Color Code Map. University 

Context in source publication

Context 1

... load daily profiles that compose the set of measurements for training and evaluation of SOM maps correspond to two typical medium users: an industry and an university located both in Spain. Data used in the training of the neural network correspond with weekly load curves -from Monday to Friday- of both customers measured in 2003. For simplicity Saturdays and Sundays profiles were not considered in this phase of study. Table I shows the mask associated to daily load profiles -a number of mask for all the load profiles for a given customer-, used in the case of the classification of customers reported in section 5, which allows to identify the data assigned to a cell with the corresponding customer. The table also shows the number of load curves considered for each customer -including anomalous days later eliminated- that determines in certain way the size of the network. An alternative labelling is used for data filtering purposes. This filtering is applied separately to both customers. By means of this labelling a number is assigned to each profile following the next criterion: the last two digits indicate the day of the month and the initial remaining ones the corresponding month (i.e., mm/dd). Thus, label maps that are obtained -see figure 2- allow to identify weekly load data assigned to each cell. As previously stated, the use of information in Time Domain was chosen. Especially, the curves interpolated to 24 values of daily consumption that were used were obtained from load curves recorded every 15 minutes. The reason was simply the good results obtained in previous works accomplished by the authors [8]. In the process of result optimization, different possible configurations were proved for the parameters of the network. A map size of 16x16 cells and a number of 2000 and 1000 steps for primary and secondary training respectively was finally applied. Some results and characteristics observed for the User 2 -University- after a randomly feeding of the network are discussed. The SOM training results obtained with weekly load profiles -from Monday to Friday- for User 2 are shown in the color map presented in figure 1. In this one it is observed the number of profiles that were presented to the SOM and were assigned to the same cell marked with a digit. Likewise, the relative size of the above mentioned cells indicates the level of similarity of the profiles in relation to the rest of the set of data employed in the training. Those and others characteristics –the zones appear marked with different colours depending of data similarity and its distance to data belonging to others clusters-, allow to realize a first draft of the location of anomalous or not typical demand profiles -left bottom corner on the map-. Also, using the label map obtained by means of the criterion already explained (mm/dd), it is possible to identify the particular days assigned to the region of interest –see figure 2-. For example, the study of figure 2 shows labels 501 and 1208, corresponding with two holidays in Spain, located in this area. Also a county holiday marked with label 1009 is located closed to the previous ones. Obviously 1231, 1225 are festive days in Christmas Season. Finally and once the network is trained, it is possible to force it to group data fixing an upper limit of clusters. In this particular case and for a maximum of two clusters the following result was obtained -figure 3-. Searching through the label map the profiles assigned to the region of minor size, it can be observed that these consumptions fit with holidays. Moreover, these distant patterns -from the typical ones- for the customer are located in the opposite corner. If the upper limit of clusters is increased -in an iterative and automatic way- it is possible to optimize the search because punctual or seasonal demand patterns are removed from standard ones. When a maximum numb er of ten clusters is allowed, the fo ur zones defined in figure 4 are found. By means of the label map and plotting the corresponding load profiles, it can be seen that the network is able to distinguish three kinds of profiles: typical consumption patterns, assigned to regions 3 and 4 -see figure 5-; profiles placed in region 1, which due to the characteristic of topographic preservation of SOM are identified as holidays –i.e., anomalous days, see figure 6-; finally profiles that denote standards of different behaviour from the usual ones lie in area 2 -see figure 7-. The previous results show a s the network is not only able of identifying anomalous days with the consistent capacity of cleanliness of the database, but it presents other usefulness or applications. Thus, the network isolates consumption profiles with erroneous measures caused by failure in demand meters and will enable the utilities to manage them in a right way -to reconstruct them or to eliminate wrong data-. Moreover, it allows the study of particular behaviors of the customer. So demand patterns of the studied university identified in cluster 2 corresponds to days of July during which relative night- time demand level growths due to the continuous use of air conditioning and a day-time peak demand reduction owed to student’s holidays period. Therefore, the exposed method is presented as a useful tool in the study and estimation of end use loads -air conditioning in this case- and its possible routes of management -control actions, introduction of new technologies, and efficiency measures-. The first purpose of application of SOM is to obtain a map for the identification and classification of electrical customers with different demand patterns that may be later employed in diverse tasks [9], [10], [11]. In the present section an example of classification is shown using data of the two customers presented at the beginning of this paper. The typical profiles of the above mentioned customers are well-known, showing features that make them notably different in their behaviors. First, the training of the network has been done usi ng all available information without data filtering. Simply it was followed the data conditioning process previously described using the network’s characteristics shown in section 4. Label map obtained using mask 1 for Medium Industry and mask 2 for the University can be seen in figure 8. In that figure it is shown how the network is able to group the customer’s profiles in the same map area separating and differentiating each one. Nevertheless, it is shown how the map presents zones where separation is not clear and labels of both customers are even mixed. This fact suggests that some profiles belonging to different customers are not well suited to typical consumption patterns and therefore the network is not able to distinguish them in a right way. For that reason it was performed a new training eliminating previously the anomalous demand profiles identified by means of the method exposed in the previous section. In this case the filtering was separately applied to both customers. The results can be seen in figure 9: the network is now able to classify correctly and besides separate ...