No abstract

Modern unorganized machines--extreme learning machines and echo state networks--provide an elegant balance between processing capability and mathematical simplicity, circumventing the difficulties associated with the conventional training approaches of feedforward/recurrent neural networks (FNNs/RNNs). This work performs a detailed investigation of the applicability of unorganized architectures to the problem of seasonal streamflow series forecasting, considering scenarios associated with four Brazilian hydroelectric plants and four distinct prediction horizons. Experimental results indicate the pertinence of these models to the focused task.

No abstract

No abstract

No abstract

Echo state networks (ESNs) are promising options for performing time series prediction, as they establish an attractive tradeoff between dynamical processing capability and simplicity in the training process. The recurrent character of these structures is essentially concentrated on a nonlinear processing stage known as dynamic reservoir, whereas the output layeralso termed readoutis allowed to assume the form of a combiner that is, as a rule, linear with respect to its free parameters. Recently, Butcher et al. and Boccato et al. proposed readout design paradigmsbased, respectively, on Volterra filtering and extreme learning machinesthat were shown to enhance the information extraction potential of ESNs without compromising the intrinsic simplicity of their training process. In this work, these approaches are comparatively analyzed, using different reservoir configurations, in the context of monthly streamflow series forecasting. The analysis is based on data from the hydroelectric plant of Furnas, covering three different periods with distinct hydrologic profiles and including the canonical linear approach for readout design as a benchmark. The results reveal that the aforementioned proposals may lead to a relevant performance improvement, thus indicating that the reservoir is not per se capable of fully exploring the nonlinear relationships underlying the focused data.

No abstract

No abstract