Performance analysis of dimensionality reduction techniques for demand side management

The advancement of the electric grid has led to tremendous growth in data generated from the installed sensors. Smart meters measure the electric energy usage of a consumer, transmit the measured data to the utility and receive pricing information. This requires a two way communication between the utility and the end user. With the projected increase in the number of deployed smart meters, utilities would be facing challenges in handling huge quantities of data, referred to as Big Data. For the analysis of the large data to be tractable, we need to extract important lower dimensional features from raw measurements. In this paper we critically analyze dimensionality reduction of smart meter data for smart grid applications. We compare performance of two dimensionality reduction techniques, Random Projection and Principal Component Analysis, on projecting smart meters data onto a linear subspace of reduced dimensions. We compute the Euclidean distance between pair of data samples in the original and reduced dimensions and obtained the mean and standard deviation of the relative error. Additionally, we cluster the users using the original data and after applying dimensionality reduction. The sum of square error (SSE), distance between datapoints and the centroid in a given cluster, is used to compare the clustering performance of the two techniques.