Abstract:
Statistical post-processing techniques are used to remove systematic biases in modeled data. Models have shortcomings in the physical parameterization of weather events and have the inability to handle sub-grid phenomena successfully. The accuracy of forecasts interpolated to station points is limited by the horizontal resolution of the model. The magnitude of the bias at a station point depends upon geographical location and season. A neural network (NN) is a statistical downscaling method that seeks to model the linear or non-linear relationship between a set of different predictors and the predictand. NN’s have a training rule whereby the weights of connections between predictors and the predictand, are adjusted on the basis of the data. NN systems have been developed by using as input, different model variables from the NCEP Ensemble Prediction System (EPS) and Eta model to forecast minimum/maximum temperature and rainfall (Quantitative Precipitation Forecast (QPF) and Probability of Precipitation (PoP)), respectively. Results show some potential for improved NN forecasts over the forecast generated by the Numerical Weather Prediction (NWP) models. The implementation of a NN system can serve as a guidance tool in operational forecasting but with one difficulty that the NWP model has to be frozen, meaning no upgrades or changes on the model.
