CLUSTERING-BASED WIND POWER SCENARIO REDUCTION TECHNIQUE J. Sumaili 1 , H. Keko 1 , V. Miranda 1 , A. Botterud 2 and J. Wang 2 1 Instituto de Engenharia de Sistemas e Computadores do Porto (INESC Porto) Faculdade de Engenharia, Universidade do Porto Porto, Portugal jean.sumaili@inescporto.pt, hrvoje.keko@inescporto.pt, vmiranda@inescporto.pt 2 Argonne National Laboratory, Decision and Information Sciences Division Argonne, IL, USA abotterud@anl.gov, jianhui.wang@anl.gov Abstract – This paper describes a new technique aimed at representing wind power forecasting uncertainty by a set of discrete scenarios able to characterize the probabili- ty density function of the wind power forecast. From an initial large set of sampled scenarios, a reduced discrete set of representative or focal scenarios associated with a prob- ability of occurrence is created using clustering techniques. The advantage is that this allows reducing the computa- tional burden in stochastic models that require scenario representation. The validity of the reduction methodology has been tested in a simplified Unit Commitment (UC) problem. Keywords: wind power, uncertainty, scenario re- duc`tion, probability. 1 INTRODUCTION The continuous growth in wind power penetration in- creases variability and volatility in the power system, thus posing new challenges to power system manage- ment and planning. Wind power forecasts are getting more and more mature, offering better predictions gen- erally characterized by a single-value forecast (or point forecast) for each look-ahead time horizon. To better employ the forecasts in practice, the users need addi- tional information concerning the uncertainty of the forecasts. One way of presenting this uncertainty is using a discrete set of scenarios, sampled according to a probability density function associated to the forecasts, and therefore able to characterize the uncertainty ac- cording to the historical error distribution. However, in order to accurately capture the entire characteristic of the wind power forecasted during the prediction period, a decision maker may need to evaluate a high number of scenarios. This is a time-consuming and computational- ly demanding task – and the computational burden is usually not compatible with stochastic programming algorithms run in useful time. This paper describes a new technique aimed at representing the uncertainty associated with wind power in a forecasting horizon by a reduced discrete set of scenarios, where each scenario becomes associated with the probability of a cluster that it represents. The ap- proach is similar to the one presented in [1]. This new set of scenarios may then be used as input in computa- tionally demanding stochastic problems (e.g. unit com- mitment, market bidding). In the proposed methodology, one departs from a wind power scenario generator of Monte Carlo type, such as [2]. This method is used to generate a very large sample: a set of scenarios, which hopefully gives a good discrete approximation to the probability density func- tion of wind power. Then, an evolutionary optimization algorithm [3] is used to cluster the scenarios, identifying the areas of high density, and replacing each cluster by its more representative (focal) scenario. The probability associated to each focal scenario will be given by the probability associated to its cluster within the sample. The motivation for this work is to adequately represent the shape of the full probability density func- tion with a reduced set of scenarios with assigned prob- abilities. This makes the reduction method adequate in risk assessment. In the literature, several scenario reduc- tion methods exist, commonly based on a scenario tree construction method [4], [5]. In this work, a clustering approach is used instead that works directly on the sce- narios generated by Monte Carlo type scenario genera- tion identifying the area of maximum density according to a defined similarity criterion. The reduced equivalent set is still a discrete represen- tation of the probability density function of wind power scenarios, in a space whose dimension is equal to the number of time steps considered, The great advantage of having available the reduced set of focal scenarios is that it allows solving stochastic programming problems in a practical and industrial environment — and this opens the door for the management of the power system based on risk assessment models, something very much needed whenever the uncertainty associated to wind power leads to possibly large hedging costs. This paper presents the details of the model devel- oped and its application in a stochastic unit commitment model, validating the interest in the approach. Unit commitment is a major step in the decision process of system operators – and stochastic unit commitment models are one logical approach to modeling a system with large uncertainty present in data, such as the case 17 th Power Systems Computation Conference Stockholm Sweden - August 22-26, 2011