Multi-step Learning Rule for Recurrent Neural Models: An Application to Time Series Forecasting

摘要: Multi-step prediction is a difficult task that has attracted increasing interest in recent years. It tries to achieve predictions several steps ahead into the future starting from current information. The this work development of nonlinear neural models for purpose building multi-step time series schemes. In context, most popular are based on traditional feedforward networks. However, kind model may present some disadvantages when long-term problem formulated because they trained predict only next sampling time. paper, partially recurrent network proposed as better alternative. For model, learning phase with imposed, which allows obtain future. order validate performance dynamic behaviour future, three different data have been used study cases. An artificial series, logistic map, and two real sunspots laser data. Models networks also compared against model. results suggest than can help improving accuracy.