Parallel algorithm for short-term time series forecasting on BSN mesh

Abstract

Time series forecasting is an efficient approach for future prediction based on past observations. It is useful for real world applications such as, predicting gain, loss and market trends in future, production planning, weather forecasting, flood forecasting, etc. Having larger network size and symmetrical behavior, the recently proposed Bi-Swapped Networks (BSN) is an excellent class of networks for the efficient parallel implementation of such applications in comparison to OTIS networks. It is well accepted that moving average is the best-suited approach for short-term time series forecasting. In this article, we present the parallel mapping of the weighted moving averages of time series forecasting over √n × √n BSN mesh. It requires 5√n-1 intra-group (electronic) and 1 inter-part (optical) moves. We also present network scalability and compare the mapping of proposed parallel algorithm over BSN mesh with its counterpart OTIS mesh network [6]. The proposed approach claims to be cost effective and demands fewer communication moves.