Deep learning in time series forecasting with transformer models and RNNs

Abstract

Given the increasing need for accurate weather forecasts, the use of neural networks, especially transformer and recurrent neural networks (RNNs), has been highlighted for their ability to capture complex patterns in time series. This study examined 14 neural network models applied to forecast weather variables, evaluated using metrics such as median absolute error (MedianAbsE), mean absolute error (MeanAbsE), maximum absolute error (MaxAbsE), root mean squared percent error (RMSPE), and root mean square error (RMSE). Transformer-based models such as Informer, iTransformer, Former, and patch time series transformer (PatchTST) stood out for their accuracy in capturing long-term patterns, with Informer showing the best performance. In contrast, RNN models such as auto-temporal convolutional networks (TCN) and bidirectional TCN (BiTCN) were better suited to short-term forecasting, despite being more prone to significant errors. Using iTransformer it was possible to achieve a MedianAbsE of 1.21, MeanAbsE of 1.24, MaxAbsE of 2.86, RMSPE de 0.66, and RMSE de 1.43. This study demonstrates the potential of neural networks, especially transformers, to improve accuracy, providing a practical and theoretical basis for selecting the most suitable models for predictive applications.