论文部分内容阅读
We introduce a suite of forecasting methods which improve traditional analog forecasting by combining ideas from state-space reconstruction in dynamical systems and kernel methods developed in harmonic analysis and machine learning.The first improvement is to use Takens' delay-coordinate maps to recover information lost through partial observations.Then,weighted ensembles of analogs are chosen according to similarity kernels featuring an explicit dependence on the dynamical vector field generating the data.