Comparative analysis of recurrent neural networks for weather prediction in the Antarctic region

Abstract

Numerical weather prediction is a well-established method that uses current atmospheric conditions as inputs to solve wind, temperature, pressure and humidity equations. This study examines the use of deep learning for meteorological forecasting using historical data from the Bharati Station, Antarctica. Different unique recurrent neural network models have been developed using a deep learning framework and explicitly trained to predict the weather conditions of the next 24 to 48 hours.The effectiveness of our proposed approach is compared against state-of-the-art neural network algorithms, and the results demonstrate better forecasting performance.In this study, the Transformer model has the lowestRoot Mean Square Error (RMSE) of 0.000478, making it one of the most efficient models in the neural networks investigated. This progress facilitates a more efficient development process, which, in turn, enhances the accuracy of weather forecasts.