Research On Reconfigurable Computing For Time Series Forecasting

Time series forcasting as a practically useful data analysis method has drawntremendous attentions in scientific research. However, complicated algorithmarchitecture and huge hardware resource occupation have greatly limitedapplications of time series forecasting. Together with increasing applications oftime series forecasting in general embedded systems, efficient forcasting methodsconsuming only restricted hardware resources have become more and morenecessary. As one of the mainstreams in the development of computing,reconfigurable computing(RC) provides a feasible solution to this problem withcustomized high performance parallel computing and flexible reconfiguration.Therefore, the fusion of time series forecasting and RC can provide a novelsolution and basic research framework to the high performance computingresearch of time series forecasting consuming only limited computing resource.So this fusion research has both theoretical and practical significance.Although a number of research works have been proposed to improve thecomputing efficiency of time series forecasting algorithms, applications of RChave surprisingly been ignored in most of these works. Meanwhile, there is anobvious conflict between the efficient computing requirement of complicatedtime series forcasting algorithms and limited RC hardware resources. Given allthese, with careful comparison and analysis, we choose least squares supportvector machine(LS-SVM) as the target algorithms in this paper. Furthermore, wefocus on the RC research of time series forecasting from the perspective of RCapplicability improvement of LS-SVM, RC method with balanced resourceoccupation and computing efficiency, and task scheduling in limited hardwareresources. We here in mainly describe the following work.(1)In LS-SVM based time series forecasting, the high computing complexityand improper linear equations solving method influence its applicability ofreconfigurable computing. To solve this problem, we propose a local modelingmethod of LS-SVM based on clustering strategy. Firstly, we adopt K-meansclustering algorithm with reasonable cluster number to get the local modelingsamples and reduces space complexity. Then, modified Cholesky factorization isused to get analytical, stable and low complex solutions of linear equations.Experiments show that CLS-SVM could obtain better training and predictionefficiency compared to LS-SVM with acceptable precision drop. Moreover, it can reduce the space complexity and get analytical, stable and low complex solutions.All these property make it suitable for reconfigurable computing.(2)To solve the conflict of hardware resources occupation and computingefficiency in the reconfigurable computing of LS-SVM, we propose an LS-SVMreconfigurable computing method based on partial dynamic self-reconfigurablesystem. First we present partial dynamical self-reconfigurable systemarchitecture. Then we propose a time multiplexing and spacial parallel computingstructure to balance the hardware resource occupation and computing efficiency.Experiment results show that when compared to non-time multiplexingarchitecture, our method has high hardware utilization while achieving highcomputing efficiency. This makes the embedded application of LS-SVM withrestricted hardware resources possible.(3)To realize efficient time series forecasting with RC, we explore the taskscheduling in FPGA based partial dynamic RC system. To solve the existingproblems including improper assumptions in reconfigurable partition(RP)divisionmethods and anti-fragments technology, ignorance of configuration prefetch andlow efficiency in optimized scheduling method, with careful study of schedulingmechanism and optimal scheduling method, we propose a heuristic based ontypical heterogeneous multi-core system list scheduling method. With staticdivision strategy of RP, RP sizes are decided under the principle of improvingflexibility and resources utilization. Then chip layout is realized with modifiedminimum horizontal method to reduce fragments. Finally, by active configurationprefetch and task insertion under the limit of serial configuration, we improve thescheduling performance. Simulation experiments and real world application toCLS-SVM algorithm all show that, our heuristic method has a good schedulingperformance and generality. Meanwhile, our heuristic also achieves an obviousefficiency improvement compared to optimal scheduling algorithm.