Before describing the proposed methodology for anomaly detection indetail, let’s first look at one customer energy consumption provided in Figure1. Figure1. Energy consumption visulaiztino with line charts The line Chart given in figure 1 shows the energy for one of the customer.Line charts are the most common visualization technique of time series.
However, with the fine-grained consumption readings provided by smart meters,this method is no longer effective to uncover the underlaying patterns inenergy consumption.In thispaper, a unique visualization technique is employed. It is built upon the work in 7.
The pixel based technique is used to visually encode numerical values into colors. It gives an overview of large time span and provides the possibility to recognize patterns or exceptions in the powerconsumption time series, and to analyze the anomaly scores. Figure 2 shows thesame energy consumption visualized with pixel based technique.
This techniqueis showing that this customer is exhibiting high consumption period from 7 A.Mto 23 P.M, and low consumption period from 23 P.
M to 7 A.M. Figure2.
Energy consumption visulaiztino with pixel based technique A. Temporal Context Energy consumption is influenced by a set of context variables such as occupancy andappliance usage. To be able to analyze such contextual variables, extra sensinginfrastructure should be used.
However, a humanactivity typically follows regular temporal. For instance, appliance usage is highly correlated totemporal context 12. Therefore, in this paper, occupancy andappliance usage are indicated by temporal context sets as follows:· High consumption period: this set contains readingsduring weekdays from (7 a.m. to 12 a.
m.).· Low consumption period: this set contains readingduring weekdays from (12 a.
m. to 7 a.m.).
· Weekend: this set contains reading taken during the weekend, when residential buildings are likelyto have increased occupancy, and commercial premises are unoccupied.Multipletemporal context sets can be also defined for each meter reading consideringthe usage of the building (residential or commercial). B.
Geographic Area Effect Based on the metadata obtained from smartmeters, the geographic area is defined as the set of buildings that areexpected to be influence similarly by the same context variables. Figure 3.represents six customers from the same geographic area. For each customer, thearea marked in red represents a deviation from the normal pattern.
However, in overall, this pattern represents atrend among different customers, therefore, flagging it as an anomaly is not necessarily correct. For example, during holidays, when residential buildingshave increased occupancy will resultin higher energyconsumption. Another example is temperature variation for a specific area can influence energy consumptionconsiderably.