Abstract

A novel method of point weather forecasting is presented in which an ensemble of 950 individually trained feed- forward back propagation neural networks were developed to produce 24-hour forecasts for nine different cities across the contiguous United States in varying climate regimes totaling 72 forecasts. The ensemble forecasts consisted of the following parameters: 24-hour maximum temperature, minimum temperature, 2-minute maximum sustained wind speed, and accumulated precipitation. The ensemble neural network was trained on dynamical weather models (comprised of the primitive equations of the atmosphere), statistical weather forecasts (dynamical guidance post- processed using linear statistical methods), and human forecasters as input. More than two years of historical weather observations served as verification/targets for the input. Performance of the ensemble is assessed through a comparative analysis between it and the five input predictors to measure relevant error reduction achieved by the network. Error metrics that were used to assess this are: root mean square error (RMSE) and bias. Results indicate significant error reduction across all forecast parameters between the ensemble network forecasts and input model forecasts. Assessing ensemble forecast performance per city shows ensemble reductions of RMSE from the input models for maximum and minimum temperature, on average, exceeding 2-3 degrees Fahrenheit, wind of 3-5 knots and precipitation of 0.10-0.20 inches.

How to Cite