Using Neural Networks to Provide Local Weather Forecasts

Artificial neural networks (ANNs) have been applied extensively to both regress and classify weather phenomena. While one of the core strengths of neural networks is rendering accurate predictions with noisy datasets, there is currently not a significant amount of research focusing on whether ANNs are capable of producing accurate forecasts of relevant weather variables from small-scale, imperfect datasets. Also, there is not a significant amount of research focusing on the forecasting performance of neural networks applied to weather datasets that have been temporally rolled-up from a base dataset. In this paper, a survey of existing research on applying ANNs to weather prediction is presented. Also, an experiment in which neural networks are used to regress and classify minimum temperature and maximum gust weather variables is presented. This experiment used a dataset containing weather variables recorded every 15 minutes over the course of a year by a personal weather collection station in Statesboro, Georgia. Data cleansing and normalization were applied to this dataset to subsequently derive three separate datasets representing 1-hour, 6-hour, and 24-hour time intervals. Three different NN structures were then applied to these datasets in order to generate minimum temperature regressions at 15-minute, 1-hour, 3-hour, 6-hour, 12-hour, and 24-hour look-ahead ranges. Maximum gust regressions were also generated for each data set at 1-hour, 3-hour, 6-hour, 12-hour, and 24-hour look-ahead ranges. Finally, neural networks were applied to these datasets to classify freezing events and gusty events at 3-hour, 6-hour, 12-hour, and 24-hour look-ahead ranges