A Novel Transfer Learning-Based Intelligent Nonintrusive Load-Monitoring With Limited Measurements

In this article, the real-time nonintrusive load-monitoring (NILM) problem with limited measurements, i.e., low sampling rate data, is investigated. NILM is a technique to identify the various types of appliances by analyzing the voltage and current features collected by sensors, such as smart outlets. It is one of the most important topics in smart grid management and optimization. Although recent NILM methods developed several general models that can be used for various scenarios, these algorithms either require high-frequency measurement devices or have been limited to use for some specific appliances. A real-time intelligent NILM algorithm that is able to be transferred among different appliances with low-frequency data is still lacking and desperately needed. In this article, a novel online learning-based intelligent NILM algorithm has been developed that can, online, infer a variety of appliances using a transferred model with limited measurements. Especially, the developed algorithm integrates the emerging transfer learning technique along with deep neural networks. Two neural networks are used in two different stages: 1) the long short-term memory (LSTM) neural network to extract the lower level spatial and temporal features from the gray-scale image generated by the measurements and 2) the probabilistic neural network (PNN) to classify the appliance type as well as transfer between appliances. Finally, the algorithm is implemented into practical smart outlet hardware to demonstrate its effectiveness.

I. Introduction

Recently, the demand for monitoring power consumption of various appliances has become a crucial issue to the smart grid measurement and optimization [1]. By analyzing appliances’ real-time electric load data, the utility companies have the potential to adjust the power supply efficiently and further optimize energy allocation. This potential has motivated the development of “smart buildings,” where the appliances’ load is dynamically monitored. The specific appliance’s real-time load data provide the opportunity to profile the characteristics of a building’s power demand, e.g., frequency [2] and scheduling [3]. As a result, the utility companies can autonomously optimize the power distribution by balancing the load demand response for every building based on the accurate individual appliance’s data. Nevertheless, obtaining such data is challenging since a single building may have hundreds of individual appliances. Due to the limit of communication and data fusion difficulties caused by large-scale sensor networks, previous studies focused on disaggregating specific appliance data from a general-purpose power meter that measures the sum of all appliances’ load [4]. Those approaches were organized into two steps: 1) disaggregating individual appliance load and 2) identify the appliance type through various load features. However, it becomes more and more difficult to disaggregate individual appliance data from a general meter with a large number of appliances.