Heterogeneous data reduction in WSN

Abstract

The transformation of existing power grids into Smart Grids (SGs) aims to facilitate grid energy automation for a better quality of service by providing fault tolerance and integrating renewable energy resources in the power market. This evolution towards a smarter electricity grid requires the ability to transmit in real time a maximum of data on the network usage. A Wireless Sensor Network (WSN) distributed across the power grid is a promising solution, given the reduced cost and ease of deployment of such networks. These advantages come up against the unstable radio links and limited resources of WSN. In order to reduce the amount of data sent over the network, and thus reduce energy consumption, data prediction is a potent solution of data reduction. It consists on predicting the values sensed by sensor nodes within certain error threshold, and resides both at the sensors and at the sink. The raw data is sent only if the desired accuracy is not satisfied, thereby reducing data transmission. We focus on time series estimation with Least Mean Square (LMS) for data prediction in WSN, in a Smart Grid context, where several applications with different data types and Quality of Service (QoS) requirements will exist on the same network. LMS proved its simplicity and robustness for a wide variety of applications, but the parameters selection (step size and filter length) can directly affect its global performance, choosing the right ones is then crucial. Having no clear and robust method on how to optimize these parameters for a variety of applications, we propose a modification of the original LMS that consists of training the filter for a certain time with the data itself in order to customize the aforementioned parameters. We consider different types of real data traces for the photo voltaic cells monitoring. Our simulation results provide a better data prediction while minimizing the mean square error compared to an existing solution in literature.