Ensemble Learning and Time-Domain Feature Mining based Appliance Power Consumption Pattern Recognition

As technology has advanced, using smart meters instead of conventional ones has changed. These smart meters are essential parts of intelligent grids, offering significant advantages to many stakeholders regarding social, environmental, and economic constraints. A detailed method of fine-grained collection and analysis of the metering data is necessary to benefit the different smart grid stakeholders. This results in collecting, transmitting, and processing a large amount of data with an exponentially raised dimension. It seriously constrains the system processing resources, latency, data management, transmission effectiveness, and analysis response time. In this regard, a new hybridization of ensemble learning and time-domain feature extraction is suggested for the automatic classification of appliances through the processing of their power consumption data. Time-domain feature mining makes up for the previously noted deficiency and can result in a considerable real-time data dimension reduction without sacrificing important information. Data in compressed form can therefore be processed, examined, stored, and transmitted with efficiency. It offers a significant reduction in the post cloud-based classification latency along with computational and transmission effectiveness. The classification is carried out using the robust ensemble learning (EL) classifiers. The performance of considered EL classifiers is evaluated for the case of a real multi-class and multi-device smart meter dataset. The devised solution achieves a highest average classification accuracy of 88.33% for a 10-class problem while securing a compression gain of 32.73.