, and h, as the Lagrange multipliers associated to TABLE5 storage balance equation, demand supply equation and cost* (amled to-go multivariate linear constraints respectively, of the 0421 0 470 $8 dispatch problem of stage I. After the solution of each -0 089 4 151 c operation dispatch problem, the Bender's cut coefficients G,,, 0 459 0 581 q,,,, ..., nAPc that will he added to the expected cost-to-go function of stage (t-l), are calculated deriving the operation storage balance, demand supply equations and the active cost in relation to the energy stored in the beginning of the linear constraint representing the cost-to-go function, besides stage I and the energy inflows in stages (t-l), ..., (t-p), the Bender's cut coefficients of Southeast subsystem from obtaining: February 2001 to October 2000. TABLE 6 ICY,= [I + (aFciav,)Y A,] X, (3) FEBRUARY 2001 OPEFATION DISPATCH PARAMETERS xAl,, =FcOPn, - ( l - y ) $ h , , +"Al,t+l$Ph, + n A 2 , t + l h ( 4 )

nA2,,= W ; z , -(1-~)$%,,, + %,,+,$;h, + nA3.,+A(5)

With the values of table 6 is possible to calculate the Bender's cut coefficients that will be aggregated to the When, in September 2000, the system dispatch and spot expected cost-to-go function of January 2001. = (0.9179 x 0.9142 x 0.641 x-726) (0.0858 x 0.641 x prices were calculated twice, using different inflow forecasts nAl,, for September and the consolidated and known energy inflows 684) + (-1663 x 0.641 x 0.992) + (0 x 0.992) for August, July, etc, the spot price associated to a smaller nA,,= -391 - 37 - 1057 + 0 = -1488 = 205 + 19 + 554 + 0 = 780 forecasted energy inflow could he smaller than the associated liA2,1 x A 3 = , -380-36- 1029 + 0 = -1449 to the higher forecasted energy inflow only if the q,,, 111. PROBLEMIDENTIFICATION


xA4,t =256 + 24 + 695 + 0 = 978 nAIZ = 662 + 63 + 1792 + 0 = 2522 nA6,, = -1 1 16 - 107 - 3019 + 0 = -4249


With the values of table 7 is possible to calculate the Bender's cut coeEcients that will be aggregated to the expected cost-to-go function of December 2000. nAl,,=-572-54- 1379+775=-1232 ne,,= 0 + 0 + 0 1437 = -1437 n = 0 + 0 + 0 + 970 = 970 nA4,, = 0 + 0 + 0 + 2503 = 2503 nAS,* = 0 + O + O - 4212 =-4212 A6.f = 0

With the values of table 10 is possible to calculate the Bender's cut coefficients that will be aggregated to the expected cost-to-go bnction of September 2000. rrAl,, = -31 1 - 28 - 1077 + 2442 = 1026 ae1=99+9+344-4115=-3657 xA3,, = -384 - 35 - 1329 + 0 = -1748 x A4.I = 0 6 AS,, = 0 n A6.t = 0 Now, the Bender's cut that will be aggregated to expected cost-to-go function of September 2000 can be Written as:


attl2 -786V,+, + 1026A, - 3657A,.,


- 1748A,.> +

C.",,+IV,+I+n,,,t+,*, +...+%, ,+,A,-,+,+coflst


With the values of table 8 is possible to calculate the Bender's cut coefficients that will be aggregated to the expected cost-to-go function ofNovember 2000. rrA,,,=-624-59-1187- 1427=-3297 re,,= 0 + 0 + 0 + 962 = 962 nA3,, = 0 + 0 + 0 + 2482 = 2482 R ~= 0 ~+ 0 , + 0 4178 = -4178 n AS,, = 0 6115.1 = 0


It can be observed that the Bender's cut coefficient associated to the forecasted energy inflow, A,, is, in this case, positive. The appearance of the positive coefficient in the Bender's cut is due to the expression of the expected energy inflow of February 2001 for the Southeast subsystem as a function of the state at September 2000, includes the inflow at September 2000 times a negative constant. This can be shown by backsubstitution of the equations of the January inflow, December inflow, November inflow, October inflow and September inflow in the equation of February inflow, as is shown in the following. Considering X, as (Z, - p,,,), the regression of February in the PAR(p) model can be written as:

