Post-processing of the Weather Research and Forecasting (WRF) mesoscale model by artificial neural networks

The Weather Research and Forecasting mesoscale model (WRF) has been used to simulate hourly 10m wind speed. This model is able to solve atmospheric equations with a resolution up to tens of meters. However, since processes as turbulence, radiation exchange, cumulus and microphysics are represented by means of physical parameterizations, WRF surface outputs are affected by systematic errors also due to uncertainties of the initial and boundary conditions provided by global models. In this study a preliminary approach to develop bias correction and reduction is presented, based on post-processing WRF output by Artificial Neural Networks (ANN). Postprocessed WRF output at a single location in the city of Taranto, in the southern part of Apulia region, has been validated against ground data from a weather monitoring station. In particular, the ANN algorithm has a feed-forward multilayer perceptron architecture. In order to achieve better correction of the bias, a feature selection has been performed. Moreover, to estimate the best ANN setup, a tuning on the representative network parameters (hidden layer number, neuron number per layer, transfer and training function) was also carried out. The performances of whole procedure (deterministic model post-processed by statistical algorithm) have been evaluated in terms of root mean squared error (RMSE) and Pearson’s correlation (PC).