Appliance Usage Prediction Using a Time Series Based Classification Approach Kaustav Basu, Vincent Debusschere, Seddik Bacha Grenoble Electrical Engineering Laboratory(G2E lab) BP 46-38402, Saint Martin d’Heres, France Email:Kaustav.Basu@g2elab.grenoble-inp.fr,Vincent.debusschere@g2elab.grenoble-inp.fr Seddik.bacha@g2elab.grenoble-inp.fr Abstract—Energy management for residential homes and of- fices require the prediction of the usage(s) or service request(s) of different appliances present in the house. The hardware requirement is more simplified and practical if the task is only based on energy consumption data and no other sensors are used. The proposed model tries to formalize such an approach using a time-series based multi-label classifier which takes into account correlation between different appliances among other factors. In this work, prediction results are shown for 1-hour in the future but this approach can be extended to predict more hours in the future as per the requirement(with restrictions). The learned models and decision tree showing the important factors in the input information is also discussed. Keywords-Appliance Usage Prediction, Learning Algorithm, Energy Management in Homes, Data Mining, Smart-Buildings, Multi-label classifiers, decision tree. I. I NTRODUCTION Buildings (residential and tertiary) represent the first en- ergy consumer and the second greenhouse emission source in France. Passive house and positive energy houses are being accepted as a standard for new buildings where the electrical part of energy consumed will be predominant. In order to achieve this goal, energy management have to be set including energy use, meteorology, inhabitants comportment, etc. But these optimization of energy savings can contradict an optimal comfort of the habitants. Home automation system basically consists of household appliances linked via a communication network allowing interactions for control purposes [1]. Thanks to this network, a load management mechanism can be carried out : it is called distributed control in [2]. Load management allows inhabitants to adjust their power consumption according to an expected comfort, the energy price variation and the CO 2 equivalent emissions. A home energy management system is able to determine the best energy assignment plan and a good compromise between energy production and energy consumption [3]. In this study, energy is restricted to the electricity con- sumption and production. [4], [3] present a three-layers (anticipative layer, reactive layer and device layer) household energy control system. This system is both able to satisfy the maximum available electrical power constraint and to optimize a compromise between user satisfaction and cost. The objective of the anticipative layer explained in [5] is to compute plans for production and consumption of services. Anticipating problematic situations requires also prediction capabilities. Even if it is easier to predict overall consumption, it is important to be able to predict the usage of each appliance because, regarding dynamic demand side management, it is also important to evaluate how much energy can be saved thanks to specific requests to the customers like unbalancing requests, load shading or energy price variations. The energy savings depend on appliances : some can be unbalanced, some can be postponed and some cannot be touched. The overall goal of the prediction is described in Fig. 1. The proposed approach is restricted to the prediction of appliance usage based on appliance consumption data, time of the event and meteorological information. learning time adaptative time predictor UI knowledge local database : household history observations, known data predictions for all services time supervised services unsupervised services Figure 1. Principle of the prediction system The problem of appliance usage prediction through con- sumption data is new. [6] deals with the problem of the user behavior prediction in a home automation system using a Bayesian network for a single appliance but a general model for appliance prediction is still lacking. Short term load forecasting (STLF) at the grid level has been studied for some time but at the appliance level, these techniques are yet to be tested. Though STLF uses regressive approaches whereas the proposed approach is based on classification but the strategies used in the domain of energy load prediction led to the choice of inputs to the predictor.