Abstract
This paper presents the preliminary results of our research activity aimed at forecasting the number of voltage sags in distribution networks. The final goal of the research is to develop proper algorithms that the network operators could use to forecast how many voltage sags will occur at a given site. The availability of four years of measurements at Italian Medium Voltage (MV) networks allowed the statistical analyses of the sample voltage sags without performing model-based simulations of the electric systems in short-circuit conditions. The challenge we faced was to overcome the barrier of the extremely long measurement times that are considered mandatory to obtain a forecast with adequate confidence. The method we have presented uses the random variable time to next event to characterize the statistics of the voltage sags instead of the variable number of sags, which usually is expressed on an annual basis. The choice of this variable allows the use of a large data set, even if only a few years of measurements are available. The statistical characterization of the measured voltage sags by the variable time to next event requires preliminary data-conditioning steps, since the voltage sags that are measured can be divided in two main categories, i.e., rare voltage sags and clusters of voltage sags. Only the rare voltage sags meet the conditions of a Poisson process, and they can be used to forecast the performance that can be expected in the future. However, the clusters do not have the characteristics of memoryless events because they are sequential, time-dependent phenomena the occurrences of which are due to exogenic factors, such as rain, lightning strikes, wind, and other adverse weather conditions. In this paper, we show that filtering the clusters out from all the measured sags is crucial for making successful forecast. In addition, we show that a filter, equal for all of the nodes of the system, represents the origin of the most important critical aspects in the successive steps of the forecasting method. In the paper, we also provide a means of tracking the main problems that are encountered. The initial results encouraged the future development of new efficient techniques of filtering on a site-by-site basis to eliminate the clusters.