No abstract

Detection of steady states and identification of small electrical loads in a household or office grid are important in efficient smart energy management. This paper proposes a method that combines two machine learning techniques, unsupervised K-means clustering, and supervised k-Nearest Neighbours classification techniques, to train a system that can effectively identify the low voltage DC electrical load, and at the same time detect whether it is in its steady state. This is done by comparing the features extracted from signatures of the electric current waveforms of equipment. The combination of Kmeans and kNN in the initialisation stage removes the need to know all the training elements beforehand, and thus, considerably simplifies the process. In the normal operation stage, kNN was used to identify the new unknown test element to the cluster that has the majority votes from its nearest neighbours. The centroids obtained from the K-means clustering aided in the determination of whether the system is in steady state. The method has been successfully implemented on a low voltage DC office grid, with commonly used office equipment.

No abstract

Aiming at the rising number of dc appliances and the growing interest in their monitoring systems, this paper describes the injection of intelligence into dc pico-grids that are made up of "dumb" appliances and loads. Due to reality of economic, dc appliances and loads are usually low in cost and lack intelligence and communication features for effective monitoring and management. This paper proposes a smart sensor design for dc pico-grid with the use of a single sensor multiple loads and states detection in monitoring the "dumb" appliances. This eliminates the need to have intelligence and communication features for every appliance. With the smart sensor, several such smart dc pico-grids can be bundled into bigger scale of smart nano-grid or micro-grid. In addition to knowing how much energy or power the pico-grid is using; the smart sensor also provides load disaggregation and state-change detection. The states of the loads can be learnt and detected via the signatures and features obtained from the transient state or the steady state of the entire grid's current waveform. Computational intelligence techniques, k-Nearest Neighbours, K-Means clustering and other algorithms are used in the system for loads classification and state-change detection. Working together with the software in the smart sensor are hardware implementation of low cost operational amplifiers and logic gates; these hardware help to share the burden on the controller and release resources for the controller to perform more advanced processes. Experimental results are presented to demonstrate the operation of the smart sensor in dc pico-grid.