Data driven inference of electrical appliances for effective home smart energy management

Demand response encourages end consumers to voluntarily alter their electricity consumption based on fluctuating supply. In order to facilitate the participation of residential consumers in demand response programs, home smart energy management mechanisms that enable understandable and effective semi-automated control of appliance usage and consumption have been widely proposed. In general, there are two broad home energy management systems techniques. The first technique aims to provide automatic control for appliances inside the house. The second technique aims to learn the energy consumption patterns and provide feedback to the users. This feedback facilitates the users' better understanding of their energy consumption and leads users to take energy efficiency actions such as scheduling the laundry washing job during the off peak period. These techniques can be implemented in three different levels namely smart plug (appliance), smart meter (home) and provider. The state-of-the-art techniques for understanding appliance consumption patterns have focused on sensor based analysis through two different main approaches: having the sensors constantly attached to the appliance, known as an intrusive approach; and predicting appliance usage from a house's total electricity consumption, known as a non-intrusive approach. In an intrusive environment, the appliance usage model is developed based on data coming from sensors that constantly provide usage information for each appliance inside a house. In a non-intrusive environment, the absence of data coming from appliance level sensors imply that the focus is on disaggregating the total electricity load consumption to recognise the contributing appliances. Given the setup/installation cost of home energy management systems, it is realistic to expect that there will always be a large number of houses where only total/aggregate consumption from smart meters will be available. However, unlike custom hardware that record very high frequency (10 kHz to 100 kHz) electricity features, smart meters only retrieve macro features such as real power and reactive power in a low sampling frequency (1Hz or less). Consequently, disaggregation accuracy of current non-intrusive approaches that use low sampling frequency data collected from smart meters is low. Thus, having a disaggregation technique of low frequency data that can provide accurate and detailed insight into appliance electricity consumption patterns as well as how and when an appliance contributes to overall consumption in houses is highly desirable. In this thesis, we proposed and developed non-intrusive techniques on low frequency data that provide a fine-grained understanding of home appliance usage. Firstly, we proposed and developed Wattzup, a computationally efficient disaggregation technique. Wattzup enables accurate prediction of the appliance usage from low frequency appliance data. Wattzup works accurately on up to one reading in every 30 minutes of data. Unlike state-of-the-art load disaggregation techniques that use computationally expensive algorithms for inferring appliance states, Wattzup uses a simple classification method that is computationally efficient yet accurate. This enables Wattzup to be implemented on demand side devices such as home gateways. Secondly, we proposed and developed Wattseal, an unsupervised appliance state modelling technique. Current appliance state modelling techniques rely on a predefined number of operational states per appliance and focuses on detecting the energy usage level per state. Such a coarse-grained view of appliance states may not be able to capture variations due to the presence of substates within a predefined state. Wattseal clusters appliance level data without a predetermined number of operational states and prior assumption of appliance consumption ranges. Thus, it allows accurate detection of distinct appliance states and substates. The experiments show that factoring distinct states and substates and their consumption level modelling both improves disaggregation accuracy and provides customers with a finer understanding of their appliance usage. Thirdly, we proposed and developed iWattseal, a novel technique for incrementally learning appliance states. iWattseal is designed to be computationally efficient for embedded resource constrained devices such as smart plugs. Unlike other appliance state modelling techniques, iWattseal incrementally recognises appliance states when the sensor is attached to the appliance during the training period. While incrementally recognising the appliance state, iWattseal automatically determines the training duration. With this capability, iWattseal can develop a similar appliance model to Wattseal's appliance model with the automatically determined training duration. In addition, iWattseal also maintains disaggregation accuracy while having a shorter training period. The efficient novel data stream clustering algorithm used in the incremental learning enables iWattseal to be embedded into small devices like smart plugs. The combination of these computationally efficient techniques provides detailed appliance states model and improves the accuracy of the load disaggregation technique. Detailed understanding of home appliance usage positively impacts users, by assisting them to control their electricity usage and reduce their electricity cost. It also positively impacts electricity providers, by balancing the energy demand and preempting costly installation of new power generators and transmitters, and society at large, by enabling and promoting sustainable behaviours