Deep learning in time series forecasting with transformer models and RNNs.

Area of Science:

• Artificial Intelligence
• Meteorology
• Data Science

Background:

• Accurate weather forecasting is crucial.
• Neural networks, including transformers and RNNs, show promise for time series pattern recognition.

Purpose of the Study:

• To evaluate 14 neural network models for weather variable forecasting.
• To compare the performance of transformer and RNN models for different forecasting horizons.

Main Methods:

• Applied 14 neural network models to weather forecasting tasks.
• Evaluated models using metrics: MedianAbsE, MeanAbsE, MaxAbsE, RMSPE, and RMSE.
• Compared transformer models (Informer, iTransformer, Former, PatchTST) against RNN models (TCN, BiTCN).

Main Results:

• Transformer models demonstrated superior accuracy for long-term pattern capture, with Informer performing best.
• RNN models were more suitable for short-term forecasting but prone to higher errors.
• iTransformer achieved specific error metrics: MedianAbsE 1.21, MeanAbsE 1.24, MaxAbsE 2.86, RMSPE 0.66, RMSE 1.43.

Conclusions:

• Neural networks, particularly transformers, offer significant potential for improving weather forecast accuracy.
• The study provides a basis for selecting appropriate models for weather prediction applications based on forecasting needs.