Skip to Main Content
This paper focuses on a locally recurrent multilayer network with internal feedback paths, the IIR-MLP. The computation of the partial derivatives of the network's output with respect to its trainable weights is achieved using backpropagation through adjoints and a second order global recursive prediction error (GRPE) training algorithm is developed. Also, a local version of the GRPE is presented in order to cope with the increased computational burden of the global version. The efficiency of the proposed learning schemes, as compared to conventional gradient based methods, is tested on the wind prediction problem from 15 min to 3 h ahead on a site, using spatial correlation and facilitating measurements from nearby sites up to 40 km away.