XFcb = 0.641X,,- 0.336X0, 1.090Xs, + 1.830XA8,

+ 0.624XN,-0.421&

With the values of table 9 is possible to calculate the Bender's cut coefficients that will be aggregated to the expected cost-to-go function of October 2000. 6,,,,=-537-51-2669+952=-2305 x m , = 0 + 0 + 0 + 2463 = 2463 ~~~,~=0+0+0-4145=-4145
AS,, = 0

Substituting the known regression of January in the PAR(p) model (XJm= 0.934 X , , , ) in the above expression, it can be \nitten as:

Xpeb = (0.641 x 0.934)X0,, - 0.336X0, 0.421&, - 1.09OXs., + 1.830X~,,


Regrouping common terms, the regression of February can be \nitten as:

XFcL = 0.2627XD, 1 .830XA,,

+ 0.624XN,, - 0.421&,

- 1.09OXs,


In the same way, December regression is given by & , = 0.971 XNov. Substituting the following expression in the previous one, we obtain:


XFeb= (0.2627 x 0.971) X , 1.09OXse, + 1.830XA,, or

0.624XN, - 0.421&,

Southeast SubSystem February

XFeb = 0.8791XN,, - 0.421XO,, - 1.090Xs, + 1.830XA,,

Substituting the known PAR@) model regression of November (XNoy = 0.816 &,) in the above expression, results in: XF, = (0.8791 x 0.816)&, 1.830X~,, or
XFe,= 0.2963&,- 0.421&,


Fig 2 February PACF


+ 1.830XA,,

Once more, the October regression is given by X, = 0.47OXs, - 0.151XA,, + 0.S81X,,l. Substituting this in the previous expression and regrouping the terms, it is obtained the regression of February/2001 inflow as a function of the September inflow forecast and past inflows: -0.95O7Xsep + 1.7853X~,, + O.l722X,,1 X F=~ ~ This regression shows a negative coefficient multiplying the September energy inflow forecast. It can be noted that the negative September coefficient which showed up after these back-substitutions has its roots on negative parameters in the Februarys model for the months Nagged (December), 4lagged (October) and 5-lagged (September).
IV. A PROPOSED SOLUTION The identification procedure for the February model was reappraised. The PACF of this month, Figure 2, shows a significant value at lag 1, non-significant values from lag 2 to lag 5, and a significant value at lag 6. Using the adopted classical identification rule, choose p as rhe largest signijkanr order i so that the partial correlalionsfor k > i are iiof signifcanf, this last significant value yielded a model of order 6. On the order hand, even a pure Box&Jenkinss identification rule application [9] would enhance the classical identification rule, by neglecting not very strong significant partial correlations. This would turn the model more parsimonious without increasing too much its residual sum of squares. After a general reappraisal of the identifications for the Brazilian subsystems, it was discovered that the same pattern occurred in two other months of South and in one month to North and Northeast subsystems.

Taking apart the occurrence of runs of non-significant partial correlations before a significant one, as in the these particular cases, in general, one may state that, with the exception in the cases with flows affected by snowmelt, there is no physical reason for the inflow in any month to affect any following inflows by a negative coefficient. This would pose a positive constraint in all the coefficients of PAR(p) models. On the other side, in a pure statistical point of view, the obtained negative coefficients are the best representation of the correlation pattem revealed by the data and the literature on PAR(p) modeling of streamflows is full of examples of estimated negative coefficients; examples can be found in [SI and [lo]. As a compromise solution, it was proposed to maintain the classical rule of identification, but to reduce the order of the PAR@) model of each stage whenever the first coefficient resulting fiom any of twelve back substitutions of the equations model turns out to be negative. This model was called PAR(6a). The result was equivalent to the above cited approach of considering those, may be, spurious, significant partial correlation, as insignificant, and considering the more parsimonious modeling. It was shown that this PAR(6a) model can also reproduce the historical occurrence of severe droughts. Tables 11 to 14 compares the order selected from models PAR(6) and PAR(6a).
Jan Feb

