Identification of electrical energy consumption patterns

In the current context of changes in terms of processes and information systems, the development of methodologies to deal with large volumes of information proves to be extremely important. Encouraged by the increasing technological development, which enables the storing and processing of large amounts of data, Data Analysis tools are increasingly present in companies like EDP Distribuição, in order to take full advantage of the available consumption and production data. INTRODUCTION Data Analysis techniques make it possible to identify patterns that otherwise would hardly be found and reveal hidden correlations, among other useful information. In this context, a methodology aiming at segmenting the universe of Primary Substations (PS) according to their patterns of power consumption is presented in this paper. The database of the study consisted on the active power values of the PS’ transformers, as well as on the active power of the Distribution Generation (DG). In both cases, the values were in 15 minutes samples. Because it was the most recent year for which the data was available, the time period analysed was the year 2014. With these two quantities (active power of PS and active power of DG), the gross load was calculated. In the time period in analysis, there were 384 PS, and so it was easy to see that the amount of data we were dealing in this study was significant. Therefore, we thought it would be convenient to use a programming software capable of store and manipulate effectively greats amounts of data. We opted to develop the project using ‘R’ programming language. Being a free software, it provides a wide variety of statistical computing tools and graphical environments. INITIAL CONSTRAINTS After a short analysis, several problems were detected concerning the quality of the data being studied. The first one consisted on a time asynchronism between the meters from the PS and the DG. To try to minimize this issue, we used a smoothing method, which consisted in a centered moving average of 5 periods. Another problem found was the missing data. Because those failures could compromise the results, the days containing missing values were eliminated. Additional problems were found, such as the inexistence of A in some meters, the incorrect match between PS and DG and also long reconfiguration periods in the medium voltage (MV) network. Due to the misleading conclusions it could result in, we decided not to consider the data from the PS affected with these problems, as the remaining universe is still significant. METHODOLOGY After identifying the problems described previously, and consequent elimination of the PS affected by them, the universe in study changed from 384 to 354 PS. The methodology developed to identify the power consumption patterns, which allowed the segmentation of the universe of PS, is described below. As a result of a first analysis, it was perceived that the power consumption behaviours vary according to the season of the year and day type (Business Day, Saturday or Sunday). Therefore, the first step of the methodology consisted in disaggregating the data according to these specifications. Having in mind that our goal was to capture typical behaviours, only the two characteristic months of each season of the year were considered. In other words, in Winter we studied January and February, in Spring, April and May, in Summer, July and August and in Autumn, October and November. As there are many similarities between the power consumption in days with the same characteristics (for example, the power consumption values in business days of a particular season of the year are very alike between each other), the second step was to determine a unique load profile representative of all the same day types. To do this, we identified and eliminated, using clustering methods, the days with an abnormal power consumption behaviour over the period of study, and calculated the mean of the remaining ones. The process described on the previous paragraphs was repeated for the several seasons of the year and, and in the end, we obtained a Resulting Load Profile which reflects the annual behaviour of each PS. In Figure 1 we can see the result for a particular PS, called ‘Alameda’. 24th International Conference on Electricity Distribution Glasgow, 12-15 June 2017