No abstract

This paper investigated the development of a hybrid model for wind speed forecast, ranging from 1 to 46 days, in the northeast of Brazil. The prediction system was linked to the widely used numerical weather prediction from the ECMWF global ensemble forecast, with neural networks (NNs) trained using local measurements. The focus of this study was on the post-processing of NNs, in terms of data structure, dimensionality, architecture, training strategy, and validation. Multilayer perceptron NNs were constructed using the following inputs: wind components, temperature, humidity, and atmospheric pressure information from ECMWF, as well as latitude, longitude, sin/cos of time, and forecast lead time. The main NN output consisted of the residue of wind speed, i.e., the difference between the arithmetic ensemble mean, derived from ECMWF, and the observations. By preserving the simplicity and small dimension of the NN model, it was possible to build an ensemble of NNs (20 members) that significantly improved the forecasts. The original ECMWF bias of −0.3 to −1.4 m/s has been corrected to values between −0.1 and 0.1 m/s, while also reducing the RMSE in 10 to 30%. The operational implementation is discussed, and a detailed evaluation shows the considerable generalization capability and robustness of the forecast system, with low computational cost.