1 I I

1 6 I

Jul A Z O Sep

PAR(6) 3
1 1

PAR(6a) 3




Jan I Feb Marl





1 4


1 1


I I 1 I 2

I Jul I I Ago I I ser, I I - - ,

I w c I I



I 1 2



I 1

I 2

] 1


1 Ds
T m L E 14



occurred in September 2000 in the Brazilian dispatch operation mode is not due to errors in formulation or implementation in the stochastic dual dynamic programming . algorithm. Otherwise, it has been shown the use of classical approach of modeling PAR(p) model can potentially produce this pmhlem The proposed solution maintain the classical rule of identificationhut reduces the order of the model whenever the fmt coefficient resulted t7om twelve hack substitutions of the regressions turns out to be negative. These more parsimonious models presented a more reliable behavior even when applied to September 2000 Brazilian dispatch problem.

Figures 3 and 4 compares the historical (seventy years) and synthetic (two thousand Years) average energy inflow produced by PAR(6) and PAR(6a) models for the Southeast subsystem starting t7om different September energy inflows. Two sets of 2000 synthetic energy inflows are presented. The pink one uses the near avenge energy inflow for September, of the first run dispatch operation model in September 2000. The yellow one uses the higher energy inflow for September, dispatch operation model in September of the second 2000. It can be seen that the parsimonious model, PAR(6a), presents a more reliable behavior.

The authors would like to thank the very important contributions of L.A.Terry, F.S.Costa and A.C.G.Melo to the development ofthis work.





M E . p . e W L.A.Tew, ] . M . D ~ ~ F .,S . C . , ~ ~A ~.c.G.M~~~, "Chain of 0pti"tion Models for Selling the Energy Dispatch and Spot Price in lhe Brazilian System", Power Sysrem Compmrion Conference - PSCC'OZ, Sevillq S p b , lune 24-28,2002, M,E,P,Maceh, C,B.Mercio, B.G,Gorenstin, S,H,F,Cunhq C,suanno, MC.Sacmento, A.S.Klige"Energy Ew.hation of The NorihINonheastem and Southisoutheastem btemnnection with W A V E Model", VI Symposium q'S'cialim in Elecfric Oprofiond ond E.~pamion Planning - SEPOPE, Salvador, Brazil, May 1998. MV.F.Penh, "Optimal Stochastic Operations Scheduling of Large Hydroelectric System", lnrernotionol Joirmol o f Eleceic Pmver and Ene~S~~f~ms,Vol.ll,No.3 161-l69,luly, ,pp. 1989. I.D.Salas, I.W.Delleur, V.Yevjcvich, W.L.Lanc, Applied Modeling o f Hydroelenie Series,Water Resources Publications, 1980. D.INoaker, A.I.McLeod, K.W.Hipel, "Forecasting Monthly Riveflow Times Send', fnfernor;omlJoumal o f forrcostmg, YOI 1, pp. 179-190,

[6] M.E.P.Maceh, C.V.Bezeq "Stachastic S t d o w Model for Hydroelectric Systems", Sh PmbobrLlic Merhodr Applied 10 Pmwr Sysrena - PMAPS-1997,Vancouver, C a m k September, 1997. [7] PSR C ~ ~ ~ l t "Analysis ~ r i a , of Spot Pnce Projection for September and October of 2000". Technical Report (in Pomguese), Draft 2, O c t o b e r ,

Fig. 3 Historical x Synthetic Average Energy Mom of Subsystem - PAR(6) Model


[SI M.V.F. Perch, L.M.V.G. Pinlo, "Stochastic Optlmvation of a Multireservoir Hydroelectric System: A Decomposition Approach'', Wafer Resotmes Rereorch, Vol. 21, No, 6, Pages 779-792,lune.l985. [9] G.E.P. BO% G.M. Jenkins, Time I r l e s Aml,.i. - Forecosring and Coned, Holdcn-Day, 1970. [IO] RM. Thompstonc, K.W. Hipel, A.I. Mcleod, "Forecasting QuarlerMonthly Rjvefflow", Technical Report, TR-8447,Department of Statistical&Actuarial Sciences, University of Westem Ontario,Londoh Ontario, Canada, May, 1984.

Fig. 4 Historical x Synlhete Averagc Energy M o w s of Southeast Subsystem - PAR(6a) Model