The electricity system inherited from the 19th and 20th centuries has been a reliable but centralized system. With the spreading of local, distributed and intermittent renewable energy resources, top-down central cont...
详细信息
The electricity system inherited from the 19th and 20th centuries has been a reliable but centralized system. With the spreading of local, distributed and intermittent renewable energy resources, top-down central control of the grid no longer meets modern requirements. For these reasons, the power grid has been equipped with smart meters integrating bi-directional communications, advanced power measurement and management capabilities. Smart meters make it possible to remotely turn power on or off to a customer, read usage information, detect a service outage and the unauthorized use of electricity. To fully exploit their capabilities, we foresee the usage of distributed supervised classification algorithms. By gathering data available from meters and other sensors, such algorithms can create local classification models for attack detection, online monitoring, privacy preservation, workload balancing, prediction of energy demand and incoming faults. In this paper we present a decentralized distributed classification algorithm based on proximal support vector machines. The method uses partial knowledge, in form of data streams, to build its local model on each meter. We demonstrate the performance of the proposed scheme on synthetic datasets.
Classification is one of the most widely used methods in data mining, with numerous applications in biomedicine. The scope and the resolution of data involved in many real life applications require very efficient impl...
详细信息
ISBN:
(纸本)9783540681052
Classification is one of the most widely used methods in data mining, with numerous applications in biomedicine. The scope and the resolution of data involved in many real life applications require very efficient implementations of classification methods, developed to run on parallel or distributed computational systems. In this study we describe SVD-ReGEC, a fully parallel implementation, for distributed memory multicomputers, of a classification algorithm with a feature reduction. The classification is based on Regularized generalized eigenvalue classifier (ReGEC) and the preprocessing stage is a filter method algorithm based on Singular Value Decomposition (SVD), that reduces the dimension of the space in which classification is accomplished. The implementation is tested on random datasets and results are discussed using standard parameters.
暂无评论