Load identification from Power Recordings at Meter Panel in Residential Households Kaustav Basu, Vincent Debusschere, Seddik Bacha Abstract—Identification of electrical appliance usage(s) from the meter panel power reading has become an area of study in its own right. Many approaches over the years have used signal processing approaches at a high sampling rate (1 second typically) to evaluate the appliance load signature and subsequently used pattern recognition techniques for identification from (a) previously trained classifier(s). The proposed approach tries to identify the usage of high power consuming appliance(s) by using the aggregate power consumption at 10 minutes interval from the meter panel. The novelty of the approach lies in using a time series windowing approach which gives addition information about an aggregate power state. The usage of hour of the day as input to the systems also takes into account the temporal behavior of residential users. The usage of Multi-label classification approach for identification is also new for this domain. The model is tested over the IRISE data set and the results are encouraging. Due to its low sampling rate with time stamped aggregate power at 10 minutes scale as the only input from the user, the proposed approach is both practical and affordable. Index Terms—Non intrusive load monitoring, Residential appliance usage, domestic load separation, Multi-label classification, Data mining. I. I NTRODUCTION The reduction of domestic household energy cost is a major area of study in the present context. In the near future, the main issue for civil engineering is the thermal insulation of buildings, but in the longer term, the issues are those of renewable energy (solar, wind, etc.) and smart buildings. Load management allows inhabitants to adjust power consumption according to expected comfort, energy price variation and CO 2 equivalent emissions [1]–[4]. It is important to be able to identify the usage of each appliance because, regarding dynamic demand side management, it is important to evaluate how much energy can be saved thanks to request to customers like unbalancing requests or energy price variations. The energy savings depend on appliances: some can be postponed and some cannot be changed. From smart grid point of view, the task requires the separation of the total load into its constituent components. The primary approach of load separation are based on identification of state transitions which in most cases is done by the ON/OFF transition identification, Fig. 1. The pioneering work in load separation was started by [5] in which he proposed methods to identify individual appliance signal from the ON/OFF transitions of appliances, resulting in corresponding change in overall power observed at the smart meter. In the last two decades there have been sizeable amount of work to this effect [6]–[9] and each method proposes to reduce the limitations of the All authors are from the Grenoble Electrical Engineering Laboratory (G2ELab) BP 46, 38 402, Saint Martin d’Heres, France. Mail K. Basu : Kaustav.Basu@g2elab.grenoble-inp.fr Mail V. Debusschere : Vincent.Debusschere@g2elab.grenoble-inp.fr Mail S. Bacha : Seddik.Bacha@g2elab.grenoble-inp.fr Fig. 1. Classifier architecture. previous method. Approaches typically consist of identifying the steady state or in some cases transient appliance state changes and these identified instances are called signatures, in the next step these signatures are matched with earlier learned models using a pattern recognition algorithm. The drawback of these approaches are mainly hardware requirement due to high sampling rate and the impracticality of the process being totally non-intrusive. In [10] it is seen that some of the high power consuming appliances, such as water heater or washing machine, can be identified with reasonable precision even at sampling rate of 15 minutes. This low rate of sampling compared to 1 second reduces the hardware complexity of the process considerably, especially considering that most of the high power consuming appliances have low frequency of usage, typically once a day (this could represent an issue for the later predictions). For these appliances, we can consider a situation where a user gives a time stamped account of his high power consuming appliances for a week and in the end gets his energy management plan for every month during the year. In cases where the user cannot monitor the appliance usage then only sensors need to be used for training. In the first case, the diagram shown in Fig. 2 present the synoptic of load separation : this is an identification of the high power consuming appliances. The proposed model tries to formalize such an appliance identification by using a time-series windowing approach where the only input after the training phase is the time stamped aggregate power from the power meter. The time is represented as hour of the day [11]. For the identification purpose this model uses multi-label classification [12]–[15]