Energy-efficient self-adapting online linear forecasting for wireless sensor network applications

Abstract

New energy-efficient linear forecasting methods are proposed for various sensor network applications, including in-network data aggregation and mining. The proposed methods are designed to minimize the number of trend changes for a given application-specified forecast quality metric. They also self-adjust the model parameters, the slope and the intercept, based on the forecast errors observed via measurements. As a result, they incur O(1) space and time overheads, a critical advantage for resource-limited wireless sensors. An extensive simulation study based on real-world and synthetic time-series data shows that the proposed methods reduce the number of trend changes by 20%~50% over the existing well-known methods for a given forecast quality metric. That is, they are more predictive than the others with the same forecast quality metric