Improving Short Term Weather Forecast Accuracy with Machine Learning Models

Abstract

Traditionally, weather predictions rely on large, complex physical models that incorporate various atmospheric conditions over extended periods. Weather system disturbances frequently cause these conditions to become unstable, which results in forecasts that are off. These models usually require massive amounts of energy, running on hundreds of nodes in a big HPC (High-Performance Computing) environment. In this research, we present a method for weather prediction that trains basic ML (Machine Learning) models using historical data from various weather stations. These models can function in far less resource-intensive contexts and produce useful forecasts for specific weather conditions in the near future in a relatively short amount of time. The evaluation findings show that these models' accuracy is adequate to supplement the most advanced methods available today. We also show the benefits of using weather station data from several nearby locations instead of depending only on data from the area under forecast.