Complex-valued neural network schemes for online processing of wind signal

A novel forecasting technique based on a complex-valued (vectorial) representation of the wind signal is proposed. Unlike with the standard univariate techniques, this way a simultaneous wind speed and wind direction forecasting is performed. To cater for the nonlinear and nonstationary nature of wind, a cascaded combination of a complex-valued recurrent neural network (CRNN) and a complex-valued linear finite impulse response (CFIR) filter is used as a computational forecasting model. Simulation results on real world measurements confirm that the proposed approach provides more accurate estimates than the commonly used individual univariate approaches.