Small-world echo state network time sequence prediction method based on MCP penalty function

Abstract

The invention discloses a time series prediction method. An MCP penalty model is adopted to optimize the output weight of a small-world echo state network in the construction of the small-world echo state network; the over-fitting problem occurring when the weight is calculated by a conventional regression method is solved; an LQA approximate decomposition MCP penalty function is selected; approximate solutions of a model are determined, the problem that the MCP penalty function cannot be derived at the original point is solved; the hyper-parameter of the MCP penalty model is optimized on thebasis of the artificial bee colony algorithm integrated with the intersection and extrusion strategy, the convergence effectiveness is improved while the global optimality of the optimized parameter is guaranteed, the effective compression capacity of the MCP penalty model is improved accordingly, and the method has high application value in nonlinear time sequence prediction.

Description

Technical field: [0001] The invention relates to a time series prediction method, in particular to a small-world echo state network time series prediction method based on an MCP penalty function. Background technique: [0002] Time series forecasting has been widely used in industry, finance, military and other fields. Since most of the time series in real life are nonlinear and unstable, the prediction of nonlinear and unstable time series has always been a hot topic. research scholars in various fields. At present, one of the main methods for forecasting nonlinear and unstable time series is to use Echo State Network (ESN). The characteristic of ESN is that it only needs to train the output weights from the reserve pool to the output layer during training, which solves the problems of traditional neural networks that are easy to fall into local optimum and complex training algorithms. Therefore, computing output weights is the key to echo state network learning. [0003...

AI Technical Summary

Problems solved by technology

[0003] Traditional ESN usually uses pseudo-inverse method, Ridge regression method or Lasso regression method when calculating the output weight, which is prone to collinearity and over-fitting problems when dealing with high-dimensional nonlinear data

Images